diff --git a/AUTHORS.md b/AUTHORS.md index 4ee0542098..8c4a113fc2 100644 --- a/AUTHORS.md +++ b/AUTHORS.md @@ -4,6 +4,7 @@ | backyes | Yan-Fei Wang | | baiyfbupt | Yi-Fan Bai | | beckett1124 | Bin Qi | +| ChengduoZH | Cheng-Duo Zhao| | chengxiaohua1105 | Xiao-Hua Cheng | | cxwangyi, yiwangbaidu, wangkuiyi | Yi Wang | | cxysteven | Xing-Yi Cheng | @@ -21,6 +22,7 @@ | jczaja | Jacek Czaja | | JiayiFeng | Jia-Yi Feng | | kbinias | Krzysztof Binias | +| kexinzhao | Ke-Xin Zhao | | kuke | Yi-Bing Liu | | lcy-seso | Ying Cao | | lipeng-unisound | Peng Li | diff --git a/CMakeLists.txt b/CMakeLists.txt index cfaab206e1..4117f07721 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -55,12 +55,14 @@ option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF) option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) -option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF) +option(WITH_DISTRIBUTE "Compile with distributed support" OFF) option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF) option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF) option(WITH_CONTRIB "Compile the third-party contributation" OFF) +option(WITH_ANAKIN "Compile with Anakin library" OFF) +option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE}) # CMAKE_BUILD_TYPE if(NOT CMAKE_BUILD_TYPE) @@ -147,7 +149,16 @@ include(external/any) # download libn::any include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/cares) -include(external/grpc) + +if(WITH_DISTRIBUTE) + if(WITH_GRPC) + include(external/grpc) + else() + include(external/leveldb) + include(external/brpc) + endif() +endif() + include(external/snappy) # download snappy include(external/snappystream) include(external/threadpool) @@ -183,7 +194,10 @@ set(EXTERNAL_LIBS if(WITH_GPU) include(cuda) include(tensorrt) -endif(WITH_GPU) + include(external/anakin) +else() + set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when GPU is set." FORCE) +endif() if(WITH_AMD_GPU) find_package(HIP) diff --git a/Dockerfile b/Dockerfile index 80a96983ec..752fea5951 100644 --- a/Dockerfile +++ b/Dockerfile @@ -24,12 +24,12 @@ COPY ./paddle/scripts/docker/root/ /root/ RUN apt-get update && \ apt-get install -y --allow-downgrades \ - git python-pip python-dev openssh-server bison \ + git python-pip python-dev python-opencv openssh-server bison \ libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \ wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \ curl sed grep graphviz libjpeg-dev zlib1g-dev \ python-matplotlib gcc-4.8 g++-4.8 \ - automake locales clang-format swig doxygen cmake \ + automake locales clang-format swig cmake \ liblapack-dev liblapacke-dev \ clang-3.8 llvm-3.8 libclang-3.8-dev \ net-tools libtool ccache && \ @@ -76,8 +76,7 @@ RUN easy_install -U pip && \ pip install sphinx-rtd-theme==0.1.9 recommonmark RUN pip install pre-commit 'ipython==5.3.0' && \ - pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ - pip install opencv-python + pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' #For docstring checker RUN pip install pylint pytest astroid isort diff --git a/benchmark/.gitignore b/benchmark/.gitignore index 7b66e8a5b5..fb4114356d 100644 --- a/benchmark/.gitignore +++ b/benchmark/.gitignore @@ -7,3 +7,6 @@ paddle/rnn/imdb.pkl caffe/image/logs tensorflow/image/logs tensorflow/rnn/logs +fluid/models/*.pyc +fluid/logs +fluid/nohup.out diff --git a/benchmark/fluid/Dockerfile b/benchmark/fluid/Dockerfile new file mode 100644 index 0000000000..b9eaca5ee6 --- /dev/null +++ b/benchmark/fluid/Dockerfile @@ -0,0 +1,22 @@ +FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04 +RUN apt-get update && apt-get install -y python python-pip iputils-ping libgtk2.0-dev wget vim net-tools iftop python-opencv +RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/lib/libnccl.so +RUN pip install -U pip +RUN pip install -U kubernetes paddlepaddle + +# IMPORTANT: +# Add "ENV http_proxy=http://ip:port" if your download is slow, and don't forget to unset it at runtime. + +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()\npaddle.dataset.flowers.fetch()" | python' +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.mnist.train()\npaddle.dataset.mnist.test()\npaddle.dataset.imdb.fetch()" | python' +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.imikolov.fetch()" | python' +RUN pip uninstall -y paddlepaddle && mkdir /workspace + +ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin +ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root + +ADD *.whl / +RUN pip install /*.whl && rm -f /*.whl && chmod +x /usr/bin/paddle_k8s + +ENV LD_LIBRARY_PATH=/usr/local/lib +ADD fluid_benchmark.py recordio_converter.py models/ /workspace/ diff --git a/benchmark/fluid/README.md b/benchmark/fluid/README.md index 7071e9fdcd..28cade4634 100644 --- a/benchmark/fluid/README.md +++ b/benchmark/fluid/README.md @@ -24,14 +24,18 @@ Currently supported `--model` argument include: * Run the following command to start a benchmark job locally: ```bash - python fluid_benchmark.py --model mnist --device GPU + python fluid_benchmark.py --model mnist --device GPU ``` You can choose to use GPU/CPU training. With GPU training, you can specify `--gpus ` to run multi GPU training. + You can set async mode parameter server. With async mode, you can specify + `--async_mode` to train model asynchronous. * Run distributed training with parameter servers: + * see [run_fluid_benchmark.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/fluid/run_fluid_benchmark.sh) as an example. * start parameter servers: ```bash PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=1 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method pserver + sleep 15 ``` * start trainers: ```bash @@ -42,13 +46,37 @@ Currently supported `--model` argument include: PADDLE_PSERVER_PORT=7164 PADDLE_TRAINER_IPS=192.168.0.2,192.168.0.3 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model mnist --device GPU --update_method nccl2 ``` +## Prepare the RecordIO file to Achieve Better Performance + +Run the following command will generate RecordIO files like "mnist.recordio" under the path +and batch_size you choose, you can use batch_size=1 so that later reader can change the batch_size +at any time using `fluid.batch`. + +```bash +python -c 'from recordio_converter import *; prepare_mnist("data", 1)' +``` + ## Run Distributed Benchmark on Kubernetes Cluster +You may need to build a Docker image before submitting a cluster job onto Kubernetes, or you will +have to start all those processes mannually on each node, which is not recommended. + +To build the Docker image, you need to choose a paddle "whl" package to run with, you may either +download it from +http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_en.html or +build it by your own. Once you've got the "whl" package, put it under the current directory and run: + +```bash +docker build -t [your docker image name]:[your docker image tag] . +``` + +Then push the image to a Docker registry that your Kubernetes cluster can reach. + We provide a script `kube_gen_job.py` to generate Kubernetes yaml files to submit distributed benchmark jobs to your cluster. To generate a job yaml, just run: ```bash -python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --parallel 1 --device GPU --update_method pserver " --disttype pserver +python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --gpus 8 --device GPU --update_method pserver " --disttype pserver ``` Then the yaml files are generated under directory `myjob`, you can run: diff --git a/benchmark/fluid/args.py b/benchmark/fluid/args.py new file mode 100644 index 0000000000..68a3d42d7a --- /dev/null +++ b/benchmark/fluid/args.py @@ -0,0 +1,126 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse + +__all__ = ['parse_args', ] + +BENCHMARK_MODELS = [ + "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm" +] + + +def parse_args(): + parser = argparse.ArgumentParser('Fluid model benchmarks.') + parser.add_argument( + '--model', + type=str, + choices=BENCHMARK_MODELS, + default='resnet', + help='The model to run benchmark with.') + parser.add_argument( + '--batch_size', type=int, default=32, help='The minibatch size.') + # args related to learning rate + parser.add_argument( + '--learning_rate', type=float, default=0.001, help='The learning rate.') + # TODO(wuyi): add "--use_fake_data" option back. + parser.add_argument( + '--skip_batch_num', + type=int, + default=5, + help='The first num of minibatch num to skip, for better performance test' + ) + parser.add_argument( + '--iterations', type=int, default=80, help='The number of minibatches.') + parser.add_argument( + '--pass_num', type=int, default=100, help='The number of passes.') + parser.add_argument( + '--data_format', + type=str, + default='NCHW', + choices=['NCHW', 'NHWC'], + help='The data data_format, now only support NCHW.') + parser.add_argument( + '--device', + type=str, + default='GPU', + choices=['CPU', 'GPU'], + help='The device type.') + parser.add_argument( + '--gpus', + type=int, + default=1, + help='If gpus > 1, will use ParallelExecutor to run, else use Executor.') + # this option is available only for vgg and resnet. + parser.add_argument( + '--cpus', + type=int, + default=1, + help='If cpus > 1, will use ParallelDo to run, else use Executor.') + parser.add_argument( + '--data_set', + type=str, + default='flowers', + choices=['cifar10', 'flowers'], + help='Optional dataset for benchmark.') + parser.add_argument( + '--infer_only', action='store_true', help='If set, run forward only.') + parser.add_argument( + '--use_cprof', action='store_true', help='If set, use cProfile.') + parser.add_argument( + '--use_nvprof', + action='store_true', + help='If set, use nvprof for CUDA.') + parser.add_argument( + '--no_test', + action='store_true', + help='If set, do not test the testset during training.') + parser.add_argument( + '--memory_optimize', + action='store_true', + help='If set, optimize runtime memory before start.') + parser.add_argument( + '--use_fake_data', + action='store_true', + help='If set ommit the actual read data operators.') + parser.add_argument( + '--profile', action='store_true', help='If set, profile a few steps.') + parser.add_argument( + '--update_method', + type=str, + default='local', + choices=['local', 'pserver', 'nccl2'], + help='Choose parameter update method, can be local, pserver, nccl2.') + parser.add_argument( + '--no_split_var', + action='store_true', + default=False, + help='Whether split variables into blocks when update_method is pserver') + parser.add_argument( + '--async_mode', + action='store_true', + default=False, + help='Whether start pserver in async mode to support ASGD') + parser.add_argument( + '--use_reader_op', + action='store_true', + help='Whether to use reader op, and must specify the data path if set this to true.' + ) + parser.add_argument( + '--data_path', + type=str, + default="", + help='Directory that contains all the training recordio files.') + args = parser.parse_args() + return args diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py index c1d458970a..aa70783ecd 100644 --- a/benchmark/fluid/fluid_benchmark.py +++ b/benchmark/fluid/fluid_benchmark.py @@ -24,90 +24,7 @@ import paddle.fluid.core as core import paddle.fluid.profiler as profiler import paddle.fluid.transpiler.distribute_transpiler as distribute_transpiler -BENCHMARK_MODELS = [ - "machine_translation", "resnet", "vgg", "mnist", "stacked_dynamic_lstm" -] - - -def parse_args(): - parser = argparse.ArgumentParser('Fluid model benchmarks.') - parser.add_argument( - '--model', - type=str, - choices=BENCHMARK_MODELS, - default='resnet', - help='The model to run benchmark with.') - parser.add_argument( - '--batch_size', type=int, default=32, help='The minibatch size.') - parser.add_argument( - '--learning_rate', - type=float, - default=0.001, - help='The minibatch size.') - # TODO(wuyi): add "--use_fake_data" option back. - parser.add_argument( - '--skip_batch_num', - type=int, - default=5, - help='The first num of minibatch num to skip, for better performance test' - ) - parser.add_argument( - '--iterations', type=int, default=80, help='The number of minibatches.') - parser.add_argument( - '--pass_num', type=int, default=100, help='The number of passes.') - parser.add_argument( - '--data_format', - type=str, - default='NCHW', - choices=['NCHW', 'NHWC'], - help='The data data_format, now only support NCHW.') - parser.add_argument( - '--device', - type=str, - default='GPU', - choices=['CPU', 'GPU'], - help='The device type.') - parser.add_argument( - '--gpus', - type=int, - default=1, - help='If gpus > 1, will use ParallelExecutor to run, else use Executor.') - parser.add_argument( - '--data_set', - type=str, - default='flowers', - choices=['cifar10', 'flowers'], - help='Optional dataset for benchmark.') - parser.add_argument( - '--infer_only', action='store_true', help='If set, run forward only.') - parser.add_argument( - '--use_cprof', action='store_true', help='If set, use cProfile.') - parser.add_argument( - '--use_nvprof', - action='store_true', - help='If set, use nvprof for CUDA.') - parser.add_argument( - '--no_test', - action='store_false', - help='If set, test the testset during training.') - parser.add_argument( - '--memory_optimize', - action='store_true', - help='If set, optimize runtime memory before start.') - parser.add_argument( - '--use_fake_data', - action='store_true', - help='If set ommit the actual read data operators.') - parser.add_argument( - '--profile', action='store_true', help='If set, profile a few steps.') - parser.add_argument( - '--update_method', - type=str, - default='local', - choices=['local', 'pserver', 'nccl2'], - help='Choose parameter update method, can be local, pserver, nccl2.') - args = parser.parse_args() - return args +from args import * def append_nccl2_prepare(trainer_id): @@ -142,7 +59,7 @@ def append_nccl2_prepare(trainer_id): "nccl-based dist train.") -def dist_transpile(trainer_id): +def dist_transpile(trainer_id, args): if trainer_id < 0: return None, None @@ -164,7 +81,12 @@ def dist_transpile(trainer_id): training_role = os.getenv("PADDLE_TRAINING_ROLE") t = distribute_transpiler.DistributeTranspiler() - t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) + t.transpile( + trainer_id, + pservers=pserver_endpoints, + trainers=trainers, + sync_mode=not args.async_mode, + slice_var_up=not args.no_split_var) if training_role == "PSERVER": pserver_program = t.get_pserver_program(current_endpoint) pserver_startup_program = t.get_startup_program(current_endpoint, @@ -208,36 +130,57 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc, place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) exe = fluid.Executor(place) exe.run(startup_prog) - feed_var_list = [ - var for var in train_prog.global_block().vars.itervalues() - if var.is_data - ] - feeder = fluid.DataFeeder(feed_var_list, place) + + if not args.use_reader_op: + feed_var_list = [ + var for var in train_prog.global_block().vars.itervalues() + if var.is_data + ] + feeder = fluid.DataFeeder(feed_var_list, place) iters, num_samples, start_time = 0, 0, time.time() for pass_id in range(args.pass_num): train_losses = [] - for batch_id, data in enumerate(train_reader()): + if not args.use_reader_op: + reader_generator = train_reader() + batch_id = 0 + data = None + while True: + if not args.use_reader_op: + data = next(reader_generator, None) + if data == None: + break + if iters == args.iterations: + break if iters == args.skip_batch_num: start_time = time.time() num_samples = 0 - if iters == args.iterations: - break - loss = exe.run(train_prog, - feed=feeder.feed(data), - fetch_list=[avg_loss]) + + if args.use_reader_op: + try: + loss = exe.run(train_prog, fetch_list=[avg_loss]) + except fluid.core.EnforceNotMet as ex: + break + else: + loss = exe.run(train_prog, + feed=feeder.feed(data), + fetch_list=[avg_loss]) iters += 1 - num_samples += len(data) + batch_id += 1 + # FIXME(wuyi): For use_reader_op, if the current + # pass is not the last, the last batch of this pass + # is also equal to args.batch_size. + if args.use_reader_op: + num_samples += args.batch_size * args.gpus + else: + num_samples += len(data) train_losses.append(loss) print("Pass: %d, Iter: %d, Loss: %f\n" % (pass_id, iters, np.mean(train_losses))) - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sec\n' % - (num_samples, train_elapsed, examples_per_sec)) - print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))) + print_train_time(start_time, time.time(), num_samples) + print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))), # evaluation - if not args.no_test and batch_acc != None: + if not args.no_test and batch_acc and not args.use_reader_op: pass_test_acc = test(exe, infer_prog, test_reader, feeder, batch_acc) print(", Test Accuracy: %f" % pass_test_acc) @@ -251,10 +194,14 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc, def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc, args, train_prog, startup_prog, nccl_id_var, num_trainers, trainer_id): - feed_var_list = [ - var for var in train_prog.global_block().vars.itervalues() - if var.is_data - ] + place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) + if not args.use_reader_op: + feed_var_list = [ + var for var in train_prog.global_block().vars.itervalues() + if var.is_data + ] + feeder = fluid.DataFeeder(feed_var_list, place) + # generate fake: if args.use_fake_data: for var in feed_var_list: @@ -271,7 +218,6 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, "value": 1.0, "dtype": var.dtype}) - place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0) if nccl_id_var and trainer_id == 0: #FIXME(wuyi): wait other trainer to start listening time.sleep(30) @@ -288,12 +234,21 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, num_trainers=num_trainers, trainer_id=trainer_id) - feeder = fluid.DataFeeder(feed_var_list, place) for pass_id in range(args.pass_num): num_samples = 0 iters = 0 start_time = time.time() - for batch_id, data in enumerate(train_reader()): + if not args.use_reader_op: + reader_generator = train_reader() + batch_id = 0 + data = None + while True: + if not args.use_reader_op: + data = next(reader_generator, None) + if data == None: + break + if iters == args.iterations: + break if args.profile and pass_id == 0 and batch_id == 5: profiler.start_profiler("All") elif args.profile and pass_id == 0 and batch_id == 10: @@ -302,39 +257,50 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader, if iters == args.skip_batch_num: start_time = time.time() num_samples = 0 - if iters == args.iterations: - break - if args.use_fake_data: - loss, = exe.run([avg_loss.name]) + if args.use_fake_data or args.use_reader_op: + try: + loss, = exe.run([avg_loss.name]) + except fluid.core.EnforceNotMet as ex: + break else: loss, = exe.run([avg_loss.name], feed=feeder.feed(data)) if args.update_method == "pserver": exe.bcast_params() - num_samples += len(data) + if args.use_reader_op: + num_samples += args.batch_size * args.gpus + else: + num_samples += len(data) iters += 1 if batch_id % 1 == 0: print("Pass %d, batch %d, loss %s" % (pass_id, batch_id, np.array(loss))) - train_elapsed = time.time() - start_time - examples_per_sec = num_samples / train_elapsed - print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % - (num_samples, train_elapsed, examples_per_sec)) - if not args.no_test and batch_acc != None: + batch_id += 1 + + print_train_time(start_time, time.time(), num_samples) + if not args.no_test and batch_acc and not args.use_reader_op: + # we have not implement record io for test + # skip test when use args.use_reader_op test_acc = test(startup_exe, infer_prog, test_reader, feeder, batch_acc) print("Pass: %d, Test Accuracy: %f\n" % (pass_id, test_acc)) - exit(0) def print_arguments(args): vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and vars(args)['device'] == 'GPU') - print('----------- resnet Configuration Arguments -----------') + print('----------- Configuration Arguments -----------') for arg, value in sorted(vars(args).iteritems()): print('%s: %s' % (arg, value)) print('------------------------------------------------') +def print_train_time(start_time, end_time, num_samples): + train_elapsed = end_time - start_time + examples_per_sec = num_samples / train_elapsed + print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' % + (num_samples, train_elapsed, examples_per_sec)) + + def main(): args = parse_args() print_arguments(args) @@ -342,7 +308,7 @@ def main(): # the unique trainer id, starting from 0, needed by trainer # only nccl_id_var, num_trainers, trainer_id = ( - None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1"))) + None, 1, int(os.getenv("PADDLE_TRAINER_ID", "0"))) if args.use_cprof: pr = cProfile.Profile() @@ -356,7 +322,7 @@ def main(): fluid.memory_optimize(fluid.default_main_program()) if args.update_method == "pserver": - train_prog, startup_prog = dist_transpile(trainer_id) + train_prog, startup_prog = dist_transpile(trainer_id, args) if not train_prog: raise Exception( "Must configure correct environments to run dist train.") diff --git a/benchmark/fluid/kube_gen_job.py b/benchmark/fluid/kube_gen_job.py index 39ba207fd9..9da8a69af1 100644 --- a/benchmark/fluid/kube_gen_job.py +++ b/benchmark/fluid/kube_gen_job.py @@ -49,7 +49,7 @@ def parse_args(): parser.add_argument( '--fluid', default=1, type=int, help='whether is fluid job') parser.add_argument( - '--rdma', action='store_ture', help='whether mount rdma libs') + '--rdma', action='store_true', help='whether mount rdma libs') parser.add_argument( '--disttype', default="pserver", diff --git a/benchmark/fluid/models/machine_translation.py b/benchmark/fluid/models/machine_translation.py index 635b3373dd..69541adf6b 100644 --- a/benchmark/fluid/models/machine_translation.py +++ b/benchmark/fluid/models/machine_translation.py @@ -197,6 +197,8 @@ def lodtensor_to_ndarray(lod_tensor): def get_model(args): + if args.use_reader_op: + raise Exception("machine_translation do not support reader op for now.") embedding_dim = 512 encoder_size = 512 decoder_size = 512 @@ -221,7 +223,7 @@ def get_model(args): train_batch_generator = paddle.batch( paddle.reader.shuffle( paddle.dataset.wmt14.train(dict_size), buf_size=1000), - batch_size=args.batch_size) + batch_size=args.batch_size * args.gpus) test_batch_generator = paddle.batch( paddle.reader.shuffle( diff --git a/benchmark/fluid/models/mnist.py b/benchmark/fluid/models/mnist.py index d264bfc12b..8e740dc689 100644 --- a/benchmark/fluid/models/mnist.py +++ b/benchmark/fluid/models/mnist.py @@ -20,6 +20,7 @@ import numpy as np import argparse import time import cProfile +import os import paddle import paddle.fluid as fluid @@ -65,19 +66,49 @@ def cnn_model(data): def get_model(args): - # Input data - images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - - # Train program - predict = cnn_model(images) - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - # Evaluator - batch_size_tensor = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size_tensor) + if args.use_reader_op: + filelist = [ + os.path.join(args.data_path, f) for f in os.listdir(args.data_path) + ] + data_file = fluid.layers.open_files( + filenames=filelist, + shapes=[[-1, 1, 28, 28], (-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int64"], + thread_num=args.gpus, + pass_num=args.pass_num) + data_file = fluid.layers.double_buffer( + fluid.layers.batch( + data_file, batch_size=args.batch_size)) + images, label = fluid.layers.read_file(data_file) + else: + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + if args.device == 'CPU' and args.cpus > 1: + places = fluid.layers.get_places(args.cpus) + pd = fluid.layers.ParallelDo(places) + with pd.do(): + predict = cnn_model(pd.read_input(images)) + label = pd.read_input(label) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + batch_acc = fluid.layers.accuracy(input=predict, label=label) + + pd.write_output(avg_cost) + pd.write_output(batch_acc) + + avg_cost, batch_acc = pd() + avg_cost = fluid.layers.mean(avg_cost) + batch_acc = fluid.layers.mean(batch_acc) + else: + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_acc = fluid.layers.accuracy(input=predict, label=label) # inference program inference_program = fluid.default_main_program().clone() @@ -88,7 +119,7 @@ def get_model(args): # Reader train_reader = paddle.batch( - paddle.dataset.mnist.train(), batch_size=args.batch_size) + paddle.dataset.mnist.train(), batch_size=args.batch_size * args.gpus) test_reader = paddle.batch( paddle.dataset.mnist.test(), batch_size=args.batch_size) return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc diff --git a/benchmark/fluid/models/resnet.py b/benchmark/fluid/models/resnet.py index 9dec8911ed..9ed1093c54 100644 --- a/benchmark/fluid/models/resnet.py +++ b/benchmark/fluid/models/resnet.py @@ -19,6 +19,7 @@ from __future__ import print_function import functools import numpy as np import time +import os import cProfile, pstats, StringIO @@ -26,6 +27,7 @@ import paddle import paddle.fluid as fluid import paddle.fluid.core as core import paddle.fluid.profiler as profiler +from recordio_converter import imagenet_train, imagenet_test def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'): @@ -122,40 +124,85 @@ def get_model(args): else: dshape = [32, 32, 3] model = resnet_cifar10 - else: + train_reader = paddle.dataset.cifar.train10() + test_reader = paddle.dataset.cifar.test10() + elif args.data_set == "flowers": class_dim = 102 if args.data_format == 'NCHW': dshape = [3, 224, 224] else: dshape = [224, 224, 3] model = resnet_imagenet - - input = fluid.layers.data(name='data', shape=dshape, dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - predict = model(input, class_dim) - cost = fluid.layers.cross_entropy(input=predict, label=label) - avg_cost = fluid.layers.mean(x=cost) - - batch_size_tensor = fluid.layers.create_tensor(dtype='int64') - batch_acc = fluid.layers.accuracy( - input=predict, label=label, total=batch_size_tensor) + train_reader = paddle.dataset.flowers.train() + test_reader = paddle.dataset.flowers.test() + elif args.data_set == "imagenet": + class_dim = 1000 + if args.data_format == 'NCHW': + dshape = [3, 224, 224] + else: + dshape = [224, 224, 3] + model = resnet_imagenet + if not args.data_path: + raise Exception( + "Must specify --data_path when training with imagenet") + train_reader = imagenet_train(args.data_path) + test_reader = imagenet_test(args.data_path) + + if args.use_reader_op: + filelist = [ + os.path.join(args.data_path, f) for f in os.listdir(args.data_path) + ] + data_file = fluid.layers.open_files( + filenames=filelist, + shapes=[[-1] + dshape, (-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int64"], + thread_num=args.gpus, + pass_num=args.pass_num) + data_file = fluid.layers.double_buffer( + fluid.layers.batch( + data_file, batch_size=args.batch_size)) + input, label = fluid.layers.read_file(data_file) + else: + input = fluid.layers.data(name='data', shape=dshape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + if args.device == 'CPU' and args.cpus > 1: + places = fluid.layers.get_places(args.cpus) + pd = fluid.layers.ParallelDo(places) + with pd.do(): + predict = model(pd.read_input(input), class_dim) + label = pd.read_input(label) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + batch_acc = fluid.layers.accuracy(input=predict, label=label) + + pd.write_output(avg_cost) + pd.write_output(batch_acc) + + avg_cost, batch_acc = pd() + avg_cost = fluid.layers.mean(avg_cost) + batch_acc = fluid.layers.mean(batch_acc) + else: + predict = model(input, class_dim) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + batch_acc = fluid.layers.accuracy(input=predict, label=label) inference_program = fluid.default_main_program().clone() with fluid.program_guard(inference_program): inference_program = fluid.io.get_inference_program( - target_vars=[batch_acc, batch_size_tensor]) + target_vars=[batch_acc]) optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9) - train_reader = paddle.batch( + batched_train_reader = paddle.batch( paddle.reader.shuffle( - paddle.dataset.cifar.train10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), - buf_size=5120), - batch_size=args.batch_size) - test_reader = paddle.batch( - paddle.dataset.cifar.test10() - if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), - batch_size=args.batch_size) - - return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc + train_reader, buf_size=5120), + batch_size=args.batch_size * args.gpus, + drop_last=True) + batched_test_reader = paddle.batch( + train_reader, batch_size=args.batch_size, drop_last=True) + + return avg_cost, inference_program, optimizer, batched_train_reader,\ + batched_test_reader, batch_acc diff --git a/benchmark/fluid/models/stacked_dynamic_lstm.py b/benchmark/fluid/models/stacked_dynamic_lstm.py index 81a28b5f3a..211869af4e 100644 --- a/benchmark/fluid/models/stacked_dynamic_lstm.py +++ b/benchmark/fluid/models/stacked_dynamic_lstm.py @@ -44,6 +44,9 @@ def crop_sentence(reader, crop_size): def get_model(args): + if args.use_reader_op: + raise Exception( + "stacked_dynamic_lstm do not support reader op for now.") lstm_size = 512 emb_dim = 512 crop_size = 1500 @@ -115,7 +118,7 @@ def get_model(args): train_reader = batch( paddle.reader.shuffle( crop_sentence(imdb.train(word_dict), crop_size), buf_size=25000), - batch_size=args.batch_size) + batch_size=args.batch_size * args.gpus) test_reader = batch( paddle.reader.shuffle( crop_sentence(imdb.test(word_dict), crop_size), buf_size=25000), diff --git a/benchmark/fluid/models/vgg.py b/benchmark/fluid/models/vgg.py index 53856c5f7a..932601302d 100644 --- a/benchmark/fluid/models/vgg.py +++ b/benchmark/fluid/models/vgg.py @@ -22,6 +22,7 @@ import paddle.fluid as fluid import paddle.fluid.core as core import argparse import functools +import os def vgg16_bn_drop(input): @@ -65,9 +66,25 @@ def get_model(args): else: data_shape = [224, 224, 3] - # Input data - images = fluid.layers.data(name='pixel', shape=data_shape, dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') + if args.use_reader_op: + filelist = [ + os.path.join(args.data_path, f) for f in os.listdir(args.data_path) + ] + data_file = fluid.layers.open_files( + filenames=filelist, + shapes=[[-1] + data_shape, (-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int64"], + thread_num=args.gpus, + pass_num=args.pass_num) + data_file = fluid.layers.double_buffer( + fluid.layers.batch( + data_file, batch_size=args.batch_size)) + images, label = fluid.layers.read_file(data_file) + else: + images = fluid.layers.data( + name='data', shape=data_shape, dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') # Train program net = vgg16_bn_drop(images) @@ -95,7 +112,7 @@ def get_model(args): paddle.dataset.cifar.train10() if args.data_set == 'cifar10' else paddle.dataset.flowers.train(), buf_size=5120), - batch_size=args.batch_size) + batch_size=args.batch_size * args.gpus) test_reader = paddle.batch( paddle.dataset.cifar.test10() if args.data_set == 'cifar10' else paddle.dataset.flowers.test(), diff --git a/benchmark/fluid/recordio_converter.py b/benchmark/fluid/recordio_converter.py new file mode 100644 index 0000000000..f2dc39109b --- /dev/null +++ b/benchmark/fluid/recordio_converter.py @@ -0,0 +1,164 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import random +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.dataset import mnist, cifar, flowers, image + + +def convert_2_recordio(py_reader, outfilepath, batch_size, shape_data, + shape_label): + num_batches = 0 + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch(py_reader(), batch_size=batch_size) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=shape_data), + fluid.layers.data( + name='label', shape=shape_label, dtype='int64'), + ], + place=fluid.CPUPlace()) + num_batches = fluid.recordio_writer.convert_reader_to_recordio_file( + outfilepath, reader, feeder) + return num_batches + + +def prepare_mnist(outpath, batch_size): + outfilepath = os.path.join(outpath, "mnist.recordio") + convert_2_recordio(mnist.train, outfilepath, batch_size, [784], [1]) + + +def prepare_cifar10(outpath, batch_size): + outfilepath = os.path.join(outpath, "cifar.recordio") + convert_2_recordio(cifar.train10, outfilepath, batch_size, [3, 32, 32], [1]) + + +def prepare_flowers(outpath, batch_size): + outfilepath = os.path.join(outpath, "flowers.recordio") + convert_2_recordio(flowers.train, outfilepath, batch_size, [3, 224, 224], + [1]) + + +def default_mapper(sample): + img, label = sample + img = image.simple_transform( + img, 256, 224, True, mean=[103.94, 116.78, 123.68]) + return img.flatten().astype('float32'), label + + +def imagenet_train(data_dir): + contents = os.listdir(data_dir) + if set(contents) != set( + ["train", "train.txt", "val", "val_set", "val.txt", "unzip.sh"]): + raise Exception("Imagenet data contents error!") + img2label = dict() + imgfilelist = [] + with open(os.path.join(data_dir, "train.txt")) as fn: + while 1: + l = fn.readline() + if not l: + break + img, lbl = l[:-1].split(" ") + img2label[img] = int(lbl) + imgfilelist.append(img) + # shuffle all, this is slow + random.shuffle(imgfilelist) + + def train_reader(): + for idx, imgfile in enumerate(imgfilelist): + data = image.load_image( + os.path.join(data_dir, "train", imgfile.lower())) + label = [img2label[imgfile], ] + yield [data, label] + + return paddle.reader.map_readers(default_mapper, train_reader) + + +def imagenet_test(data_dir): + contents = os.listdir(data_dir) + if set(contents) != set( + ["train", "train.txt", "val", "val_set", "val.txt", "unzip.sh"]): + raise Exception("Imagenet data contents error!") + img2label = dict() + imgfilelist = [] + with open(os.path.join(data_dir, "val.txt")) as fn: + while 1: + l = fn.readline() + if not l: + break + img, lbl = l[:-1].split(" ") + img2label[img] = int(lbl) + imgfilelist.append(img) + + def test_reader(): + for idx, imgfile in enumerate(imgfilelist): + base_path = os.path.join(data_dir, "val", imgfile.split(".")[0]) + image_path = ".".join([base_path, "jpeg"]) + data = image.load_image(image_path) + label = [img2label[imgfile], ] + yield [data, label] + + return paddle.reader.map_readers(default_mapper, test_reader) + + +# FIXME(wuyi): delete this when https://github.com/PaddlePaddle/Paddle/pull/11066 is merged +def convert_reader_to_recordio_files( + filename, + batch_per_file, + reader_creator, + feeder, + compressor=core.RecordIOWriter.Compressor.Snappy, + max_num_records=1000, + feed_order=None): + if feed_order is None: + feed_order = feeder.feed_names + f_name, f_ext = os.path.splitext(filename) + assert (f_ext == ".recordio") + + lines = [] + f_idx = 0 + counter = 0 + for idx, batch in enumerate(reader_creator()): + lines.append(batch) + if idx >= batch_per_file and idx % batch_per_file == 0: + filename = "%s-%05d%s" % (f_name, f_idx, f_ext) + with fluid.recordio_writer.create_recordio_writer( + filename, compressor, max_num_records) as writer: + for l in lines: + res = feeder.feed(l) + for each in feed_order: + writer.append_tensor(res[each]) + writer.complete_append_tensor() + counter += 1 + lines = [] + f_idx += 1 + print("written file: ", filename) + return counter + + +def prepare_imagenet(inpath, outpath, batch_size): + r = paddle.batch(imagenet_train(inpath), batch_size=batch_size) + feeder = fluid.DataFeeder( + feed_list=[ + fluid.layers.data( + name="image", shape=[3, 224, 224]), fluid.layers.data( + name="label", shape=[1], dtype='int64') + ], + place=fluid.CPUPlace()) + outpath = os.path.join(outpath, "imagenet.recordio") + convert_reader_to_recordio_files(outpath, 10000, r, feeder) diff --git a/benchmark/fluid/run.sh b/benchmark/fluid/run.sh index f6dfd20bf2..5d9b2db871 100644 --- a/benchmark/fluid/run.sh +++ b/benchmark/fluid/run.sh @@ -2,6 +2,7 @@ # This script benchmarking the PaddlePaddle Fluid on # single thread single GPU. +mkdir -p logs #export FLAGS_fraction_of_gpu_memory_to_use=0.0 export CUDNN_PATH=/paddle/cudnn_v5 @@ -35,71 +36,74 @@ nohup stdbuf -oL nvidia-smi \ --format=csv \ --filename=mem.log \ -l 1 & + # mnist # mnist gpu mnist 128 -FLAGS_benchmark=true stdbuf -oL python fluid/mnist.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=mnist \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ --iterations=500 \ - 2>&1 | tee -a mnist_gpu_128.log + 2>&1 | tee -a logs/mnist_gpu_128.log # vgg16 # gpu cifar10 128 -FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=vgg16 \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a vgg16_gpu_128.log + 2>&1 | tee -a logs/vgg16_gpu_128.log # flowers gpu 128 -FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=vgg16 \ --device=GPU \ --batch_size=32 \ --data_set=flowers \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a vgg16_gpu_flowers_32.log + 2>&1 | tee -a logs/vgg16_gpu_flowers_32.log # resnet50 # resnet50 gpu cifar10 128 -FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=resnet \ --device=GPU \ --batch_size=128 \ --data_set=cifar10 \ - --model=resnet_cifar10 \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a resnet50_gpu_128.log + 2>&1 | tee -a logs/resnet50_gpu_128.log # resnet50 gpu flowers 64 -FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=resnet \ --device=GPU \ --batch_size=64 \ --data_set=flowers \ - --model=resnet_imagenet \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a resnet50_gpu_flowers_64.log + 2>&1 | tee -a logs/resnet50_gpu_flowers_64.log # lstm # lstm gpu imdb 32 # tensorflow only support batch=32 -FLAGS_benchmark=true stdbuf -oL python fluid/stacked_dynamic_lstm.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=stacked_dynamic_lstm \ --device=GPU \ --batch_size=32 \ --skip_batch_num=5 \ --iterations=30 \ - --hidden_dim=512 \ - --emb_dim=512 \ - --crop_size=1500 \ - 2>&1 | tee -a lstm_gpu_32.log + 2>&1 | tee -a logs/lstm_gpu_32.log # seq2seq # seq2seq gpu wmb 128 -FLAGS_benchmark=true stdbuf -oL python fluid/machine_translation.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=machine_translation \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ --iterations=30 \ - 2>&1 | tee -a lstm_gpu_128.log + 2>&1 | tee -a logs/lstm_gpu_128.log diff --git a/benchmark/fluid/run_fluid_benchmark.sh b/benchmark/fluid/run_fluid_benchmark.sh new file mode 100644 index 0000000000..4309a3126c --- /dev/null +++ b/benchmark/fluid/run_fluid_benchmark.sh @@ -0,0 +1,9 @@ +#!/bin/bash + +PADDLE_TRAINING_ROLE=PSERVER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model resnet --device CPU --update_method pserver --iterations=10000 & + +sleep 15 + +CUDA_VISIBLE_DEVICES=0,1 PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=0 python fluid_benchmark.py --model resnet --device GPU --update_method pserver --iterations=10000 --gpus 2 & + +CUDA_VISIBLE_DEVICES=2,3 PADDLE_TRAINING_ROLE=TRAINER PADDLE_PSERVER_PORT=7164 PADDLE_PSERVER_IPS=127.0.0.1 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=127.0.0.1 PADDLE_TRAINER_ID=1 python fluid_benchmark.py --model resnet --device GPU --update_method pserver --iterations=10000 --gpus 2 & diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 682614742c..6a8b15a6b6 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -92,6 +92,9 @@ if(WITH_GPU) if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7) message(FATAL_ERROR "TensorRT needs CUDNN >= 7.0 to compile") endif() + if(${TENSORRT_MAJOR_VERSION} VERSION_LESS 4) + message(FATAL_ERROR "Paddle needs TensorRT >= 4.0 to compile") + endif() include_directories(${TENSORRT_INCLUDE_DIR}) endif() elseif(WITH_AMD_GPU) @@ -115,6 +118,10 @@ endif() set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}") +if(WITH_DISTRIBUTE) + add_definitions(-DPADDLE_WITH_DISTRIBUTE) +endif() + if(WITH_GOLANG) # we need to symlink Paddle directory into GOPATH. If we # don't do it and we have code that depends on Paddle, go @@ -163,3 +170,7 @@ if(WITH_GOLANG) endif() endif(WITH_GOLANG) + +if(WITH_GRPC) + add_definitions(-DPADDLE_WITH_GRPC) +endif(WITH_GRPC) diff --git a/cmake/external/anakin.cmake b/cmake/external/anakin.cmake new file mode 100644 index 0000000000..f1cd9c99eb --- /dev/null +++ b/cmake/external/anakin.cmake @@ -0,0 +1,42 @@ +if (NOT WITH_ANAKIN) + return() +endif() + +set(ANAKIN_INSTALL_DIR "${THIRD_PARTY_PATH}/install/anakin" CACHE PATH + "Anakin install path." FORCE) +set(ANAKIN_INCLUDE "${ANAKIN_INSTALL_DIR}" CACHE STRING "root of Anakin header files") +set(ANAKIN_LIBRARY "${ANAKIN_INSTALL_DIR}" CACHE STRING "path of Anakin library") + +set(ANAKIN_COMPILE_EXTRA_FLAGS -Wno-error=unused-variable -Wno-error=format-extra-args -Wno-error=comment -Wno-error=format -Wno-error=switch -Wno-error=return-type -Wno-error=non-virtual-dtor -Wno-reorder -Wno-error=cpp) + +set(ANAKIN_LIBRARY_URL "https://github.com/pangge/Anakin/releases/download/3.0/anakin_release_simple.tar.gz") + +# A helper function used in Anakin, currently, to use it, one need to recursively include +# nearly all the header files. +function(fetch_include_recursively root_dir) + if (IS_DIRECTORY ${root_dir}) + include_directories(${root_dir}) + endif() + + file(GLOB ALL_SUB RELATIVE ${root_dir} ${root_dir}/*) + foreach(sub ${ALL_SUB}) + if (IS_DIRECTORY ${root_dir}/${sub}) + fetch_include_recursively(${root_dir}/${sub}) + endif() + endforeach() +endfunction() + +# download library +message(STATUS "Download Anakin library from ${ANAKIN_LIBRARY_URL}") +execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}") +execute_process(COMMAND bash -c "rm -rf ${ANAKIN_INSTALL_DIR}/*") +execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget -q ${ANAKIN_LIBRARY_URL}") +execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}") +execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz") + +if (WITH_ANAKIN) + message(STATUS "Anakin for inference is enabled") + message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}") + fetch_include_recursively(${ANAKIN_INCLUDE}) + link_directories(${ANAKIN_LIBRARY}) +endif() diff --git a/cmake/external/brpc.cmake b/cmake/external/brpc.cmake new file mode 100644 index 0000000000..8e2c913b2c --- /dev/null +++ b/cmake/external/brpc.cmake @@ -0,0 +1,58 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +INCLUDE(ExternalProject) + +SET(BRPC_SOURCES_DIR ${THIRD_PARTY_PATH}/brpc) +SET(BRPC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/brpc) +SET(BRPC_INCLUDE_DIR "${BRPC_INSTALL_DIR}/include" CACHE PATH "brpc include directory." FORCE) +SET(BRPC_LIBRARIES "${BRPC_INSTALL_DIR}/lib/libbrpc.a" CACHE FILEPATH "brpc library." FORCE) + +INCLUDE_DIRECTORIES(${BRPC_INCLUDE_DIR}) + +# Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args +set(prefix_path "${THIRD_PARTY_PATH}/install/gflags|${THIRD_PARTY_PATH}/install/leveldb|${THIRD_PARTY_PATH}/install/snappy|${THIRD_PARTY_PATH}/install/gtest|${THIRD_PARTY_PATH}/install/protobuf") + +# If minimal .a is need, you can set WITH_DEBUG_SYMBOLS=OFF +ExternalProject_Add( + extern_brpc + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/brpc/brpc" + GIT_TAG "6d153dd7ff00f960ae6895c9c5fff0ce9f07aff2" + PREFIX ${BRPC_SOURCES_DIR} + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS} + -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS} + -DCMAKE_INSTALL_PREFIX=${BRPC_INSTALL_DIR} + -DCMAKE_INSTALL_LIBDIR=${BRPC_INSTALL_DIR}/lib + -DCMAKE_POSITION_INDEPENDENT_CODE=ON + -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE} + -DCMAKE_PREFIX_PATH=${prefix_path} + -DBRPC_WITH_GLOG=ON + ${EXTERNAL_OPTIONAL_ARGS} + LIST_SEPARATOR | + CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${BRPC_INSTALL_DIR} + -DCMAKE_INSTALL_LIBDIR:PATH=${BRPC_INSTALL_DIR}/lib + -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON + -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE} +) +ADD_DEPENDENCIES(extern_brpc protobuf leveldb gflags glog gtest snappy) +ADD_LIBRARY(brpc STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES}) +ADD_DEPENDENCIES(brpc extern_brpc) + + +LIST(APPEND external_project_dependencies brpc) diff --git a/cmake/external/grpc.cmake b/cmake/external/grpc.cmake index 9459f1ddfe..ffdf91a354 100644 --- a/cmake/external/grpc.cmake +++ b/cmake/external/grpc.cmake @@ -33,10 +33,19 @@ ELSE() SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin) ENDIF() +# FIXME(wuyi): do not build zlib cares protobuf twice, find a way to build grpc with them ExternalProject_Add( extern_grpc DEPENDS protobuf zlib - URL "http://paddlepaddledeps.bj.bcebos.com/grpc.tar.xz" + # NOTE(wuyi): + # this package is generated by following steps: + # 1. git clone -b v1.8.x https://github.com/grpc/grpc.git + # 2. submodule update --init + # 3. keep only zlib, cares, protobuf, boringssl under "third_party", + # checkout and clean other dirs under third_party + # 4. remove .git, and package the directory. + URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.8.x.tar.gz" + URL_MD5 "c9c58ee7d0e8929a63155af6a2ecdbd0" PREFIX ${GRPC_SOURCES_DIR} UPDATE_COMMAND "" CONFIGURE_COMMAND "" @@ -49,7 +58,6 @@ ExternalProject_Add( INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install ) -# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them. ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION "${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a") diff --git a/cmake/external/leveldb.cmake b/cmake/external/leveldb.cmake new file mode 100644 index 0000000000..fb5091731d --- /dev/null +++ b/cmake/external/leveldb.cmake @@ -0,0 +1,44 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +INCLUDE(ExternalProject) + +SET(LEVELDB_SOURCES_DIR ${THIRD_PARTY_PATH}/leveldb) +SET(LEVELDB_INSTALL_DIR ${THIRD_PARTY_PATH}/install/leveldb) +SET(LEVELDB_INCLUDE_DIR "${LEVELDB_INSTALL_DIR}/include" CACHE PATH "leveldb include directory." FORCE) +SET(LEVELDB_LIBRARIES "${LEVELDB_INSTALL_DIR}/lib/libleveldb.a" CACHE FILEPATH "leveldb library." FORCE) +INCLUDE_DIRECTORIES(${LEVELDB_INCLUDE_DIR}) + +ExternalProject_Add( + extern_leveldb + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${LEVELDB_SOURCES_DIR} + URL "https://github.com/google/leveldb/archive/v1.18.tar.gz" + URL_MD5 "73770de34a2a5ab34498d2e05b2b7fa0" + CONFIGURE_COMMAND "" + BUILD_COMMAND CXXFLAGS=-fPIC make -j ${NUM_OF_PROCESSOR} libleveldb.a + INSTALL_COMMAND mkdir -p ${LEVELDB_INSTALL_DIR}/lib/ + && cp ${LEVELDB_SOURCES_DIR}/src/extern_leveldb/libleveldb.a ${LEVELDB_LIBRARIES} + && cp -r ${LEVELDB_SOURCES_DIR}/src/extern_leveldb/include ${LEVELDB_INSTALL_DIR}/ + BUILD_IN_SOURCE 1 +) + +ADD_DEPENDENCIES(extern_leveldb snappy) + +ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL) +SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES}) +ADD_DEPENDENCIES(leveldb extern_leveldb) + +LIST(APPEND external_project_dependencies leveldb) + diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 8af2765f58..4a49a92f2b 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -29,6 +29,8 @@ IF(NOT ${CBLAS_FOUND}) "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE FILEPATH "openblas library." FORCE) + ADD_DEFINITIONS(-DPADDLE_USE_OPENBLAS) + SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable") SET(OPENBLAS_COMMIT "v0.2.20") diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 9ddd05b3d9..0e2df86c19 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -610,3 +610,21 @@ function(grpc_library TARGET_NAME) COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") cc_library("${TARGET_NAME}" SRCS "${grpc_library_SRCS}" DEPS "${TARGET_NAME}_grpc" "${TARGET_NAME}_proto" "${grpc_library_DEPS}") endfunction() + + +function(brpc_library TARGET_NAME) + set(oneValueArgs PROTO) + set(multiValueArgs SRCS DEPS) + set(options "") + cmake_parse_arguments(brpc_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + message(STATUS "generating brpc ${brpc_library_PROTO}") + + get_filename_component(ABS_PROTO ${brpc_library_PROTO} ABSOLUTE) + get_filename_component(PROTO_WE ${brpc_library_PROTO} NAME_WE) + get_filename_component(PROTO_PATH ${ABS_PROTO} PATH) + + protobuf_generate_cpp(brpc_proto_srcs brpc_proto_hdrs "${ABS_PROTO}") + cc_library("${TARGET_NAME}_proto" SRCS "${brpc_proto_srcs}") + cc_library("${TARGET_NAME}" SRCS "${brpc_library_SRCS}" DEPS "${TARGET_NAME}_proto" "${brpc_library_DEPS}") +endfunction() diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 236a55d332..cd44fe2542 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -39,7 +39,7 @@ function(copy TARGET) message(FATAL_ERROR "${TARGET} source numbers are not equal to destination numbers") endif() math(EXPR len "${copy_lib_SRCS_len} - 1") - + add_custom_target(${TARGET} DEPENDS ${copy_lib_DEPS}) foreach(index RANGE ${len}) list(GET copy_lib_SRCS ${index} src) @@ -155,6 +155,15 @@ copy(inference_lib DEPS paddle_fluid_shared paddle_fluid DSTS ${dst_dir}/${module} ${dst_dir}/${module} ) +if(WITH_CONTRIB) + set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference") + copy(contrib_inference_lib DEPS paddle_inference_api + SRCS ${PADDLE_SOURCE_DIR}/paddle/contrib/inference/paddle_inference_api.h + ${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api.* + DSTS ${contrib_dst_dir} ${contrib_dst_dir} + ) +endif() + set(module "platform") copy(platform_lib DEPS profiler_py_proto SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h diff --git a/doc/fluid/api/detection.rst b/doc/fluid/api/detection.rst new file mode 100644 index 0000000000..e69de29bb2 diff --git a/doc/fluid/api/gen_doc.sh b/doc/fluid/api/gen_doc.sh index 0f05393555..27f2419c06 100755 --- a/doc/fluid/api/gen_doc.sh +++ b/doc/fluid/api/gen_doc.sh @@ -1,5 +1,5 @@ #!/bin/bash -python gen_doc.py layers --submodules control_flow device io nn ops tensor > layers.rst +python gen_doc.py layers --submodules control_flow device io nn ops tensor detection learning_rate_scheduler > layers.rst for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer do diff --git a/doc/fluid/api/layers.rst b/doc/fluid/api/layers.rst index f53da4d194..8d1c9247b1 100644 --- a/doc/fluid/api/layers.rst +++ b/doc/fluid/api/layers.rst @@ -342,6 +342,12 @@ conv2d .. autofunction:: paddle.fluid.layers.conv2d :noindex: +conv3d +------ + +.. autofunction:: paddle.fluid.layers.conv3d + :noindex: + sequence_pool ------------- @@ -366,6 +372,12 @@ pool2d .. autofunction:: paddle.fluid.layers.pool2d :noindex: +pool3d +------ + +.. autofunction:: paddle.fluid.layers.pool3d + :noindex: + batch_norm ---------- @@ -384,6 +396,13 @@ conv2d_transpose .. autofunction:: paddle.fluid.layers.conv2d_transpose :noindex: +conv3d_transpose +---------------- + +.. autofunction:: paddle.fluid.layers.conv2d_transpose + :noindex: + + sequence_expand --------------- @@ -594,7 +613,6 @@ roi_pool .. autofunction:: paddle.fluid.layers.roi_pool :noindex: - ops === @@ -991,21 +1009,93 @@ zeros .. autofunction:: paddle.fluid.layers.zeros :noindex: -topk ----- +detection +========= -.. autofunction:: paddle.fluid.layers.topk +multi_box_head +-------------- + +.. autofunction:: paddle.fluid.layers.multi_box_head :noindex: -dice_loss ----- +bipartite_match +--------------- + +.. autofunction:: paddle.fluid.layers.bipartite_match + :noindex: + +target_assign +------------- + +.. autofunction:: paddle.fluid.layers.target_assign + :noindex: + +detection_output +---------------- + +.. autofunction:: paddle.fluid.layers.detection_output + :noindex: + +ssd_loss +-------- -.. autofunction:: paddle.fluid.layers.dice_loss +.. autofunction:: paddle.fluid.layers.ssd_loss :noindex: -upsampling_bilinear2d -____ +detection_map +------------- + +.. autofunction:: paddle.fluid.layers.detection_map + :noindex: + +iou_similarity +-------------- + +.. autofunction:: paddle.fluid.layers.iou_similarity + :noindex: + +box_coder +--------- + +.. autofunction:: paddle.fluid.layers.box_coder + :noindex: + +learning_rate_scheduler +======================= + +exponential_decay +----------------- + +.. autofunction:: paddle.fluid.layers.exponential_decay + :noindex: + +natural_exp_decay +----------------- + +.. autofunction:: paddle.fluid.layers.natural_exp_decay + :noindex: + +inverse_time_decay +------------------ + +.. autofunction:: paddle.fluid.layers.inverse_time_decay + :noindex: + +polynomial_decay +---------------- + +.. autofunction:: paddle.fluid.layers.polynomial_decay + :noindex: + +piecewise_decay +--------------- + +.. autofunction:: paddle.fluid.layers.piecewise_decay + :noindex: + +noam_decay +---------- -.. autofunction:: paddle.fluid.layers.upsampling_bilinear2d +.. autofunction:: paddle.fluid.layers.noam_decay :noindex: diff --git a/doc/fluid/api/optimizer.rst b/doc/fluid/api/optimizer.rst index df2bd2eace..79a0995fce 100644 --- a/doc/fluid/api/optimizer.rst +++ b/doc/fluid/api/optimizer.rst @@ -47,28 +47,6 @@ DecayedAdagrad :members: :noindex: -Adadelta ------------------ - -.. autoclass:: paddle.fluid.optimizer.Adadelta - :members: - :noindex: - -RMSProp ------------------ - -.. autoclass:: paddle.fluid.optimizer.RMSProp - :members: - :noindex: - -ModelAverage ------------------ - -.. autoclass:: paddle.fluid.optimizer.ModelAverage - :members: - :noindex: - - SGDOptimizer ------------ @@ -111,25 +89,24 @@ DecayedAdagradOptimizer :members: :noindex: +Adadelta +-------- -AdadeltaOptimizer ------------------ - -.. autoclass:: paddle.fluid.optimizer.AdadeltaOptimizer +.. autoclass:: paddle.fluid.optimizer.Adadelta :members: :noindex: +ModelAverage +------------ -RMSPropOptimizer ------------------ - -.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer +.. autoclass:: paddle.fluid.optimizer.ModelAverage :members: :noindex: - + Optimizer --------- .. autoclass:: paddle.fluid.optimizer.Optimizer :members: :noindex: + diff --git a/doc/fluid/design/concepts/var_desc.md b/doc/fluid/design/concepts/var_desc.md index 6750323c01..8db67f6703 100644 --- a/doc/fluid/design/concepts/var_desc.md +++ b/doc/fluid/design/concepts/var_desc.md @@ -35,7 +35,7 @@ The computation `Program` consists of nested `Blocks`. Each `Block` will consist ## Definition of VarType -A VarDesc should have a name, type and whether or not it is persistable. The are different kinds of variable types supported in PaddlePaddle, apart from the POD_Types like: `LOD_TENSOR`, `SELECTED_ROWS`, `FEED_MINIBATCH`, `FETCH_LIST`, `STEP_SCOPES`, `LOD_RANK_TABLE`, `LOD_TENSOR_ARRAY`, `PLACE_LIST`, `READER` and `CHANNEL`. These are declared inside `VarType`. A `VarDesc` then looks as the following: +A VarDesc should have a name, type and whether or not it is persistable. There are different kinds of variable types supported in PaddlePaddle, apart from the POD_Types like: `LOD_TENSOR`, `SELECTED_ROWS`, `FEED_MINIBATCH`, `FETCH_LIST`, `STEP_SCOPES`, `LOD_RANK_TABLE`, `LOD_TENSOR_ARRAY`, `PLACE_LIST`, `READER` and `CHANNEL`. These are declared inside `VarType`. A `VarDesc` then looks as the following: ```proto message VarDesc { diff --git a/doc/fluid/dev/api_doc_std_cn.md b/doc/fluid/dev/api_doc_std_cn.md index b50f18f21d..7d39b8de1e 100644 --- a/doc/fluid/dev/api_doc_std_cn.md +++ b/doc/fluid/dev/api_doc_std_cn.md @@ -1,8 +1,9 @@ # API注释撰写标准 -- [API注释模块](#API注释模块) -- [格式及示例](#格式及示例) -- [完整示例](#完整示例) +- [API注释撰写标准](#api) + - [API注释模块](#api) + - [格式及示例](#) + - [完整示例](#) ## API注释模块 @@ -217,4 +218,4 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 ## 完整示例 -fc 的完整注释见[示例](src/fc.py)。 +fc 的完整注释见[示例](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。 diff --git a/doc/fluid/dev/api_doc_std_en.md b/doc/fluid/dev/api_doc_std_en.md index e57072d52f..f175b21975 100644 --- a/doc/fluid/dev/api_doc_std_en.md +++ b/doc/fluid/dev/api_doc_std_en.md @@ -1,8 +1,9 @@ # API Doc Standard -- [API Doc Structure](#API Doc Structure) -- [Format and Examples](#Format and Examples) -- [Complete Example](#Complete Example) +- [API Doc Standard](#api-doc-standard) + - [API Doc Structure](#api-doc-structure) + - [Format and Examples](#format-and-examples) + - [Complete Example](#complete-example) ## API Doc Structure @@ -223,4 +224,4 @@ Format and examples of each part of API documantation are as follows: (take fc f ## Complete Example -Complete Example of fc please see [here](src/fc.py)。 +Complete Example of fc please see [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/src/fc.py)。 diff --git a/doc/fluid/howto/cluster/fluid_recordio.md b/doc/fluid/howto/cluster/fluid_recordio.md new file mode 100644 index 0000000000..55ce63ec19 --- /dev/null +++ b/doc/fluid/howto/cluster/fluid_recordio.md @@ -0,0 +1,127 @@ +# How to use RecordIO in Fluid + +If you want to use RecordIO as your training data format, you need to convert to your training data +to RecordIO files and reading them in the process of training, PaddlePaddle Fluid provides some +interface to deal with the RecordIO files. + +## Generate RecordIO File + +Before start training with RecordIO files, you need to convert your training data +to RecordIO format by `fluid.recordio_writer.convert_reader_to_recordio_file`, the sample codes +as follows: + +```python + reader = paddle.batch(mnist.train(), batch_size=1) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=[784]), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + fluid.recordio_writer.convert_reader_to_recordio_file('./mnist.recordio', reader, feeder) +``` + +The above code snippet would generate a RecordIO `./mnist.recordio` on your host. + +**NOTE**: we recommend users to set `batch_size=1` when generating the recordio files so that users can +adjust it flexibly while reading it. + +## Use the RecordIO file in a Local Training Job + +PaddlePaddle Fluid provides an interface `fluid.layers.io.open_recordio_file` to load your RecordIO file +and then you can use them as a Layer in your network configuration, the sample codes as follows: + +```python + data_file = fluid.layers.io.open_recordio_file( + filename="./mnist.recordio", + shapes=[(-1, 784),(-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int32"]) + data_file = fluid.layers.io.batch(data_file, batch_size=4) + + img, label = fluid.layers.io.read_file(data_file) + hidden = fluid.layers.fc(input=img, size=100, act='tanh') + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + avg_loss = fluid.layers.mean(loss) + + fluid.optimizer.Adam(learning_rate=1e-3).minimize(avg_loss) + + place = fluid.CPUPlace() + + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + avg_loss_np = [] + + # train a pass + batch_id = 0 + while True: + tmp, = exe.run(fetch_list=[avg_loss]) + + avg_loss_np.append(tmp) + print(batch_id) + batch_id += 1 +``` + +## Use the RecordIO files in Distributed Training + +1. generate multiple RecordIO files + +For a distributed training job, you may have multiple trainer nodes, +and one or more RecordIO files for one trainer node, you can use the interface +`fluid.recordio_writer.convert_reader_to_recordio_files` to convert your training data +into multiple RecordIO files, the sample codes as follows: + +```python + reader = paddle.batch(mnist.train(), batch_size=1) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=[784]), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + fluid.recordio_writer.convert_reader_to_recordio_files( + filename_suffix='./mnist.recordio', batch_per_file=100, reader, feeder) +``` + +The above codes would generate multiple RecordIO files on your host like: + +```bash +. + \_mnist-00000.recordio + |-mnist-00001.recordio + |-mnist-00002.recordio + |-mnist-00003.recordio + |-mnist-00004.recordio +``` + +2. open multiple RecordIO files by `fluid.layers.io.open_files` + +For a distributed training job, the distributed operator system will schedule trainer process on multiple nodes, +each trainer process reads parts of the whole training data, we usually take the following approach to make the training +data allocated by each trainer process as uniform as possiable: + +```python +def gen_train_list(file_pattern, trainers, trainer_id): + file_list = glob.glob(file_pattern) + ret_list = [] + for idx, f in enumerate(file_list): + if (idx + trainers) % trainers == trainer_id: + ret_list.append(f) + return ret_list + +trainers = int(os.getenv("TRAINERS")) +trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID")) +data_file = fluid.layers.io.open_files( + filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0), + thread_num=1, + shapes=[(-1, 784),(-1, 1)], + lod_levels=[0, 0], + dtypes=["float32", "int32"]) +img, label = fluid.layers.io.read_file(data_files) +... +``` diff --git a/doc/fluid/howto/index_cn.rst b/doc/fluid/howto/index_cn.rst index b7c6201797..b57af64f44 100644 --- a/doc/fluid/howto/index_cn.rst +++ b/doc/fluid/howto/index_cn.rst @@ -4,5 +4,5 @@ .. toctree:: :maxdepth: 1 + inference/index_cn.rst optimization/index_cn.rst - inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/index_en.rst b/doc/fluid/howto/index_en.rst index f3ca41cdbf..fd21e167ce 100644 --- a/doc/fluid/howto/index_en.rst +++ b/doc/fluid/howto/index_en.rst @@ -5,4 +5,3 @@ HOW TO :maxdepth: 1 optimization/index_en.rst - inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/inference/build_and_install_lib_cn.rst b/doc/fluid/howto/inference/build_and_install_lib_cn.rst new file mode 100644 index 0000000000..c8d9992fcc --- /dev/null +++ b/doc/fluid/howto/inference/build_and_install_lib_cn.rst @@ -0,0 +1,96 @@ +安装与编译C++预测库 +=========================== + +直接下载安装 +------------- + +====================== ======================================== +版本说明 C++预测库 +====================== ======================================== +cpu_avx_mkl `fluid.tgz `_ +cpu_avx_openblas `fluid.tgz `_ +cpu_noavx_openblas `fluid.tgz `_ +cuda7.5_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn7_avx_mkl `fluid.tgz `_ +====================== ======================================== + +从源码编译 +---------- +用户也可以从 PaddlePaddle 核心代码编译C++预测库，只需在编译时配制下面这些编译选项： + +================= ========= +选项 值 +================= ========= +CMAKE_BUILD_TYPE Release +FLUID_INSTALL_DIR 安装路径 +WITH_FLUID_ONLY ON（推荐） +WITH_SWIG_PY OFF（推荐 +WITH_PYTHON OFF（推荐） +WITH_GPU ON/OFF +WITH_MKL ON/OFF +================= ========= + +建议按照推荐值设置，以避免链接不必要的库。其它可选编译选项按需进行设定。 + +下面的代码片段从github拉取最新代码，配制编译选项（需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径）： + + .. code-block:: bash + + pip install paddlepaddle-gpu + PADDLE_ROOT=/path/of/capi + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + mkdir build + cd build + cmake -DFLUID_INSTALL_DIR=$PADDLE_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_FLUID_ONLY=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_PYTHON=OFF \ + -DWITH_MKL=OFF \ + -DWITH_GPU=OFF \ + .. + make + make inference_lib_dist + +成功编译后，使用C++预测库所需的依赖（包括：（1）编译出的PaddlePaddle预测库和头文件；（2）第三方链接库和头文件；（3）版本信息与编译选项信息） +均会存放于PADDLE_ROOT目录中。目录结构如下： + + .. code-block:: text + + PaddleRoot/ + ├── CMakeCache.txt + ├── paddle + │   └── fluid + │   ├── framework + │   ├── inference + │   ├── memory + │   ├── platform + │   ├── pybind + │   └── string + ├── third_party + │   ├── boost + │   │   └── boost + │   ├── eigen3 + │   │   ├── Eigen + │   │   └── unsupported + │   └── install + │   ├── gflags + │   ├── glog + │   ├── mklml + │   ├── protobuf + │   ├── snappy + │   ├── snappystream + │   └── zlib + └── version.txt + +version.txt 中记录了该预测库的版本信息，包括Git Commit ID、使用OpenBlas或MKL数学库、CUDA/CUDNN版本号，如： + + .. code-block:: text + + GIT COMMIT ID: c95cd4742f02bb009e651a00b07b21c979637dc8 + WITH_MKL: ON + WITH_GPU: ON + CUDA version: 8.0 + CUDNN version: v5 diff --git a/doc/fluid/howto/inference/index_cn.rst b/doc/fluid/howto/inference/index_cn.rst new file mode 100644 index 0000000000..a903423548 --- /dev/null +++ b/doc/fluid/howto/inference/index_cn.rst @@ -0,0 +1,8 @@ +预测库 +------------ + +.. toctree:: + :maxdepth: 1 + + build_and_install_lib_cn.rst + inference_support_in_fluid_cn.md diff --git a/doc/fluid/howto/inference/inference_support_in_fluid.md b/doc/fluid/howto/inference/inference_support_in_fluid_cn.md similarity index 90% rename from doc/fluid/howto/inference/inference_support_in_fluid.md rename to doc/fluid/howto/inference/inference_support_in_fluid_cn.md index d272cd3e3b..309b17fccd 100644 --- a/doc/fluid/howto/inference/inference_support_in_fluid.md +++ b/doc/fluid/howto/inference/inference_support_in_fluid_cn.md @@ -1,9 +1,8 @@ -# Fluid Inference使用指南 +# 使用指南 ## 目录： - Python Inference API -- 编译Fluid Inference库 - Inference C++ API - Inference实例 - Inference计算优化 @@ -55,62 +54,6 @@ return [program, feed_target_names, fetch_targets] ``` - -## 编译Fluid Inference库 - - - **不需要额外的CMake选项** - - 1、 配置CMake命令，更多配置请参考[源码编译PaddlePaddle](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/build_from_source_cn.html) - ```bash - $ git clone https://github.com/PaddlePaddle/Paddle.git - $ cd Paddle - $ mkdir build - $ cd build - $ cmake -DCMAKE_INSTALL_PREFIX=your/path/to/paddle_inference_lib \ - -DCMAKE_BUILD_TYPE=Release \ - -DWITH_PYTHON=ON \ - -DWITH_MKL=OFF \ - -DWITH_GPU=OFF \ - .. - ``` - - - 2、 编译PaddlePaddle - ```bash - $ make - ``` - - - 3、 部署。执行如下命令将PaddlePaddle Fluid Inference库部署到`your/path/to/paddle_inference_lib`目录。 - ```bash - $ make inference_lib_dist - ``` - -- 目录结构 - - ```bash - $ cd your/path/to/paddle_inference_lib - $ tree - . - |-- paddle - | `-- fluid - | |-- framework - | |-- inference - | | |-- io.h - | | `-- libpaddle_fluid.so - | |-- memory - | |-- platform - | `-- string - |-- third_party - | |-- eigen3 - | `-- install - | |-- gflags - | |-- glog - | `-- protobuf - `-- ... - ``` - - 假设`PADDLE_ROOT=your/path/to/paddle_inference_lib`。 - - - ## 链接Fluid Inference库 - 示例项目([链接](https://github.com/luotao1/fluid_inference_example.git)) diff --git a/doc/fluid/howto/optimization/benchmark/README.md b/doc/fluid/howto/optimization/benchmark/README.md deleted file mode 120000 index db30af7f53..0000000000 --- a/doc/fluid/howto/optimization/benchmark/README.md +++ /dev/null @@ -1 +0,0 @@ -../../../../../benchmark/cluster/README.md \ No newline at end of file diff --git a/doc/fluid/howto/optimization/benchmark/vgg16/README.md b/doc/fluid/howto/optimization/benchmark/vgg16/README.md deleted file mode 120000 index ca963ef5f0..0000000000 --- a/doc/fluid/howto/optimization/benchmark/vgg16/README.md +++ /dev/null @@ -1 +0,0 @@ -../../../../../../benchmark/cluster/vgg16/README.md \ No newline at end of file diff --git a/doc/fluid/howto/optimization/host_memory_profiling_cn.md b/doc/fluid/howto/optimization/host_memory_profiling_cn.md new file mode 100644 index 0000000000..9b55a66ded --- /dev/null +++ b/doc/fluid/howto/optimization/host_memory_profiling_cn.md @@ -0,0 +1,89 @@ +## 堆内存分析和优化 + +计算机程序都可能有内存泄漏的风险。**内存泄漏**一般是由于程序在堆(heap)上分配了内存而没有释放，随着程序的运行占用的内存越来越大，一方面会影响程序的稳定性，可能让运行速度越来越慢，或者造成oom，甚至会影响运行程序的机器的稳定性，造成宕机。 + + +目前有很多内存泄漏分析工具，比较经典的有[valgrind](http://valgrind.org/docs/manual/quick-start.html#quick-start.intro), [gperftools](https://gperftools.github.io/gperftools/)。 + +因为Fluid是用Python驱动C++ core来运行，valgrind直接分析非常困难，需要自己编译debug版本的、带valgrind支持的专用Python版本，而且输出的信息中大部分是Python自己的符号和调用信息，分析起来很困难，另外使用valgrind会让程序运行速度变得非常慢，所以不建议使用。 + +本教程主要介绍[gperftools](https://gperftools.github.io/gperftools/)的使用。 + +gperftool主要支持以下四个功能： + +- thread-caching malloc +- heap-checking using tcmalloc +- heap-profiling using tcmalloc +- CPU profiler + +Paddle也提供了基于gperftool的[CPU性能分析教程](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/cpu_profiling_cn.md)。 + +对于堆内存的分析，主要用到thread-caching malloc和heap-profiling using tcmalloc。 + +## 使用流程 +#### 环境 +本教程基于paddle提供的Docker开发环境paddlepaddle/paddle:latest-dev，基于Ubuntu 16.04.4 LTS环境。 + +#### 使用流程 + +- 安装google-perftools + +``` +apt-get install libunwind-dev +apt-get install google-perftools +``` + +- 安装pprof + +``` +go get -u github.com/google/pprof +``` + +- 设置运行环境 + +``` +export PPROF_PATH=/root/gopath/bin/pprof +export PPROF_BINARY_PATH=/root/gopath/bin/pprof +export LD_PRELOAD=/usr/lib/libtcmalloc.so.4 +``` + +- 使用heap profile来运行python程序。本质上是周期性的对堆的分配情况做一次快照。 + +``` +# HEAPPROFILE 设置生成的堆分析文件的目录和文件前缀 +# HEAP_PROFILE_ALLOCATION_INTERVAL 设置每分配多少存储dump一次dump，默认1GB +env HEAPPROFILE="./perf_log/test.log" HEAP_PROFILE_ALLOCATION_INTERVAL=209715200 python trainer.py +``` + +随着程序的运行，会在perf_log这个文件夹下生成很多文件，如下： + +``` +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0001.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0002.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0003.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0004.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0005.heap +-rw-r--r-- 1 root root 1.0M Jun 1 15:00 test.log.0006.heap +``` + +- 使用pprof对heap文件进行分析。分析有两种模式： + - 完整模式。会对当前heap做一个分析，显示目前分配内存一些调用路径。 + + ``` + pprof --pdf python test.log.0012.heap + ``` + 上述命令会生成一个profile00x.pdf的文件，可以直接打开，例如：[memory_cpu_allocator](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_cpu_allocator.pdf)。从下图可以看出，在CPU版本fluid的运行过程中，分配存储最多的模块式CPUAllocator. 而别的模块相对而言分配内存较少，所以被忽略了，这对于分配内存泄漏是很不方便的，因为泄漏是一个缓慢的过程，在这种图中是无法看到的。 + + ![result](https://user-images.githubusercontent.com/3048612/40964027-a54033e4-68dc-11e8-836a-144910c4bb8c.png) + + - Diff模式。可以对两个时刻的heap做diff，把一些内存分配没有发生变化的模块去掉，而把增量部分显示出来。 + ``` + pprof --pdf --base test.log.0010.heap python test.log.1045.heap + ``` + 生成的结果为：[`memory_leak_protobuf`](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_leak_protobuf.pdf) + + 从图中可以看出：ProgramDesc这个结构，在两个版本之间增长了200MB+，所以这里有很大的内存泄漏的可能性，最终结果也确实证明是这里造成了泄漏。 + + ![result](https://user-images.githubusercontent.com/3048612/40964057-b434d5e4-68dc-11e8-894b-8ab62bcf26c2.png) + ![result](https://user-images.githubusercontent.com/3048612/40964063-b7dbee44-68dc-11e8-9719-da279f86477f.png) + diff --git a/doc/mobile/cross_compiling_for_android_cn.md b/doc/mobile/cross_compiling_for_android_cn.md index cdd6917239..0607748b75 100644 --- a/doc/mobile/cross_compiling_for_android_cn.md +++ b/doc/mobile/cross_compiling_for_android_cn.md @@ -63,16 +63,16 @@ Android的Docker开发镜像向用户提供两个可配置的参数： - 编译`armeabi-v7a`，`Android API 21`的PaddlePaddle库 ```bash -$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev +$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev ./paddle/scripts/paddle_build.sh build_android ``` - 编译`arm64-v8a`，`Android API 21`的PaddlePaddle库 ```bash -$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev +$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev ./paddle/scripts/paddle_build.sh build_android ``` -执行上述`docker run`命令时，容器默认执行[paddle/scripts/docker/build_android.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。 +执行上述`docker run`命令时，容器执行[paddle/scripts/paddle_build.sh build_android](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/paddle_build.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。 ## 基于Linux交叉编译环境的编译方式 本文档将以Linux x86-64平台为例，介绍交叉编译Android平台上适用的PaddlePaddle库的方法和步骤。 diff --git a/doc/mobile/cross_compiling_for_android_en.md b/doc/mobile/cross_compiling_for_android_en.md index 6af16fc114..572063e801 100644 --- a/doc/mobile/cross_compiling_for_android_en.md +++ b/doc/mobile/cross_compiling_for_android_en.md @@ -36,7 +36,7 @@ $ docker pull docker.paddlepaddlehub.com/paddle:latest-dev-android We can run the Docker image we just created to build the inference library of PaddlePaddle for Android using the command below: ```bash -$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" paddle:dev-android +$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" paddle:dev-android ./paddle/scripts/paddle_build.sh build_android ``` The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`: @@ -70,7 +70,7 @@ The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`: The ARM-64 architecture (`arm64-v8a`) requires at least level 21 of Android API. -The default entry-point of the Docker image, [`paddle/scripts/docker/build_android.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh) generates the [Android cross-compiling standalone toolchain](https://developer.android.com/ndk/guides/standalone_toolchain.html) based on the argument: `ANDROID_ABI` or `ANDROID_API`. For information about other configuration arguments, please continue reading. +The build command, [`paddle/scripts/paddle_build.sh build_android`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/paddle_build.sh) generates the [Android cross-compiling standalone toolchain](https://developer.android.com/ndk/guides/standalone_toolchain.html) based on the argument: `ANDROID_ABI` or `ANDROID_API`. For information about other configuration arguments, please continue reading. The above command generates and outputs the inference library in `$PWD/install_android` and puts third-party libraries in `$PWD/install_android/third_party`. diff --git a/doc/survey/dynamic_graph.md b/doc/survey/dynamic_graph.md new file mode 100644 index 0000000000..6b80b014b1 --- /dev/null +++ b/doc/survey/dynamic_graph.md @@ -0,0 +1,378 @@ +# Automatic Differentiation with the Tape + +## Automatic Differentiation + +A key challenge in the field of deep learning is to automatically derive the backward pass from the forward pass described algorithmically by researchers. Such a derivation, or a transformation of the forward pass program, has been long studied before the recent prosperity of deep learning in the field known as [automatic differentiation](https://arxiv.org/pdf/1502.05767.pdf). + +## The Tape + +Given the forward pass program (usually in Python in practices), there are two strategies to derive the backward pass: + +1. from the forward pass program itself, or +1. from the execution trace of the forward pass program, which is often known as the *tape*. + +This article surveys systems that follow the latter strategy. + +## Dynamic Network + +When we train a deep learning model, the tape changes every iteration as the input data change, so we have to re-derive the backward pass every iteration. This is known as *dynamic network*. + +Deep learning systems that utilize the idea of dynamic network gained their popularities in recent years. This article surveys two representative systems: [PyTorch](https://pytorch.org/) and [DyNet](https://dynet.readthedocs.io/en/latest/). + +## An Overview + +Both frameworks record a ‘tape’ of the computation and interpreting (or run-time compiling) a transformation of the tape played back in reverse. This tape is a different kind of entity than the original program.[[link]](http://www.bcl.hamilton.ie/~barak/papers/toplas-reverse.pdf) + +Consider the following code feedforward model. + +```python +x = Variable(randn(20, 1))) +label = Variable(randint(1)) +W_1, W_2 = Variable(randn(20, 20)), Variable(randn(10, 20)) +h = matmul(W_1, x) +pred = matmul(W_2, x) +loss = softmax(pred, label) +loss.backward() +``` + +### 1) Dynet uses List to encode the Tape + +During the forward execution, a list of operators, in this case `matmul`, `matmul` and `softmax`, are recorded in the tape, along with the necessary information needed to do the backward such as pointers to the inputs and outputs. Then the tape is played in reverse order at `loss.backward()`. + +
+ +digraph g { + graph [ + rankdir = "LR" + ]; + node [ + fontsize = "16" + shape = "ellipse" + ]; + edge []; + "node0" [ + label = " type: matmul | input: W_1, x | output: h" + shape = "record" + ]; + "node1" [ + label = " type: matmul | input: W_2, h | output: pred" + shape = "record" + ]; + "node2" [ + label = " type: softmax | input: pred, label | output: loss" + shape = "record" + ]; + "node0":f0 -> "node1":f0 []; + "node1":f0 -> "node2":f0 []; +} +
+ +![Alt text](https://g.gravizo.com/svg?digraph%20g%20{%20graph%20[%20rankdir%20=%20%22LR%22%20];%20node%20[%20fontsize%20=%20%2216%22%20shape%20=%20%22ellipse%22%20];%20edge%20[];%20%22node0%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20%3Cf1%3E%20input:%20W_1,%20x%20|%20%3Cf2%3E%20output:%20h%22%20shape%20=%20%22record%22%20];%20%22node1%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20%3Cf1%3E%20input:%20W_2,%20h%20|%20%3Cf2%3E%20output:%20pred%22%20shape%20=%20%22record%22%20];%20%22node2%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20softmax%20|%20%3Cf1%3E%20input:%20pred,%20label%20|%20%3Cf2%3E%20output:%20loss%22%20shape%20=%20%22record%22%20];%20%22node0%22:f0%20-%3E%20%22node1%22:f0%20[%20id%20=%200%20];%20%22node1%22:f0%20-%3E%20%22node2%22:f0%20[%20id%20=%201%20];%20}) + +### 2) Pytorch uses Node Graph to encode the Tape + +The graph is composed of `Variable`s and `Function`s. During the forward execution, a `Variable` records its creator function, e.g. `h.creator = matmul`. And a Function records its inputs' previous/dependent functions `prev_func` through `creator`, e.g. `matmul.prev_func = matmul1`. At `loss.backward()`, a topological sort is performed on all `prev_func`s. Then the grad op is performed by the sorted order. + +
+ +digraph g { + graph [ + rankdir = "LR" + ]; + + subgraph function { + node [ + fontsize = "16" + style = filled + shape = "record" + ]; + "matmul0" [ label = " type: matmul | prev_func: None" ]; + "matmul1" [ label = " type: matmul | prev_func: matmul" ]; + "softmax" [ label = " type: softmax | prev_func: matmul" ]; + } + + subgraph variable { + node [ + fontsize = "16" + shape = "Mrecord" + style = filled + fillcolor = white + ]; + "x" [ label = " x | creator: None" ]; + "label" [ label = " label | creator: None" ]; + "W_1" [ label = " W_1 | creator: None" ]; + "W_2" [ label = " W_2 | creator: None" ]; + "h" [ label = " h | creator: None" ]; + "pred" [ label = " pred | creator: matmul" ]; + "loss" [ label = " loss | creator: softmax" ]; + } + + subgraph data_flow { + "x":f0 -> "matmul0":f0; + "W_1":f0 -> "matmul0":f0; + "matmul0":f0 -> "h":f0; + + "h":f0 -> "matmul1":f0; + "W_2":f0 -> "matmul1":f0; + "matmul1":f0 -> "pred":f0; + + "pred":f0 -> "softmax":f0; + "label":f0 -> "softmax":f0; + "softmax":f0 -> "loss":f0; + } + + subgraph prev_func { + edge [color="red", arrowsize="0.6", penwidth="1", constraint=false]; + "matmul1":f1 -> "matmul0":f0; + "softmax":f1 -> "matmul1":f0; + label = "prev_func"; + } +} +
+ +![Alt text](https://g.gravizo.com/svg?digraph%20g%20{%20graph%20[%20rankdir%20=%20%22LR%22%20];%20subgraph%20function%20{%20node%20[%20fontsize%20=%20%2216%22%20style%20=%20filled%20shape%20=%20%22record%22%20];%20%22matmul0%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20prev_func:%20None%22%20];%20%22matmul1%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20matmul%20|%20prev_func:%20matmul%22%20];%20%22softmax%22%20[%20label%20=%20%22%3Cf0%3E%20type:%20softmax%20|%20prev_func:%20matmul%22%20];%20}%20subgraph%20variable%20{%20node%20[%20fontsize%20=%20%2216%22%20shape%20=%20%22Mrecord%22%20style%20=%20filled%20fillcolor%20=%20white%20];%20%22x%22%20[%20label%20=%20%22%3Cf0%3E%20x%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22label%22%20[%20label%20=%20%22%3Cf0%3E%20label%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22W_1%22%20[%20label%20=%20%22%3Cf0%3E%20W_1%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22W_2%22%20[%20label%20=%20%22%3Cf0%3E%20W_2%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22h%22%20[%20label%20=%20%22%3Cf0%3E%20h%20|%20%3Cf1%3E%20creator:%20None%22%20];%20%22pred%22%20[%20label%20=%20%22%3Cf0%3E%20pred%20|%20%3Cf1%3E%20creator:%20matmul%22%20];%20%22loss%22%20[%20label%20=%20%22%3Cf0%3E%20loss%20|%20%3Cf1%3E%20creator:%20softmax%22%20];%20}%20subgraph%20data_flow%20{%20%22x%22:f0%20-%3E%20%22matmul0%22:f0;%20%22W_1%22:f0%20-%3E%20%22matmul0%22:f0;%20%22matmul0%22:f0%20-%3E%20%22h%22:f0;%20%22h%22:f0%20-%3E%20%22matmul1%22:f0;%20%22W_2%22:f0%20-%3E%20%22matmul1%22:f0;%20%22matmul1%22:f0%20-%3E%20%22pred%22:f0;%20%22pred%22:f0%20-%3E%20%22softmax%22:f0;%20%22label%22:f0%20-%3E%20%22softmax%22:f0;%20%22softmax%22:f0%20-%3E%20%22loss%22:f0;%20}%20subgraph%20prev_func%20{%20edge%20[color=%22red%22,%20arrowsize=%220.6%22,%20penwidth=%221%22,%20constraint=false];%20%22matmul1%22:f1%20-%3E%20%22matmul0%22:f0;%20%22softmax%22:f1%20-%3E%20%22matmul1%22:f0;%20label%20=%20%22prev_func%22;%20}%20}) + +Chainer and Autograd uses the similar techniques to record the forward pass. For details please refer to the appendix. + +## Design choices + +### 1) Dynet's List vs Pytorch's Node Graph + +What's good about List: +1. It avoids a topological sort. One only needs to traverse the list of operators in reverse and calling the corresponding backward operator. +1. It promises effient data parallelism implementations. One could count the time of usage of a certain variable during the construction list. Then in the play back, one knows the calculation of a variable has completed. This enables communication and computation overlapping. + +What's good about Node Graph: +1. More flexibility. PyTorch users can mix and match independent graphs however they like, in whatever threads they like (without explicit synchronization). An added benefit of structuring graphs this way is that when a portion of the graph becomes dead, it is automatically freed. [[2]](https://openreview.net/pdf?id=BJJsrmfCZ) Consider the following example, Pytorch only does backward on SmallNet while Dynet does both BigNet and SmallNet. +```python +result = BigNet(data) +loss = SmallNet(data) +loss.backward() +``` + +### 2) Dynet's Lazy evaluation vs Pytorch's Immediate evaluation + +Dynet builds the list in a symbolic matter. Consider the following example +```python +for epoch in range(num_epochs): + for in_words, out_label in training_data: + dy.renew_cg() + W = dy.parameter(W_p) + b = dy.parameter(b_p) + score_sym = dy.softmax(W*dy.concatenate([E[in_words[0]],E[in_words[1]]])+b) + loss_sym = dy.pickneglogsoftmax(score_sym, out_label) + loss_val = loss_sym.value() + loss_sym.backward() +``` +The computation of `lookup`, `concat`, `matmul` and `softmax` didn't happen until the call of `loss_sym.value()`. This defered execution is useful because it allows some graph-like optimization possible, e.g. kernel fusion. + +Pytorch chooses immediate evaluation. It avoids ever materializing a "forward graph"/"tape" (no need to explicitly call `dy.renew_cg()` to reset the list), recording only what is necessary to differentiate the computation, i.e. `creator` and `prev_func`. + + +## What can fluid learn from them? + +Please refer to `paddle/contrib/dynamic/`. + +# Appendix + +### Overview + +| Framework | Has Tape | Core in C++ | First Release Date | +|-----------|----------|-------------|--------------------| +| Autograd | No | No | Mar 5, 2015 | +| Chainer | No | No | Jun 5, 2015 | +| Pytorch | No | Yes | Aug 31, 2016 | +| Dynet | Yes | Yes | Oct 12, 2016 | + +### Source Code +#### Autograd +[Backward code](https://github.com/HIPS/autograd/blob/442205dfefe407beffb33550846434baa90c4de7/autograd/core.py#L8-L40). In the forward pass, a graph of VJPNode is constructed. +```python +# User API +def make_grad(fun, x): + start_node = VJPNode.new_root() + end_value, end_node = trace(start_node, fun, x) + return backward_pass(g, end_node), end_value + +# trace the forward pass by creating VJPNodes +def trace(start_node, fun, x): + with trace_stack.new_trace() as t: + start_box = new_box(x, t, start_node) + end_box = fun(start_box) + return end_box._value, end_box._node + +def backward_pass(g, end_node): + outgrads = {end_node : (g, False)} + for node in toposort(end_node): + outgrad = outgrads.pop(node) + ingrads = node.vjp(outgrad[0]) + for parent, ingrad in zip(node.parents, ingrads): + outgrads[parent] = add_outgrads(outgrads.get(parent), ingrad) + return outgrad[0] + +# Every VJPNode corresponds to a op_grad +class VJPNode(Node): + __slots__ = ['parents', 'vjp'] + def __init__(self, value, fun, args, kwargs, parent_argnums, parents): + self.parents = parents + vjpmaker = primitive_vjps[fun] + self.vjp = vjpmaker(parent_argnums, value, args, kwargs) +``` +#### Chainer +Example Code +```python +# (1) Function Set definition, creates FunctionNode +model = FunctionSet( + l1=F.Linear(784, 100), + l2=F.Linear(100, 100), + l3=F.Linear(100, 10)).to_gpu() + +# (2) Optimizer Setup +opt = optimizers.SGD() +opt.setup(model) + +# (3) Forward computation +def forward(x, t): + h1 = F.relu(model.l1(x)) + h2 = F.relu(model.l2(h1)) + y = model.l3(h2) + return F.softmax_cross_entropy(y, t) + +# (4) Training loop +for epoch in xrange(n_epoch): + for i in xrange(0, N, b_size): + x = Variable(to_gpu(...)) + t = Variable(to_gpu(...)) + opt.zero_grads() + loss = forward(x, t) + loss.backward() + opt.update() +``` +In `forward(x, t)`, a graph of [`VariableNode`](https://github.com/chainer/chainer/blob/master/chainer/variable.py#L110) and [`FunctionNode`](https://github.com/chainer/chainer/blob/a69103a4aa59d5b318f39b01dbcb858d465b89cf/chainer/function_node.py#L19) is constructed. Every output's `VariableNode.creator` is pointed to the `FunctionNode`. +```python +class FunctionNode(object): + ... + def apply(self, inputs): + outputs = self.forward(inputs) + ret = tuple([variable.Variable(y, requires_grad=requires_grad) + for y in outputs]) + # Topological ordering + self.rank = max([x.rank for x in inputs]) if input_vars else 0 + # Add backward edges + for y in ret: + y.creator_node = self + self.inputs = tuple([x.node for x in input_vars]) + self.outputs = tuple([y.node for y in ret]) + + return ret +``` +`loss.backward()` will calculate the accumulated gradient of all variables. All the backward of `FunctionNode`s will be called based on the topological order. +```python +class VariableNode(object): + ... + def backward(self, retain_grad, loss_scale): + if self.creator_node is None: + return + + cand_funcs = [] + seen_set = set() + grads = {} + + # Initialize error by 1, if this is a loss variable + if self.data.size == 1 and self._grad_var is None: + self.grad = numpy.ones_like(self.data) + grads[self._node] = self._grad_var + + def add_cand(cand): + if cand not in seen_set: + # Negate since heapq is min-heap. This is a global variable + heapq.heappush(cand_funcs, (-cand.rank, len(seen_set), cand)) + seen_set.add(cand) + + add_cand(self.creator_node) + + while cand_funcs: + _, _, func = heapq.heappop(cand_funcs) + gxs = func.backward_accumulate(func.inputs, func.outputs, func.outputs.grad) + + for x, gx in enumerate(gxs): + if x in grads: + grads[x] += gx + else: + grads[x] = gx + + if x.creator_node is not None: + add_cand(x.creator_node) +``` + +#### PyTorch +Example Code +```python +x = Variable(torch.ones(5, 5)) +y = Variable(torch.ones(5, 5) * 4) +z = x ** 2 + x * 2 + x * y + y +z.backward(torch.ones(5, 5)) +``` +The trace is done by `Variable.creator` and `Function.previous_functions`. +```python +class Variable(object): + def __init__(self, tensor, creator=None, requires_grad=True): + if creator is None: + creator = Leaf(self, requires_grad) + self.data = tensor + self.creator = creator + self._grad = None + + def backward(self, gradient=None): + if gradient is None: + if self.data.numel() != 1: + raise RuntimeError('backward should be called only on a scalar (i.e. 1-element tensor) or with gradient w.r.t. the variable') + gradient = self.data.new(1).fill_(1) + self._execution_engine.run_backward(self, gradient) + +class Function(obejct): + # ... + def _do_forward(self, *input): + unpacked_input = tuple(arg.data for arg in input) + raw_output = self.forward(*unpacked_input) + + # mark output.creator = self for backward trace + output = tuple(Variable(tensor, self) for tensor in raw_output) + + self.previous_functions = [(arg.creator, id(arg)) for arg in input] + self.output_ids = {id(var): i for i, var in enumerate(output)} + return output + + def _do_backward(self, grad_output): + return self.backwaerd(grad_output) +``` +The [backward](https://github.com/pytorch/pytorch/blob/v0.1.1/torch/autograd/engine.py) is similar to Autograd. + +#### DyNet +Example code +```python +model = dy.model() +W_p = model.add_parameters((20, 100)) +b_p = model.add_parameters(20) +E = model.add_lookup_parameters((20000, 50)) +for epoch in range(num_epochs): + for in_words, out_label in training_data: + dy.renew_cg() # init tape + W = dy.parameter(W_p) + b = dy.parameter(b_p) + score_sym = dy.softmax(W*dy.concatenate([E[in_words[0]],E[in_words[1]]])+b) + loss_sym = dy.pickneglogsoftmax(score_sym, out_label) + loss_val = loss_sym.value() + loss_sym.backward() +``` +[forward](https://github.com/clab/dynet/blob/740a9626a13a2732544de142e256ad0d0a166658/dynet/exec.cc#L84-L158), [backward](https://github.com/clab/dynet/blob/740a9626a13a2732544de142e256ad0d0a166658/dynet/exec.cc#L166-L284). The trace is done by creating a tape of expressions in every iteration. Backward is done by traverse the tape in the reverse order. +```c++ +void SimpleExecutionEngine::backward(VariableIndex from_where, bool full) { + ... + for (int i = num_nodes - 1; i >= 0; --i) { + // each node corresponds to an op + node->backward(xs, node_fx, node_dEdfx, ai, node_dEdxai); + } + ... +} +``` diff --git a/doc/v2/api/config/evaluators.rst b/doc/v2/api/config/evaluators.rst index 9ac972fb19..458d892e82 100644 --- a/doc/v2/api/config/evaluators.rst +++ b/doc/v2/api/config/evaluators.rst @@ -101,7 +101,7 @@ value_printer :noindex: Detection -===== +========== detection_map ------------- diff --git a/doc/v2/api/config/layer.rst b/doc/v2/api/config/layer.rst index 1a6496968c..5a0cfadfce 100644 --- a/doc/v2/api/config/layer.rst +++ b/doc/v2/api/config/layer.rst @@ -11,7 +11,7 @@ Data layer data ---- -.. autoclass:: paddle.v2.layer.data +.. autofunction:: paddle.v2.layer.data :noindex: Fully Connected Layers @@ -21,12 +21,12 @@ Fully Connected Layers fc -- -.. autoclass:: paddle.v2.layer.fc +.. autofunction:: paddle.v2.layer.fc :noindex: selective_fc ------------ -.. autoclass:: paddle.v2.layer.selective_fc +.. autofunction:: paddle.v2.layer.selective_fc :noindex: Conv Layers @@ -34,34 +34,34 @@ Conv Layers conv_operator ------------- -.. autoclass:: paddle.v2.layer.conv_operator +.. autofunction:: paddle.v2.layer.conv_operator :noindex: conv_projection --------------- -.. autoclass:: paddle.v2.layer.conv_projection +.. autofunction:: paddle.v2.layer.conv_projection :noindex: conv_shift ---------- -.. autoclass:: paddle.v2.layer.conv_shift +.. autofunction:: paddle.v2.layer.conv_shift :noindex: img_conv -------- -.. autoclass:: paddle.v2.layer.img_conv +.. autofunction:: paddle.v2.layer.img_conv :noindex: .. _api_v2.layer_context_projection: context_projection ------------------ -.. autoclass:: paddle.v2.layer.context_projection +.. autofunction:: paddle.v2.layer.context_projection :noindex: row_conv -------- -.. autoclass:: paddle.v2.layer.row_conv +.. autofunction:: paddle.v2.layer.row_conv :noindex: Image Pooling Layer @@ -69,27 +69,27 @@ Image Pooling Layer img_pool -------- -.. autoclass:: paddle.v2.layer.img_pool +.. autofunction:: paddle.v2.layer.img_pool :noindex: spp --- -.. autoclass:: paddle.v2.layer.spp +.. autofunction:: paddle.v2.layer.spp :noindex: maxout ------ -.. autoclass:: paddle.v2.layer.maxout +.. autofunction:: paddle.v2.layer.maxout :noindex: roi_pool -------- -.. autoclass:: paddle.v2.layer.roi_pool +.. autofunction:: paddle.v2.layer.roi_pool :noindex: pad ---- -.. autoclass:: paddle.v2.layer.pad +.. autofunction:: paddle.v2.layer.pad :noindex: Norm Layer @@ -97,27 +97,27 @@ Norm Layer img_cmrnorm ----------- -.. autoclass:: paddle.v2.layer.img_cmrnorm +.. autofunction:: paddle.v2.layer.img_cmrnorm :noindex: batch_norm ---------- -.. autoclass:: paddle.v2.layer.batch_norm +.. autofunction:: paddle.v2.layer.batch_norm :noindex: sum_to_one_norm --------------- -.. autoclass:: paddle.v2.layer.sum_to_one_norm +.. autofunction:: paddle.v2.layer.sum_to_one_norm :noindex: cross_channel_norm ------------------ -.. autoclass:: paddle.v2.layer.cross_channel_norm +.. autofunction:: paddle.v2.layer.cross_channel_norm :noindex: row_l2_norm ----------- -.. autoclass:: paddle.v2.layer.row_l2_norm +.. autofunction:: paddle.v2.layer.row_l2_norm :noindex: Recurrent Layers @@ -125,22 +125,22 @@ Recurrent Layers recurrent --------- -.. autoclass:: paddle.v2.layer.recurrent +.. autofunction:: paddle.v2.layer.recurrent :noindex: lstmemory --------- -.. autoclass:: paddle.v2.layer.lstmemory +.. autofunction:: paddle.v2.layer.lstmemory :noindex: grumemory --------- -.. autoclass:: paddle.v2.layer.grumemory +.. autofunction:: paddle.v2.layer.grumemory :noindex: gated_unit ----------- -.. autoclass:: paddle.v2.layer.gated_unit +.. autofunction:: paddle.v2.layer.gated_unit :noindex: Recurrent Layer Group @@ -148,32 +148,32 @@ Recurrent Layer Group memory ------ -.. autoclass:: paddle.v2.layer.memory +.. autofunction:: paddle.v2.layer.memory :noindex: recurrent_group --------------- -.. autoclass:: paddle.v2.layer.recurrent_group +.. autofunction:: paddle.v2.layer.recurrent_group :noindex: lstm_step --------- -.. autoclass:: paddle.v2.layer.lstm_step +.. autofunction:: paddle.v2.layer.lstm_step :noindex: gru_step -------- -.. autoclass:: paddle.v2.layer.gru_step +.. autofunction:: paddle.v2.layer.gru_step :noindex: beam_search ------------ -.. autoclass:: paddle.v2.layer.beam_search +.. autofunction:: paddle.v2.layer.beam_search :noindex: get_output ---------- -.. autoclass:: paddle.v2.layer.get_output +.. autofunction:: paddle.v2.layer.get_output :noindex: Mixed Layer @@ -183,54 +183,54 @@ Mixed Layer mixed ----- -.. autoclass:: paddle.v2.layer.mixed +.. autofunction:: paddle.v2.layer.mixed :noindex: .. _api_v2.layer_embedding: embedding --------- -.. autoclass:: paddle.v2.layer.embedding +.. autofunction:: paddle.v2.layer.embedding :noindex: scaling_projection ------------------ -.. autoclass:: paddle.v2.layer.scaling_projection +.. autofunction:: paddle.v2.layer.scaling_projection :noindex: dotmul_projection ----------------- -.. autoclass:: paddle.v2.layer.dotmul_projection +.. autofunction:: paddle.v2.layer.dotmul_projection :noindex: dotmul_operator --------------- -.. autoclass:: paddle.v2.layer.dotmul_operator +.. autofunction:: paddle.v2.layer.dotmul_operator :noindex: full_matrix_projection ---------------------- -.. autoclass:: paddle.v2.layer.full_matrix_projection +.. autofunction:: paddle.v2.layer.full_matrix_projection :noindex: identity_projection ------------------- -.. autoclass:: paddle.v2.layer.identity_projection +.. autofunction:: paddle.v2.layer.identity_projection :noindex: slice_projection ------------------- -.. autoclass:: paddle.v2.layer.slice_projection +.. autofunction:: paddle.v2.layer.slice_projection :noindex: table_projection ---------------- -.. autoclass:: paddle.v2.layer.table_projection +.. autofunction:: paddle.v2.layer.table_projection :noindex: trans_full_matrix_projection ---------------------------- -.. autoclass:: paddle.v2.layer.trans_full_matrix_projection +.. autofunction:: paddle.v2.layer.trans_full_matrix_projection :noindex: Aggregate Layers @@ -245,51 +245,46 @@ AggregateLevel pooling ------- -.. autoclass:: paddle.v2.layer.pooling +.. autofunction:: paddle.v2.layer.pooling :noindex: .. _api_v2.layer_last_seq: last_seq -------- -.. autoclass:: paddle.v2.layer.last_seq +.. autofunction:: paddle.v2.layer.last_seq :noindex: .. _api_v2.layer_first_seq: first_seq --------- -.. autoclass:: paddle.v2.layer.first_seq +.. autofunction:: paddle.v2.layer.first_seq :noindex: sub_seq --------- -.. autoclass:: paddle.v2.layer.sub_seq +.. autofunction:: paddle.v2.layer.sub_seq :noindex: concat ------ -.. autoclass:: paddle.v2.layer.concat +.. autofunction:: paddle.v2.layer.concat :noindex: seq_concat ---------- -.. autoclass:: paddle.v2.layer.seq_concat +.. autofunction:: paddle.v2.layer.seq_concat :noindex: seq_slice --------- -.. autoclass:: paddle.v2.layer.seq_slice - :noindex: - -kmax_sequence_score -------------------- -.. autoclass:: paddle.v2.layer.kmax_sequence_score +.. autofunction:: paddle.v2.layer.seq_slice :noindex: sub_nested_seq -------------- -.. autoclass:: paddle.v2.layer.sub_nested_seq +.. autofunction:: paddle.v2.layer.sub_nested_seq :noindex: Reshaping Layers @@ -297,7 +292,7 @@ Reshaping Layers block_expand ------------ -.. autoclass:: paddle.v2.layer.block_expand +.. autofunction:: paddle.v2.layer.block_expand :noindex: .. _api_v2.layer_expand: @@ -309,22 +304,22 @@ ExpandLevel expand ------ -.. autoclass:: paddle.v2.layer.expand +.. autofunction:: paddle.v2.layer.expand :noindex: repeat ------ -.. autoclass:: paddle.v2.layer.repeat +.. autofunction:: paddle.v2.layer.repeat :noindex: rotate ------ -.. autoclass:: paddle.v2.layer.rotate +.. autofunction:: paddle.v2.layer.rotate :noindex: seq_reshape ----------- -.. autoclass:: paddle.v2.layer.seq_reshape +.. autofunction:: paddle.v2.layer.seq_reshape :noindex: Math Layers @@ -332,94 +327,94 @@ Math Layers addto ----- -.. autoclass:: paddle.v2.layer.addto +.. autofunction:: paddle.v2.layer.addto :noindex: linear_comb ----------- -.. autoclass:: paddle.v2.layer.linear_comb +.. autofunction:: paddle.v2.layer.linear_comb :noindex: interpolation ------------- -.. autoclass:: paddle.v2.layer.interpolation +.. autofunction:: paddle.v2.layer.interpolation :noindex: bilinear_interp --------------- -.. autoclass:: paddle.v2.layer.bilinear_interp +.. autofunction:: paddle.v2.layer.bilinear_interp :noindex: dropout -------- -.. autoclass:: paddle.v2.layer.dropout +.. autofunction:: paddle.v2.layer.dropout :noindex: dot_prod --------- -.. autoclass:: paddle.v2.layer.dot_prod +.. autofunction:: paddle.v2.layer.dot_prod :noindex: out_prod -------- -.. autoclass:: paddle.v2.layer.out_prod +.. autofunction:: paddle.v2.layer.out_prod :noindex: power ----- -.. autoclass:: paddle.v2.layer.power +.. autofunction:: paddle.v2.layer.power :noindex: scaling ------- -.. autoclass:: paddle.v2.layer.scaling +.. autofunction:: paddle.v2.layer.scaling :noindex: clip ---- -.. autoclass:: paddle.v2.layer.clip +.. autofunction:: paddle.v2.layer.clip :noindex: resize ------ -.. autoclass:: paddle.v2.layer.resize +.. autofunction:: paddle.v2.layer.resize :noindex: slope_intercept --------------- -.. autoclass:: paddle.v2.layer.slope_intercept +.. autofunction:: paddle.v2.layer.slope_intercept :noindex: tensor ------ -.. autoclass:: paddle.v2.layer.tensor +.. autofunction:: paddle.v2.layer.tensor :noindex: .. _api_v2.layer_cos_sim: cos_sim ------- -.. autoclass:: paddle.v2.layer.cos_sim +.. autofunction:: paddle.v2.layer.cos_sim :noindex: l2_distance ----------- -.. autoclass:: paddle.v2.layer.l2_distance +.. autofunction:: paddle.v2.layer.l2_distance :noindex: trans ----- -.. autoclass:: paddle.v2.layer.trans +.. autofunction:: paddle.v2.layer.trans :noindex: scale_shift ----------- -.. autoclass:: paddle.v2.layer.scale_shift +.. autofunction:: paddle.v2.layer.scale_shift :noindex: factorization_machine --------------------- -.. autoclass:: paddle.v2.layer.factorization_machine +.. autofunction:: paddle.v2.layer.factorization_machine :noindex: Sampling Layers @@ -427,17 +422,17 @@ Sampling Layers maxid ----- -.. autoclass:: paddle.v2.layer.max_id +.. autofunction:: paddle.v2.layer.max_id :noindex: sampling_id ----------- -.. autoclass:: paddle.v2.layer.sampling_id +.. autofunction:: paddle.v2.layer.sampling_id :noindex: multiplex --------- -.. autoclass:: paddle.v2.layer.multiplex +.. autofunction:: paddle.v2.layer.multiplex :noindex: .. _api_v2.layer_costs: @@ -447,97 +442,97 @@ Cost Layers cross_entropy_cost ------------------ -.. autoclass:: paddle.v2.layer.cross_entropy_cost +.. autofunction:: paddle.v2.layer.cross_entropy_cost :noindex: cross_entropy_with_selfnorm_cost -------------------------------- -.. autoclass:: paddle.v2.layer.cross_entropy_with_selfnorm_cost +.. autofunction:: paddle.v2.layer.cross_entropy_with_selfnorm_cost :noindex: multi_binary_label_cross_entropy_cost ------------------------------------- -.. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost +.. autofunction:: paddle.v2.layer.multi_binary_label_cross_entropy_cost :noindex: classification_cost ------------------- -.. autoclass:: paddle.v2.layer.classification_cost +.. autofunction:: paddle.v2.layer.classification_cost :noindex: huber_regression_cost ------------------------- -.. autoclass:: paddle.v2.layer.huber_regression_cost +.. autofunction:: paddle.v2.layer.huber_regression_cost :noindex: huber_classification_cost ------------------------- -.. autoclass:: paddle.v2.layer.huber_classification_cost +.. autofunction:: paddle.v2.layer.huber_classification_cost :noindex: lambda_cost ----------- -.. autoclass:: paddle.v2.layer.lambda_cost +.. autofunction:: paddle.v2.layer.lambda_cost :noindex: square_error_cost ----------------- -.. autoclass:: paddle.v2.layer.square_error_cost +.. autofunction:: paddle.v2.layer.square_error_cost :noindex: rank_cost --------- -.. autoclass:: paddle.v2.layer.rank_cost +.. autofunction:: paddle.v2.layer.rank_cost :noindex: sum_cost --------- -.. autoclass:: paddle.v2.layer.sum_cost +.. autofunction:: paddle.v2.layer.sum_cost :noindex: crf --- -.. autoclass:: paddle.v2.layer.crf +.. autofunction:: paddle.v2.layer.crf :noindex: crf_decoding ------------ -.. autoclass:: paddle.v2.layer.crf_decoding +.. autofunction:: paddle.v2.layer.crf_decoding :noindex: ctc --- -.. autoclass:: paddle.v2.layer.ctc +.. autofunction:: paddle.v2.layer.ctc :noindex: warp_ctc -------- -.. autoclass:: paddle.v2.layer.warp_ctc +.. autofunction:: paddle.v2.layer.warp_ctc :noindex: nce --- -.. autoclass:: paddle.v2.layer.nce +.. autofunction:: paddle.v2.layer.nce :noindex: hsigmoid --------- -.. autoclass:: paddle.v2.layer.hsigmoid +.. autofunction:: paddle.v2.layer.hsigmoid :noindex: smooth_l1_cost -------------- -.. autoclass:: paddle.v2.layer.smooth_l1_cost +.. autofunction:: paddle.v2.layer.smooth_l1_cost :noindex: multibox_loss -------------- -.. autoclass:: paddle.v2.layer.multibox_loss +.. autofunction:: paddle.v2.layer.multibox_loss :noindex: detection_output ---------------- -.. autoclass:: paddle.v2.layer.detection_output +.. autofunction:: paddle.v2.layer.detection_output :noindex: Check Layer @@ -545,7 +540,7 @@ Check Layer eos --- -.. autoclass:: paddle.v2.layer.eos +.. autofunction:: paddle.v2.layer.eos :noindex: Activation @@ -553,5 +548,5 @@ Activation prelu -------- -.. autoclass:: paddle.v2.layer.prelu +.. autofunction:: paddle.v2.layer.prelu :noindex: diff --git a/doc/v2/api/index_en.rst b/doc/v2/api/index_en.rst index b11cd449af..70c5c524aa 100644 --- a/doc/v2/api/index_en.rst +++ b/doc/v2/api/index_en.rst @@ -8,4 +8,3 @@ API model_configs.rst data.rst run_logic.rst - fluid/index.rst diff --git a/doc/v2/build_and_install/build_from_source_cn.rst b/doc/v2/build_and_install/build_from_source_cn.rst index 741c01ce54..6421c53082 100644 --- a/doc/v2/build_and_install/build_from_source_cn.rst +++ b/doc/v2/build_and_install/build_from_source_cn.rst @@ -23,7 +23,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 在 `这里 `__ 找到 paddle_manylinux_devel 镜像的编译以及使用方法。或者参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。 -如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。 +如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 :ref:`编译依赖 <_compile_deps>` 之后才能开始编译的步骤。 编译PaddlePaddle，需要执行： @@ -106,7 +106,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 - 学习 Docker 有多难？ - 理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 + 理解 Docker 并不难，大概花十分钟看一下 `如何使用Docker `_ 。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。 - 我可以用 IDE 吗？ @@ -123,7 +123,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 - 可以并行编译吗？ - 是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 + 是的。我们的 Docker image 运行一个 `Paddle编译Bash脚本 `_ 。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。 - Docker 需要 sudo @@ -131,11 +131,11 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安 - 在 Windows/MacOS 上编译很慢 - Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。 + Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考 `如何为Windows/Mac计算机上的Docker增加内存和虚拟机 `_ 。 - 磁盘不够 - 本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。 + 本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考 `如何删除Docker Container `_ 来清理这些内容。 .. _compile_deps: @@ -195,7 +195,7 @@ BLAS PaddlePaddle支持 `MKL `_ 和 `OpenBlAS `_ 两种BLAS库。默认使用MKL。如果使用MKL并且机器含有AVX2指令集， -还会下载MKL-DNN数学库，详细参考 `这里 `_ 。 +还会下载MKL-DNN数学库，详细参考 `mkldnn设计文档 `_ 。 如果关闭MKL，则会使用OpenBLAS作为BLAS库。 @@ -211,7 +211,7 @@ PaddlePaddle可以使用cuDNN v5.1之后的任何一个版本来编译运行， 编译选项的设置 ++++++++++++++ -PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 +PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 .. code-block:: bash diff --git a/doc/v2/build_and_install/build_from_source_en.rst b/doc/v2/build_and_install/build_from_source_en.rst index b06c43e19d..b08b45d43e 100644 --- a/doc/v2/build_and_install/build_from_source_en.rst +++ b/doc/v2/build_and_install/build_from_source_en.rst @@ -11,7 +11,7 @@ To build PaddlePaddle, you need 1. A computer -- Linux, Windows, MacOS. 2. Docker. -Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. +Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. We run all the tools by running this image. .. _build_step: @@ -26,6 +26,8 @@ you can also find how to build and use paddle_manylinux_devel Docker image from `here `__ Or you can build your own image from source as the optional step below: +If you don't wish to use docker，you need to install several compile dependencies manually as :ref:`Compile Dependencies <_compile_deps>` shows to start compilation. + .. code-block:: bash # 1. clone the source code @@ -108,7 +110,7 @@ Frequently Asked Questions - How difficult is it to learn Docker? - It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have. + It takes you ten minutes to read `an introductory article `_ and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have. - Can I use my favorite IDE? @@ -125,7 +127,7 @@ Frequently Asked Questions - Does Docker do parallel building? - Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores. + Our building Docker image runs a `Bash script `_ , which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores. - Docker requires sudo @@ -133,11 +135,11 @@ Frequently Asked Questions - Docker on Windows/MacOS builds slowly - On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details. + On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to `this issue `_ for details. - Not enough disk space - Examples in this article use option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/). + Examples in this article use option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to `this article `_ . .. _compile_deps: diff --git a/doc/v2/build_and_install/pip_install_cn.rst b/doc/v2/build_and_install/pip_install_cn.rst index 853bdb21bb..095da19cd4 100644 --- a/doc/v2/build_and_install/pip_install_cn.rst +++ b/doc/v2/build_and_install/pip_install_cn.rst @@ -60,6 +60,7 @@ paddlepaddle-gpu==0.11.0 使用CUDA 7.5和cuDNN 5编译的0.11.0版 "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `_" "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" + "cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" .. _pip_dependency: diff --git a/doc/v2/build_and_install/pip_install_en.rst b/doc/v2/build_and_install/pip_install_en.rst index fecf6d3712..8406e4aa1f 100644 --- a/doc/v2/build_and_install/pip_install_en.rst +++ b/doc/v2/build_and_install/pip_install_en.rst @@ -63,6 +63,7 @@ If the links below shows up the login form, just click "Log in as guest" to star "cpu_noavx_openblas", "`paddlepaddle-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle-latest-cp27-cp27m-linux_x86_64.whl `__" "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" + "cuda9.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-latest-cp27-cp27mu-linux_x86_64.whl `__", "`paddlepaddle_gpu-latest-cp27-cp27m-linux_x86_64.whl `__" .. _pip_dependency: diff --git a/doc/v2/dev/contribute_to_paddle_cn.md b/doc/v2/dev/contribute_to_paddle_cn.md index add06e42f1..3244eedf91 100644 --- a/doc/v2/dev/contribute_to_paddle_cn.md +++ b/doc/v2/dev/contribute_to_paddle_cn.md @@ -104,7 +104,7 @@ no changes added to commit (use "git add" and/or "git commit -a") ➜ docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest" ``` -关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。 +关于构建和测试的更多信息，请参见[使用Docker安装运行](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/v2/build_and_install/docker_install_cn.rst)。 ## 提交（commit） diff --git a/paddle/contrib/CMakeLists.txt b/paddle/contrib/CMakeLists.txt index 4b19256ef4..70e3a0583d 100644 --- a/paddle/contrib/CMakeLists.txt +++ b/paddle/contrib/CMakeLists.txt @@ -14,3 +14,4 @@ # add_subdirectory(inference) +add_subdirectory(tape) diff --git a/paddle/contrib/inference/CMakeLists.txt b/paddle/contrib/inference/CMakeLists.txt index 9c55f189bc..0f56d648b1 100644 --- a/paddle/contrib/inference/CMakeLists.txt +++ b/paddle/contrib/inference/CMakeLists.txt @@ -17,46 +17,52 @@ if(APPLE) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") endif(APPLE) -function(inference_api_test TARGET_NAME TEST_SRC) - set(options "") - set(oneValueArgs "") - set(multiValueArgs ARGS) - cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - - set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) - set(arg_list "") - if(inference_test_ARGS) - foreach(arg ${inference_test_ARGS}) - list(APPEND arg_list "_${arg}") - endforeach() - else() - list(APPEND arg_list "_") - endif() - foreach(arg ${arg_list}) - string(REGEX REPLACE "^_$" "" arg "${arg}") + +set(inference_deps paddle_inference_api paddle_fluid_api) + +function(inference_api_test TARGET_NAME) + if (WITH_TESTING) + set(options "") + set(oneValueArgs "") + set(multiValueArgs ARGS) + cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + + set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) cc_test(${TARGET_NAME} - SRCS ${TEST_SRC} - DEPS paddle_fluid_api paddle_inference_api paddle_inference_api_impl + SRCS ${TARGET_NAME}.cc + DEPS "${inference_deps}" ARGS --dirname=${PYTHON_TESTS_DIR}/book/) - # TODO(panyx0178): Figure out how to add word2vec and image_classification - # as deps. - # set_tests_properties(${TARGET_NAME} - # PROPERTIES DEPENDS ${DEP_TEST}) - endforeach() + if(inference_test_ARGS) + set_tests_properties(${TARGET_NAME} + PROPERTIES DEPENDS "${inference_test_ARGS}") + endif() + endif(WITH_TESTING) endfunction(inference_api_test) - cc_library(paddle_inference_api - SRCS paddle_inference_api.cc + SRCS paddle_inference_api.cc paddle_inference_api_impl.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) -cc_library(paddle_inference_api_impl - SRCS paddle_inference_api_impl.cc - DEPS paddle_inference_api paddle_fluid_api) - cc_test(test_paddle_inference_api SRCS test_paddle_inference_api.cc DEPS paddle_inference_api) inference_api_test(test_paddle_inference_api_impl - test_paddle_inference_api_impl.cc) + ARGS test_word2vec test_image_classification) + +if (WITH_ANAKIN AND WITH_TESTING) # only needed in CI + # Due to Anakin do not have official library releases and the versions of protobuf and cuda do not match Paddle's, + # so anakin library will not be merged to our official inference library. To use anakin prediction API, one need to + # compile the libinference_anakin_api.a and compile with anakin.so. + nv_library(inference_anakin_api SHARED SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc) + target_compile_options(inference_anakin_api BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) + target_link_libraries(inference_anakin_api anakin anakin_saber_common) + cc_test(inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc + ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin + DEPS inference_anakin_api) + target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) +endif() + +if(WITH_TESTING) + add_subdirectory(demo) +endif() diff --git a/paddle/contrib/inference/demo/CMakeLists.txt b/paddle/contrib/inference/demo/CMakeLists.txt new file mode 100644 index 0000000000..7b0fa77ad1 --- /dev/null +++ b/paddle/contrib/inference/demo/CMakeLists.txt @@ -0,0 +1,16 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +inference_api_test(simple_on_word2vec ARGS test_word2vec) diff --git a/paddle/contrib/inference/demo/simple_on_word2vec.cc b/paddle/contrib/inference/demo/simple_on_word2vec.cc new file mode 100644 index 0000000000..192a641426 --- /dev/null +++ b/paddle/contrib/inference/demo/simple_on_word2vec.cc @@ -0,0 +1,128 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file contains a simple demo for how to take a model for inference. + */ + +#include +#include +#include +#include +#include "paddle/contrib/inference/paddle_inference_api.h" +namespace paddle { +namespace demo { + +DEFINE_string(dirname, "", "Directory of the inference model."); + +void Main(bool use_gpu) { + //# 1. Create PaddlePredictor with a config. + NativeConfig config; + config.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config.use_gpu = use_gpu; + config.fraction_of_gpu_memory = 0.15; + config.device = 0; + auto predictor = + CreatePaddlePredictor(config); + + for (int batch_id = 0; batch_id < 3; batch_id++) { + //# 2. Prepare input. + int64_t data[4] = {1, 2, 3, 4}; + + PaddleBuf buf{.data = data, .length = sizeof(data)}; + PaddleTensor tensor{.name = "", + .shape = std::vector({4, 1}), + .data = buf, + .dtype = PaddleDType::INT64}; + + // For simplicity, we set all the slots with the same data. + std::vector slots(4, tensor); + + //# 3. Run + std::vector outputs; + CHECK(predictor->Run(slots, &outputs)); + + //# 4. Get output. + ASSERT_EQ(outputs.size(), 1UL); + LOG(INFO) << "output buffer size: " << outputs.front().data.length; + const size_t num_elements = outputs.front().data.length / sizeof(float); + // The outputs' buffers are in CPU memory. + for (size_t i = 0; i < std::min(5UL, num_elements); i++) { + LOG(INFO) << static_cast(outputs.front().data.data)[i]; + } + // TODO(Superjomn): this is should be free automatically + free(outputs[0].data.data); + } +} + +void MainThreads(int num_threads, bool use_gpu) { + // Multi-threads only support on CPU + // 0. Create PaddlePredictor with a config. + NativeConfig config; + config.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config.use_gpu = use_gpu; + config.fraction_of_gpu_memory = 0.15; + config.device = 0; + auto main_predictor = + CreatePaddlePredictor(config); + + std::vector threads; + for (int tid = 0; tid < num_threads; ++tid) { + threads.emplace_back([&, tid]() { + // 1. clone a predictor which shares the same parameters + auto predictor = main_predictor->Clone(); + constexpr int num_batches = 3; + for (int batch_id = 0; batch_id < num_batches; ++batch_id) { + // 2. Dummy Input Data + int64_t data[4] = {1, 2, 3, 4}; + PaddleBuf buf{.data = data, .length = sizeof(data)}; + PaddleTensor tensor{.name = "", + .shape = std::vector({4, 1}), + .data = buf, + .dtype = PaddleDType::INT64}; + std::vector inputs(4, tensor); + std::vector outputs; + // 3. Run + CHECK(predictor->Run(inputs, &outputs)); + + // 4. Get output. + ASSERT_EQ(outputs.size(), 1UL); + LOG(INFO) << "TID: " << tid << ", " + << "output buffer size: " << outputs.front().data.length; + const size_t num_elements = outputs.front().data.length / sizeof(float); + // The outputs' buffers are in CPU memory. + for (size_t i = 0; i < std::min(5UL, num_elements); i++) { + LOG(INFO) << static_cast(outputs.front().data.data)[i]; + } + free(outputs[0].data.data); + } + }); + } + for (int i = 0; i < num_threads; ++i) { + threads[i].join(); + } +} + +TEST(demo, word2vec_cpu) { Main(false /*use_gpu*/); } +TEST(demo_multi_threads, word2vec_cpu_1) { MainThreads(1, false /*use_gpu*/); } +TEST(demo_multi_threads, word2vec_cpu_4) { MainThreads(4, false /*use_gpu*/); } + +#ifdef PADDLE_WITH_CUDA +TEST(demo, word2vec_gpu) { Main(true /*use_gpu*/); } +TEST(demo_multi_threads, word2vec_gpu_1) { MainThreads(1, true /*use_gpu*/); } +TEST(demo_multi_threads, word2vec_gpu_4) { MainThreads(4, true /*use_gpu*/); } +#endif + +} // namespace demo +} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api.h b/paddle/contrib/inference/paddle_inference_api.h index f804d9b286..77e2d77b6b 100644 --- a/paddle/contrib/inference/paddle_inference_api.h +++ b/paddle/contrib/inference/paddle_inference_api.h @@ -1,16 +1,16 @@ /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 +http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ /* * This file contains the definition of a simple Inference API for Paddle. @@ -40,20 +40,30 @@ struct PaddleBuf { struct PaddleTensor { std::string name; // variable name. std::vector shape; + // TODO(Superjomn) for LoD support, add a vector> field if needed. PaddleBuf data; // blob of data. PaddleDType dtype; }; +enum class PaddleEngineKind { + kNative = 0, // Use the native Fluid facility. + kAnakin, // Use Anakin for inference. + // TODO(Superjomn) support following engines latter. + // kTensorRT, // Use TensorRT for inference. + // kAutoMixedAnakin, // Automatically mix Fluid with Anakin. + // kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. +}; + /* -* A simple Inference API for Paddle. Currently this API might just be used by -* non-sequence scenerios. -* TODO(Superjomn) Prepare another API for NLP-related usages. -*/ + * A simple Inference API for Paddle. Currently this API can be used by + * non-sequence scenerios. + */ class PaddlePredictor { public: struct Config; PaddlePredictor() = default; PaddlePredictor(const PaddlePredictor&) = delete; + PaddlePredictor& operator=(const PaddlePredictor&) = delete; // Predict an record. // The caller should be responsible for allocating and releasing the memory of @@ -66,34 +76,41 @@ class PaddlePredictor { // be thread-safe. virtual std::unique_ptr Clone() = 0; - virtual bool InitShared() { return false; } // Destroy the Predictor. - virtual ~PaddlePredictor() {} - - friend std::unique_ptr CreatePaddlePredictor( - const PaddlePredictor::Config& config); + virtual ~PaddlePredictor() = default; // The common configs for all the predictors. struct Config { - enum class EngineKind; - std::string model_dir; // path to the model directory. bool enable_engine{false}; // Enable to execute (part of) the model on - // third-party engines. - EngineKind engine_kind{Config::EngineKind::kNone}; - - enum class EngineKind { - kNone = -1, // Use the native Fluid facility. - kAnakin, // Use Anakin for inference. - kTensorRT, // Use TensorRT for inference. - kAutoMixedAnakin, // Automatically mix Fluid with Anakin. - kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. - }; }; }; -// A factory to help create difference predictor. -template -std::unique_ptr CreatePaddlePredictor(const ConfigT& config); +struct NativeConfig : public PaddlePredictor::Config { + // GPU related fields. + bool use_gpu{false}; + int device{0}; + float fraction_of_gpu_memory{-1.f}; // Negative to notify initialization. + std::string prog_file; + std::string param_file; +}; + +// Configurations for Anakin engine. +struct AnakinConfig : public PaddlePredictor::Config { + int device; + std::string model_file; + int max_batch_size{-1}; +}; + +// A factory to help create different predictors. +// +// FOR EXTENSION DEVELOPER: +// Different predictors are designated by config type and engine kind. Similar +// configs can be merged, but there shouldn't be a huge config containing +// different fields for more than one kind of predictors. +// +// Similarly, each engine kind should map to a unique predictor implementation. +template +std::unique_ptr CreatePaddlePredictor(const ConfigT& config); } // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_anakin_engine.cc b/paddle/contrib/inference/paddle_inference_api_anakin_engine.cc new file mode 100644 index 0000000000..5bafc58fa5 --- /dev/null +++ b/paddle/contrib/inference/paddle_inference_api_anakin_engine.cc @@ -0,0 +1,113 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/contrib/inference/paddle_inference_api_anakin_engine.h" +#include + +namespace paddle { + +PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor( + const AnakinConfig &config) { + CHECK(Init(config)); +} + +bool PaddleInferenceAnakinPredictor::Init(const AnakinConfig &config) { + if (!(graph_.load(config.model_file))) { + return false; + } + graph_.ResetBatchSize("input_0", config.max_batch_size); + // optimization for graph + if (!(graph_.Optimize())) { + return false; + } + // construct executer + executor_.init(graph_); + return true; +} + +bool PaddleInferenceAnakinPredictor::Run( + const std::vector &inputs, + std::vector *output_data) { + for (const auto &input : inputs) { + if (input.dtype != PaddleDType::FLOAT32) { + LOG(ERROR) << "Only support float type inputs. " << input.name + << "'s type is not float"; + return false; + } + auto d_tensor_in_p = executor_.get_in(input.name); + float *d_data_p = d_tensor_in_p->mutable_data(); + if (cudaMemcpy(d_data_p, + static_cast(input.data.data), + d_tensor_in_p->valid_size() * sizeof(float), + cudaMemcpyHostToDevice) != 0) { + LOG(ERROR) << "copy data from CPU to GPU error"; + return false; + } + } + + executor_.prediction(); + + if (output_data->empty()) { + LOG(ERROR) << "At least one output should be set with tensors' names."; + return false; + } + for (auto &output : *output_data) { + auto *tensor = executor_.get_out(output.name); + output.shape = tensor->shape(); + // Copy data from GPU -> CPU + if (cudaMemcpy(output.data.data, + tensor->mutable_data(), + tensor->valid_size() * sizeof(float), + cudaMemcpyDeviceToHost) != 0) { + LOG(ERROR) << "copy data from GPU to CPU error"; + return false; + } + } + return true; +} + +anakin::Net + &PaddleInferenceAnakinPredictor::get_executer() { + return executor_; +} + +// the cloned new Predictor of anakin share the same net weights from original +// Predictor +std::unique_ptr PaddleInferenceAnakinPredictor::Clone() { + VLOG(3) << "Anakin Predictor::clone"; + std::unique_ptr cls(new PaddleInferenceAnakinPredictor()); + // construct executer from other graph + auto anakin_predictor_p = + dynamic_cast(cls.get()); + if (!anakin_predictor_p) { + LOG(ERROR) << "fail to call Init"; + return nullptr; + } + anakin_predictor_p->get_executer().init(graph_); + + return std::move(cls); +} + +// A factory to help create difference predictor. +template <> +std::unique_ptr +CreatePaddlePredictor( + const AnakinConfig &config) { + VLOG(3) << "Anakin Predictor create."; + std::unique_ptr x( + new PaddleInferenceAnakinPredictor(config)); + return x; +}; + +} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_anakin_engine.h b/paddle/contrib/inference/paddle_inference_api_anakin_engine.h new file mode 100644 index 0000000000..212ba41cdf --- /dev/null +++ b/paddle/contrib/inference/paddle_inference_api_anakin_engine.h @@ -0,0 +1,60 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file contains the implementation of inference API with Anakin engine + * embeded, this API can only support Anakin models. + */ + +#pragma once + +#include "paddle/contrib/inference/paddle_inference_api.h" + +// from anakin +#include "framework/core/net/net.h" +#include "saber/saber_types.h" + +namespace paddle { + +class PaddleInferenceAnakinPredictor : public PaddlePredictor { + public: + PaddleInferenceAnakinPredictor() {} + + PaddleInferenceAnakinPredictor(const AnakinConfig& config); + + // NOTE Unlike the native engine, the buffers of anakin engine's output_data + // should be allocated first. + bool Run(const std::vector& inputs, + std::vector* output_data) override; + + std::unique_ptr Clone() override; + + anakin::Net& + get_executer(); + + ~PaddleInferenceAnakinPredictor() override{}; + + private: + bool Init(const AnakinConfig& config); + + anakin::graph::Graph + graph_; + anakin::Net + executor_; + AnakinConfig config_; +}; + +} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc b/paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc new file mode 100644 index 0000000000..1d41a5c73e --- /dev/null +++ b/paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc @@ -0,0 +1,67 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include + +#include "paddle/contrib/inference/paddle_inference_api.h" + +DEFINE_string(model, "", "Directory of the inference model."); + +namespace paddle { + +AnakinConfig GetConfig() { + AnakinConfig config; + config.model_file = FLAGS_model; + config.device = 0; + config.max_batch_size = 1; + return config; +} + +TEST(inference, anakin) { + AnakinConfig config = GetConfig(); + auto predictor = + CreatePaddlePredictor(config); + + float data[1 * 3 * 224 * 224] = {1.0f}; + + PaddleBuf buf{.data = data, .length = sizeof(data)}; + PaddleTensor tensor{.name = "input_0", + .shape = std::vector({1, 3, 224, 224}), + .data = buf, + .dtype = PaddleDType::FLOAT32}; + + // For simplicity, we set all the slots with the same data. + std::vector paddle_tensor_feeds(1, tensor); + + float data_out[1000]; + + PaddleBuf buf_out{.data = data_out, .length = sizeof(data)}; + PaddleTensor tensor_out{.name = "prob_out", + .shape = std::vector({1000, 1}), + .data = buf_out, + .dtype = PaddleDType::FLOAT32}; + + std::vector outputs(1, tensor_out); + + ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs)); + + float* data_o = static_cast(outputs[0].data.data); + for (size_t j = 0; j < 1000; ++j) { + LOG(INFO) << "output[" << j << "]: " << data_o[j]; + } +} + +} // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_impl.cc b/paddle/contrib/inference/paddle_inference_api_impl.cc index ebe4c32918..bda2981a14 100644 --- a/paddle/contrib/inference/paddle_inference_api_impl.cc +++ b/paddle/contrib/inference/paddle_inference_api_impl.cc @@ -1,16 +1,16 @@ /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 +http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ #include #include @@ -54,18 +54,24 @@ std::string num2str(T a) { } } // namespace -bool PaddlePredictorImpl::Init() { +bool NativePaddlePredictor::Init( + std::shared_ptr parent_scope) { VLOG(3) << "Predictor::init()"; - // TODO(panyx0718): Should CPU vs GPU device be decided by id? - if (config_.device >= 0) { + if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); } else { place_ = paddle::platform::CPUPlace(); } - paddle::framework::InitDevices(false); + if (parent_scope) { + scope_ = parent_scope; + sub_scope_ = &(parent_scope->NewScope()); + } else { + paddle::framework::InitDevices(false); + scope_.reset(new paddle::framework::Scope()); + } + executor_.reset(new paddle::framework::Executor(place_)); - scope_.reset(new paddle::framework::Scope()); // Initialize the inference program if (!config_.model_dir.empty()) { @@ -84,20 +90,24 @@ bool PaddlePredictorImpl::Init() { return false; } ctx_ = executor_->Prepare(*inference_program_, 0); + executor_->CreateVariables( + *inference_program_, sub_scope_ ? sub_scope_ : scope_.get(), 0); - // Create variables - // TODO(panyx0718): Why need to test share_variables here? - if (config_.share_variables) { - executor_->CreateVariables(*inference_program_, scope_.get(), 0); - } // Get the feed_target_names and fetch_target_names feed_target_names_ = inference_program_->GetFeedTargetNames(); fetch_target_names_ = inference_program_->GetFetchTargetNames(); return true; } -bool PaddlePredictorImpl::Run(const std::vector &inputs, - std::vector *output_data) { +NativePaddlePredictor::~NativePaddlePredictor() { + if (sub_scope_) { + PADDLE_ENFORCE_NOT_NULL(scope_, "Should have parent scope!"); + scope_->DeleteScope(sub_scope_); + } +}; + +bool NativePaddlePredictor::Run(const std::vector &inputs, + std::vector *output_data) { VLOG(3) << "Predictor::predict"; Timer timer; timer.tic(); @@ -120,11 +130,12 @@ bool PaddlePredictorImpl::Run(const std::vector &inputs, } // Run the inference program // if share variables, we need not create variables - executor_->RunPreparedContext(ctx_.get(), - scope_.get(), - &feed_targets, - &fetch_targets, - !config_.share_variables); + executor_->RunPreparedContext( + ctx_.get(), + sub_scope_ != nullptr ? sub_scope_ : scope_.get(), + &feed_targets, + &fetch_targets, + false /* don't create variable eatch time */); if (!GetFetch(fetchs, output_data)) { LOG(ERROR) << "fail to get fetchs"; return false; @@ -133,59 +144,20 @@ bool PaddlePredictorImpl::Run(const std::vector &inputs, return true; } -std::unique_ptr PaddlePredictorImpl::Clone() { +std::unique_ptr NativePaddlePredictor::Clone() { VLOG(3) << "Predictor::clone"; - std::unique_ptr cls(new PaddlePredictorImpl(config_)); - if (!cls->InitShared()) { - LOG(ERROR) << "fail to call InitShared"; + std::unique_ptr cls(new NativePaddlePredictor(config_)); + + if (!dynamic_cast(cls.get())->Init(scope_)) { + LOG(ERROR) << "fail to call Init"; return nullptr; } // fix manylinux compile error. return std::move(cls); } -// TODO(panyx0718): Consider merge with Init()? -bool PaddlePredictorImpl::InitShared() { - VLOG(3) << "Predictor::init_shared"; - // 1. Define place, executor, scope - if (this->config_.device >= 0) { - place_ = platform::CUDAPlace(); - } else { - place_ = platform::CPUPlace(); - } - this->executor_.reset(new framework::Executor(this->place_)); - this->scope_.reset(new framework::Scope()); - // Initialize the inference program - if (!this->config_.model_dir.empty()) { - // Parameters are saved in separate files sited in - // the specified `dirname`. - this->inference_program_ = inference::Load( - this->executor_.get(), this->scope_.get(), this->config_.model_dir); - } else if (!this->config_.prog_file.empty() && - !this->config_.param_file.empty()) { - // All parameters are saved in a single file. - // The file names should be consistent with that used - // in Python API `fluid.io.save_inference_model`. - this->inference_program_ = inference::Load(this->executor_.get(), - this->scope_.get(), - this->config_.prog_file, - this->config_.param_file); - } - this->ctx_ = this->executor_->Prepare(*this->inference_program_, 0); - // 3. create variables - // TODO(panyx0718): why test share_variables. - if (config_.share_variables) { - this->executor_->CreateVariables( - *this->inference_program_, this->scope_.get(), 0); - } - // 4. Get the feed_target_names and fetch_target_names - this->feed_target_names_ = this->inference_program_->GetFeedTargetNames(); - this->fetch_target_names_ = this->inference_program_->GetFetchTargetNames(); - return true; -} - -bool PaddlePredictorImpl::SetFeed(const std::vector &inputs, - std::vector *feeds) { +bool NativePaddlePredictor::SetFeed(const std::vector &inputs, + std::vector *feeds) { VLOG(3) << "Predictor::set_feed"; if (inputs.size() != feed_target_names_.size()) { LOG(ERROR) << "wrong feed input size."; @@ -213,7 +185,7 @@ bool PaddlePredictorImpl::SetFeed(const std::vector &inputs, return true; } -bool PaddlePredictorImpl::GetFetch( +bool NativePaddlePredictor::GetFetch( const std::vector &fetchs, std::vector *outputs) { VLOG(3) << "Predictor::get_fetch"; @@ -280,23 +252,31 @@ bool PaddlePredictorImpl::GetFetch( } template <> -std::unique_ptr CreatePaddlePredictor( - const ConfigImpl &config) { - VLOG(3) << "create PaddlePredictorImpl"; - // 1. GPU memeroy - std::vector flags; - if (config.fraction_of_gpu_memory >= 0.0f || - config.fraction_of_gpu_memory <= 0.95f) { - flags.push_back("dummpy"); - std::string flag = "--fraction_of_gpu_memory_to_use=" + - num2str(config.fraction_of_gpu_memory); - flags.push_back(flag); - VLOG(3) << "set flag: " << flag; - framework::InitGflags(flags); +std::unique_ptr +CreatePaddlePredictor( + const NativeConfig &config) { + VLOG(3) << "create NativePaddlePredictor"; + if (config.use_gpu) { + // 1. GPU memeroy + PADDLE_ENFORCE_GT( + config.fraction_of_gpu_memory, + 0.f, + "fraction_of_gpu_memory in the config should be set to range (0., 1.]"); + PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device); + std::vector flags; + if (config.fraction_of_gpu_memory >= 0.0f || + config.fraction_of_gpu_memory <= 0.95f) { + flags.push_back("dummpy"); + std::string flag = "--fraction_of_gpu_memory_to_use=" + + num2str(config.fraction_of_gpu_memory); + flags.push_back(flag); + VLOG(3) << "set flag: " << flag; + framework::InitGflags(flags); + } } - std::unique_ptr predictor(new PaddlePredictorImpl(config)); - if (!dynamic_cast(predictor.get())->Init()) { + std::unique_ptr predictor(new NativePaddlePredictor(config)); + if (!dynamic_cast(predictor.get())->Init(nullptr)) { return nullptr; } return std::move(predictor); diff --git a/paddle/contrib/inference/paddle_inference_api_impl.h b/paddle/contrib/inference/paddle_inference_api_impl.h index c545461680..86d1db7bcc 100644 --- a/paddle/contrib/inference/paddle_inference_api_impl.h +++ b/paddle/contrib/inference/paddle_inference_api_impl.h @@ -29,42 +29,37 @@ namespace paddle { -struct ConfigImpl : public PaddlePredictor::Config { - int device; - float fraction_of_gpu_memory; - std::string prog_file; - std::string param_file; - bool share_variables; -}; - -class PaddlePredictorImpl : public PaddlePredictor { +class NativePaddlePredictor : public PaddlePredictor { public: - explicit PaddlePredictorImpl(const ConfigImpl &config) : config_(config) {} + explicit NativePaddlePredictor(const NativeConfig &config) + : config_(config) {} - bool Init(); + // will only create sub scope if have global scope + bool Init(std::shared_ptr parent_scope); bool Run(const std::vector &inputs, std::vector *output_data) override; std::unique_ptr Clone() override; - ~PaddlePredictorImpl() override{}; + ~NativePaddlePredictor() override; private: - bool InitShared() override; bool SetFeed(const std::vector &input_datas, std::vector *feeds); bool GetFetch(const std::vector &fetchs, std::vector *output_data); - ConfigImpl config_; + NativeConfig config_; platform::Place place_; std::unique_ptr executor_; - std::unique_ptr scope_; + std::shared_ptr scope_; std::unique_ptr ctx_; std::unique_ptr inference_program_; std::vector feed_target_names_; std::vector fetch_target_names_; + // Do not use unique_ptr, use parent scope to delete + framework::Scope *sub_scope_{nullptr}; }; } // namespace paddle diff --git a/paddle/contrib/inference/test_paddle_inference_api_impl.cc b/paddle/contrib/inference/test_paddle_inference_api_impl.cc index 096293a4e2..5d843010e0 100644 --- a/paddle/contrib/inference/test_paddle_inference_api_impl.cc +++ b/paddle/contrib/inference/test_paddle_inference_api_impl.cc @@ -15,6 +15,8 @@ limitations under the License. */ #include #include +#include + #include "gflags/gflags.h" #include "paddle/contrib/inference/paddle_inference_api_impl.h" #include "paddle/fluid/inference/tests/test_helper.h" @@ -40,19 +42,24 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { return pt; } -ConfigImpl GetConfig() { - ConfigImpl config; +NativeConfig GetConfig() { + NativeConfig config; config.model_dir = FLAGS_dirname + "word2vec.inference.model"; LOG(INFO) << "dirname " << config.model_dir; config.fraction_of_gpu_memory = 0.15; +#ifdef PADDLE_WITH_CUDA + config.use_gpu = true; +#else + config.use_gpu = false; +#endif config.device = 0; - config.share_variables = true; return config; } -TEST(paddle_inference_api_impl, word2vec) { - ConfigImpl config = GetConfig(); - std::unique_ptr predictor = CreatePaddlePredictor(config); +void MainWord2Vec(bool use_gpu) { + NativeConfig config = GetConfig(); + auto predictor = CreatePaddlePredictor(config); + config.use_gpu = use_gpu; framework::LoDTensor first_word, second_word, third_word, fourth_word; framework::LoD lod{{0, 1}}; @@ -74,7 +81,7 @@ TEST(paddle_inference_api_impl, word2vec) { ASSERT_EQ(outputs.size(), 1UL); size_t len = outputs[0].data.length; float* data = static_cast(outputs[0].data.data); - for (int j = 0; j < len / sizeof(float); ++j) { + for (size_t j = 0; j < len / sizeof(float); ++j) { ASSERT_LT(data[j], 1.0); ASSERT_GT(data[j], -1.0); } @@ -92,7 +99,7 @@ TEST(paddle_inference_api_impl, word2vec) { TestInference(config.model_dir, cpu_feeds, cpu_fetchs1); float* lod_data = output1.data(); - for (size_t i = 0; i < output1.numel(); ++i) { + for (int i = 0; i < output1.numel(); ++i) { EXPECT_LT(lod_data[i] - data[i], 1e-3); EXPECT_GT(lod_data[i] - data[i], -1e-3); } @@ -100,11 +107,11 @@ TEST(paddle_inference_api_impl, word2vec) { free(outputs[0].data.data); } -TEST(paddle_inference_api_impl, image_classification) { +void MainImageClassification(bool use_gpu) { int batch_size = 2; - bool use_mkldnn = false; bool repeat = false; - ConfigImpl config = GetConfig(); + NativeConfig config = GetConfig(); + config.use_gpu = use_gpu; config.model_dir = FLAGS_dirname + "image_classification_resnet.inference.model"; @@ -126,14 +133,10 @@ TEST(paddle_inference_api_impl, image_classification) { std::vector cpu_fetchs1; cpu_fetchs1.push_back(&output1); - TestInference(config.model_dir, - cpu_feeds, - cpu_fetchs1, - repeat, - is_combined, - use_mkldnn); + TestInference( + config.model_dir, cpu_feeds, cpu_fetchs1, repeat, is_combined); - std::unique_ptr predictor = CreatePaddlePredictor(config); + auto predictor = CreatePaddlePredictor(config); std::vector paddle_tensor_feeds; paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&input)); @@ -144,10 +147,148 @@ TEST(paddle_inference_api_impl, image_classification) { float* data = static_cast(outputs[0].data.data); float* lod_data = output1.data(); for (size_t j = 0; j < len / sizeof(float); ++j) { - EXPECT_LT(lod_data[j] - data[j], 1e-10); - EXPECT_GT(lod_data[j] - data[j], -1e-10); + EXPECT_NEAR(lod_data[j], data[j], 1e-3); } free(data); } +void MainThreadsWord2Vec(bool use_gpu) { + NativeConfig config = GetConfig(); + config.use_gpu = use_gpu; + auto main_predictor = CreatePaddlePredictor(config); + + // prepare inputs data and reference results + constexpr int num_jobs = 3; + std::vector> jobs(num_jobs); + std::vector> paddle_tensor_feeds(num_jobs); + std::vector refs(num_jobs); + for (size_t i = 0; i < jobs.size(); ++i) { + // each job has 4 words + jobs[i].resize(4); + for (size_t j = 0; j < 4; ++j) { + framework::LoD lod{{0, 1}}; + int64_t dict_size = 2073; // The size of dictionary + SetupLoDTensor(&jobs[i][j], lod, static_cast(0), dict_size - 1); + paddle_tensor_feeds[i].push_back(LodTensorToPaddleTensor(&jobs[i][j])); + } + + // get reference result of each job + std::vector ref_feeds; + std::vector ref_fetches(1, &refs[i]); + for (auto& word : jobs[i]) { + ref_feeds.push_back(&word); + } + TestInference(config.model_dir, ref_feeds, ref_fetches); + } + + // create threads and each thread run 1 job + std::vector threads; + for (int tid = 0; tid < num_jobs; ++tid) { + threads.emplace_back([&, tid]() { + auto predictor = main_predictor->Clone(); + auto& local_inputs = paddle_tensor_feeds[tid]; + std::vector local_outputs; + ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); + + // check outputs range + ASSERT_EQ(local_outputs.size(), 1UL); + const size_t len = local_outputs[0].data.length; + float* data = static_cast(local_outputs[0].data.data); + for (size_t j = 0; j < len / sizeof(float); ++j) { + ASSERT_LT(data[j], 1.0); + ASSERT_GT(data[j], -1.0); + } + + // check outputs correctness + float* ref_data = refs[tid].data(); + EXPECT_EQ(refs[tid].numel(), static_cast(len / sizeof(float))); + for (int i = 0; i < refs[tid].numel(); ++i) { + EXPECT_NEAR(ref_data[i], data[i], 1e-3); + } + free(data); + }); + } + for (int i = 0; i < num_jobs; ++i) { + threads[i].join(); + } +} + +void MainThreadsImageClassification(bool use_gpu) { + constexpr int num_jobs = 4; // each job run 1 batch + constexpr int batch_size = 1; + NativeConfig config = GetConfig(); + config.use_gpu = use_gpu; + config.model_dir = + FLAGS_dirname + "image_classification_resnet.inference.model"; + + auto main_predictor = CreatePaddlePredictor(config); + std::vector jobs(num_jobs); + std::vector> paddle_tensor_feeds(num_jobs); + std::vector refs(num_jobs); + for (size_t i = 0; i < jobs.size(); ++i) { + // prepare inputs + std::vector> feed_target_shapes = + GetFeedTargetShapes(config.model_dir, /*is_combined*/ false); + feed_target_shapes[0][0] = batch_size; + framework::DDim input_dims = framework::make_ddim(feed_target_shapes[0]); + SetupTensor(&jobs[i], input_dims, 0.f, 1.f); + paddle_tensor_feeds[i].push_back(LodTensorToPaddleTensor(&jobs[i])); + + // get reference result of each job + std::vector ref_feeds(1, &jobs[i]); + std::vector ref_fetches(1, &refs[i]); + TestInference(config.model_dir, ref_feeds, ref_fetches); + } + + // create threads and each thread run 1 job + std::vector threads; + for (int tid = 0; tid < num_jobs; ++tid) { + threads.emplace_back([&, tid]() { + auto predictor = main_predictor->Clone(); + auto& local_inputs = paddle_tensor_feeds[tid]; + std::vector local_outputs; + ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); + + // check outputs correctness + ASSERT_EQ(local_outputs.size(), 1UL); + const size_t len = local_outputs[0].data.length; + float* data = static_cast(local_outputs[0].data.data); + float* ref_data = refs[tid].data(); + EXPECT_EQ(refs[tid].numel(), len / sizeof(float)); + for (int i = 0; i < refs[tid].numel(); ++i) { + EXPECT_NEAR(ref_data[i], data[i], 1e-3); + } + free(data); + }); + } + for (int i = 0; i < num_jobs; ++i) { + threads[i].join(); + } +} + +TEST(inference_api_native, word2vec_cpu) { MainWord2Vec(false /*use_gpu*/); } +TEST(inference_api_native, word2vec_cpu_threads) { + MainThreadsWord2Vec(false /*use_gpu*/); +} +TEST(inference_api_native, image_classification_cpu) { + MainThreadsImageClassification(false /*use_gpu*/); +} +TEST(inference_api_native, image_classification_cpu_threads) { + MainThreadsImageClassification(false /*use_gpu*/); +} + +#ifdef PADDLE_WITH_CUDA +TEST(inference_api_native, word2vec_gpu) { MainWord2Vec(true /*use_gpu*/); } +TEST(inference_api_native, word2vec_gpu_threads) { + MainThreadsWord2Vec(true /*use_gpu*/); +} +TEST(inference_api_native, image_classification_gpu) { + MainThreadsImageClassification(true /*use_gpu*/); +} +TEST(inference_api_native, image_classification_gpu_threads) { + MainThreadsImageClassification(true /*use_gpu*/); +} + +#endif + } // namespace paddle diff --git a/paddle/contrib/tape/CMakeLists.txt b/paddle/contrib/tape/CMakeLists.txt new file mode 100644 index 0000000000..5450359d85 --- /dev/null +++ b/paddle/contrib/tape/CMakeLists.txt @@ -0,0 +1,25 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# + +if(APPLE) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") +endif(APPLE) + +cc_library(tape_variable SRCS variable.cc DEPS ${FLUID_CORE_MODULES} device_context framework_proto proto_desc operator) +cc_library(tape SRCS tape.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} tape_variable) + +cc_test(test_tape + SRCS test_tape.cc + DEPS tape tape_variable) diff --git a/paddle/contrib/tape/README.md b/paddle/contrib/tape/README.md new file mode 100644 index 0000000000..16c22a45d5 --- /dev/null +++ b/paddle/contrib/tape/README.md @@ -0,0 +1,252 @@ +# Dynamic Graph on Fluid + +PaddlePaddle Fluid is targeting the autodiff without tape, which, however, is very +challenging and we are still way from there. DyNet and PyTorch provide a good design +idea, the *tape*, that significantly eases the challenge. Also, DyNet provides +a C++ API that is as convenient as Python but with higher efficiency and could +conveniently integrate with industrial/production systems. This package, `tape`, +combines the good of + +1. tape from PyTorch and DyNet +2. C++ API and core from DyNet +3. rich set of operators from PaddlePaddle + +## Overview + +We can implement Dynet-like Tape(See this [survey](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/survey/dynamic_graph.md)) +by wrapping Paddle Fluid's `Operator` and `Variable`. + +The user API is straight forward since + +1. it is imperative. And it uses host language's control flow logic. +1. it avoids extra concepts such as `Scope` and `Executor`. + +All of these benefits come at the cost of just adding one line `reset_global_tape` +at every iteration. + +## Code Structure + +In short, the `Tape` contains a vector of `OpHandle`s. And an `OpHandle` contains its +`type`, the pointers to the `Variable`s, and necessary attributes. + +```c++ +class Variable { +public: + VriableHandle Grad(); // returns its gradient variable +private: + framework::VarDesc desc_; // compile time infershape, necessary for lazy execution + framework::Variable var_; // run time variable, holds data memory +}; + +using VariableHandle = shared_ptr; + +struct OpHandle { + string type_; + map> inputs_; + map> outputs_; + AttributeMap attrs_; +}; + +class Tape { +public: + void AddOp(OpHandle); // add op + void Forward(); // execute the tape_ + void Backward(); // execute the backward of the tape_ +private: + vector tape_; +}; +``` + +We uses `Function` to indicate layers. It takes care of parameter +initialization and `AddOp` to the Tape when it is called. + +```c++ +class Linear { + public: + Linear(int in_dim, int out_dim, const std::string &act) + : w_(new Variable("LinearWeight")), + b_(new Variable("LinearBias")), + act_(act) { + Tape init_tape; + + std::string initializer = "fill_constant"; + framework::AttributeMap attrs; + attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; + attrs["shape"] = std::vector{in_dim, out_dim}; + attrs["value"] = 1.0f; + init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs); + + attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; + attrs["shape"] = std::vector{out_dim}; + attrs["value"] = 1.0f; + init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs); + + init_tape.Forward(); + } + + VariableHandle operator()(VariableHandle input) { + VariableHandle pre_bias(new Variable("linear")); + get_global_tape().AddOp("mul", + {{"X", {input}}, {"Y", {w_}}}, + {{"Out", {pre_bias}}}, + {{"x_num_col_dims", 1}, {"y_num_col_dims", 1}}); + VariableHandle pre_act(new Variable("linear")); + get_global_tape().AddOp("elementwise_add", + {{"X", {pre_bias}}, {"Y", {b_}}}, + {{"Out", {pre_act}}}, + {{"axis", 1}}); + VariableHandle post_act(new Variable("linear")); + get_global_tape().AddOp(act_, + {{"X", {pre_act}}}, + {{"Out", {post_act}}}, + {}); + return post_act; + } + + std::vector Params() { return {w_, b_}; } + + private: + VariableHandle w_; + VariableHandle b_; + std::string act_; +}; +``` + +## User API + +```c++ +// Model function +paddle::tape::Linear linear1(3, 3, "relu"); // init weight and bias +paddle::tape::Linear linear2(3, 3, "relu"); // init weight and bias +paddle::tape::Mean mean; + +// Optimizer +paddle::tape::SGD sgd(0.001); + +// Data Feeder +paddle::tape::Fill data_feeder(...); +VariableHandle input(new paddle::tape::Variable("input")); +VariableHandle label(new paddle::tape::Variable("label")); + +for (int i = 0; i < 2; ++i) { + reset_global_tape(); + + data_feeder(input, label); + + auto loss = softmax(linear2(linear1(input)), label); // compile time InferShape & InferVarType + LOG(INFO) << loss.value(); // Run forward up to loss + + // Run backward, store gradient of w at w->Grad() + get_global_tape.Backward(loss); + + // Update w + sgd(linear1.Params()); + sgd(linear2.Params()); +} +``` + +
+ +digraph G { + + subgraph cluster_0 { + node [shape=record,style=filled]; + style=filled; + color=lightgrey; + linear1 [label="{type: mul | {input | {X: before_mul1 | Y: weight1}} | {output | Out: before_bias1}}"]; + elementwise_add1 [label="{type: elementwise_add | {input | {X: before_bias1 | Y: bias1}} | {output | Out: before_act1}}"]; + relu1 [label="{type: relu | {input | {X: before_act1 }} | {output | Out: after_act1}}"]; + + linear1 -> elementwise_add1->relu1; + label = "forward tape"; + } + + linear1:before_mul1->before_mul1 + linear1:weight1->weight1 + linear1:before_bias1->before_bias1 + + elementwise_add1:bias1->bias1 + elementwise_add1:before_bias1->before_bias1 + elementwise_add1:before_act1->before_act1 + + relu1:before_act1->before_act1 + relu1:after_act1->after_act1 + + subgraph cluster_1 { + node [shape=record,style=filled]; + style=filled; + color=lightgrey; + linear1_grad [label="{type: mul_grad | {input | {X: before_mul1 | Y: weight1| Out_grad: before_bias1_grad}} | {output |{X_grad: before_mul1_grad | Y_grad: weight1_grad}}}"]; + + elementwise_add1_grad [label="{type: elementwise_add_grad | {input | Out_grad: before_act1_grad} | {output |{X_grad: before_bias1_grad | Y_grad: bias1_grad}}}"]; + + relu1_grad [label="{type: relu_grad | {input | Out_grad: after_act1_grad} | {ouput | {X_grad: before_act1_grad }}}"]; + + linear1_grad -> elementwise_add1_grad ->relu1_grad [dir=back]; + label = "backward tape"; + } + + relu1_grad:after_act1_grad->after_act1_grad + relu1_grad:before_act1_grad->before_act1_grad + + elementwise_add1_grad:before_act1_grad->before_act1_grad + elementwise_add1_grad:before_bias1_grad->before_bias1_grad + elementwise_add1_grad:bias1_grad->bias1_grad + + linear1_grad:before_mul1->before_mul1 + linear1_grad:weight1->weight1 + linear1_grad:before_bias1_grad->before_bias1_grad + linear1_grad:before_mul1_grad->before_mul1_grad + linear1_grad:weight1_grad->weight1_grad + + + subgraph cluster_2 { + node [shape=record]; + label = "Linear1"; + weight1 + bias1 + } + + weight1 -> weight1_grad [ label="Grad()", style="dashed" ]; + bias1 -> bias1_grad [ label="Grad()", style="dashed"]; + + + +} +
+ +![Image](https://github.com/tonyyang-svail/Paddle/blob/cpp_tap/paddle/contrib/tape/computation_graph.png) + +## Code Reuse + +We want to stay close to Paddle Fluid as much as possible. + +### Reuse All Operators + +As all Ops are registered at `OpInfoMap`, the effort of adding a new `Function` +is about 10 lines of code, similar to expose an operator to Python. + +### Reuse Compile Time InferShape and InferVarType + +Note that all the symbolic information is stored at `tape::Varaible::desc_`, instead +of `ProgramDesc.block.vars`, we create a temporary `BlockDesc` to do `InferShape` and +`InferVarType` every time we `AddOp` to the tape. + +### Reuse Operator::Run + +We use smart pointer, instead of `Scope`, to manage memory. So we create a temporary +`Scope` for every `Operator::Run()`. + +## Possible Feature + +### Release Memory on Backward + +We can release memory aggressively. During backward, we can delete the OpHandle once +we have finished its backward. Since all the variable is managed by smart pointer, the +memory is automatically released when its `ref_count` goes to 0. + +### Kernel Fusion + +As a symbolic representation of the Tape is constructed first before the actual +execution, it would be possible to perform graph optimization. One use case is kernel +fusion. diff --git a/paddle/contrib/tape/computation_graph.png b/paddle/contrib/tape/computation_graph.png new file mode 100644 index 0000000000..6cf5ead735 Binary files /dev/null and b/paddle/contrib/tape/computation_graph.png differ diff --git a/paddle/contrib/tape/function.h b/paddle/contrib/tape/function.h new file mode 100644 index 0000000000..8c9694d9a2 --- /dev/null +++ b/paddle/contrib/tape/function.h @@ -0,0 +1,131 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include "paddle/contrib/tape/tape.h" +#include "paddle/contrib/tape/variable.h" +#include "paddle/fluid/framework/type_defs.h" + +namespace paddle { +namespace tape { + +class Function {}; + +class Fill { + public: + Fill(const std::string &initializer, const framework::AttributeMap &attrs) + : initializer_(initializer), attrs_(attrs) {} + + void operator()(VariableHandle var) { + get_global_tape().AddOp(initializer_, {}, {{"Out", {var}}}, attrs_); + } + + private: + const std::string initializer_; + const framework::AttributeMap attrs_; +}; + +class Mean { + public: + VariableHandle operator()(VariableHandle var) { + VariableHandle out(new Variable("mean")); + get_global_tape().AddOp("mean", {{"X", {var}}}, {{"Out", {out}}}, {}); + return out; + } +}; + +class Linear { + public: + Linear(int in_dim, int out_dim, const std::string &act) + : w_(new Variable("LinearWeight")), + b_(new Variable("LinearBias")), + act_(act) { + Tape init_tape; + + std::string initializer = "fill_constant"; + framework::AttributeMap attrs; + attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; + attrs["shape"] = std::vector{in_dim, out_dim}; + attrs["value"] = 1.0f; + init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs); + + attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; + attrs["shape"] = std::vector{out_dim}; + attrs["value"] = 1.0f; + init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs); + + init_tape.Forward(); + } + + VariableHandle operator()(VariableHandle input) { + VariableHandle pre_bias(new Variable("linear")); + get_global_tape().AddOp("mul", + {{"X", {input}}, {"Y", {w_}}}, + {{"Out", {pre_bias}}}, + {{"x_num_col_dims", 1}, {"y_num_col_dims", 1}}); + VariableHandle pre_act(new Variable("linear")); + get_global_tape().AddOp("elementwise_add", + {{"X", {pre_bias}}, {"Y", {b_}}}, + {{"Out", {pre_act}}}, + {{"axis", 1}}); + VariableHandle post_act(new Variable("linear")); + get_global_tape().AddOp( + act_, {{"X", {pre_act}}}, {{"Out", {post_act}}}, {}); + return post_act; + } + + std::vector Params() { return {w_, b_}; } + + private: + VariableHandle w_; + VariableHandle b_; + std::string act_; +}; + +class SGD { + public: + SGD(float learning_rate) : learning_rate_(new Variable("sgd")) { + Tape init_tape; + + std::string initializer = "fill_constant"; + framework::AttributeMap attrs; + attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; + attrs["shape"] = std::vector{1}; + attrs["value"] = learning_rate; + init_tape.AddOp(initializer, {}, {{"Out", {learning_rate_}}}, attrs); + + init_tape.Forward(); + } + + void operator()(VariableHandle input) { + PADDLE_ENFORCE(get_global_tape().HasBeenBackwarded(), + "optimization must happen after the backward"); + Tape temp_tape; + temp_tape.AddOp("sgd", + {{"Param", {input}}, + {"LearningRate", {learning_rate_}}, + {"Grad", {input->Grad()}}}, + {{"ParamOut", {input}}}, + {}); + temp_tape.Forward(); + } + + private: + VariableHandle learning_rate_; +}; +} +} diff --git a/paddle/contrib/tape/tape.cc b/paddle/contrib/tape/tape.cc new file mode 100644 index 0000000000..531499b6fe --- /dev/null +++ b/paddle/contrib/tape/tape.cc @@ -0,0 +1,265 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/contrib/tape/tape.h" + +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/dim.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/place.h" +#include "paddle/fluid/pybind/pybind.h" + +namespace paddle { +namespace tape { + +// borrowed from +// https://stackoverflow.com/questions/874134/find-if-string-ends-with-another-string-in-c +inline bool ends_with(std::string const &value, std::string const &ending) { + if (ending.size() > value.size()) return false; + return std::equal(ending.rbegin(), ending.rend(), value.rbegin()); +} + +std::ostream &operator<<(std::ostream &os, const framework::VarDesc &var_desc) { + os << var_desc.Name(); + os << "[" << var_desc.GetType() << "]"; + os << "[" << var_desc.GetDataType() << "]"; + os << "{"; + for (auto &i : var_desc.GetShape()) { + os << i << ","; + } + os << "}"; + return os; +} + +std::string to_string(const std::string &type, + const VariableHandleMap &in_vars, + const VariableHandleMap &out_vars, + const framework::AttributeMap &attrs) { + std::stringstream ss; + ss << type << " "; + for (auto ¶m_name : in_vars) { + for (auto &var : param_name.second) { + ss << param_name.first << ":(" << var->Desc() << ") "; + } + } + for (auto ¶m_name : out_vars) { + for (auto &var : param_name.second) { + ss << param_name.first << ":(" << var->Desc() << ") "; + } + } + return ss.str(); +} + +framework::OpDesc CreateOpDesc(const std::string &type, + const VariableHandleMap &in_vars, + const VariableHandleMap &out_vars, + const framework::AttributeMap &attrs) { + framework::VariableNameMap inputs; + for (auto ¶m_name : in_vars) { + for (auto &var : param_name.second) { + inputs[param_name.first].emplace_back(var->Name()); + } + } + framework::VariableNameMap outputs; + for (auto ¶m_name : out_vars) { + for (auto &var : param_name.second) { + outputs[param_name.first].emplace_back(var->Name()); + } + } + return framework::OpDesc(type, inputs, outputs, attrs); +} + +void InferShapeAndVarType(const std::string &type, + const VariableHandleMap &in_vars, + VariableHandleMap *out_vars, + const framework::AttributeMap &attrs) { + framework::OpDesc op_desc = CreateOpDesc(type, in_vars, *out_vars, attrs); + + // Create a temporary block for compile-time + framework::ProgramDesc program_desc; + framework::BlockDesc *block_desc = program_desc.MutableBlock(0); + PADDLE_ENFORCE(block_desc); + + for (auto ¶m_name : in_vars) { + for (auto &var : param_name.second) { + *block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto(); + } + } + for (auto ¶m_name : *out_vars) { + for (auto &var : param_name.second) { + *block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto(); + } + } + + LOG(INFO) << "- " << to_string(type, in_vars, *out_vars, attrs); + op_desc.InferShape(*block_desc); + op_desc.InferVarType(block_desc); + for (auto ¶m_name : *out_vars) { + for (auto &var : param_name.second) { + *var->MutableDesc()->Proto() = *block_desc->Var(var->Name())->Proto(); + } + } + LOG(INFO) << "+ " << to_string(type, in_vars, *out_vars, attrs); +} + +void Tape::AddOp(const std::string &type, + const VariableHandleMap &in_vars, + VariableHandleMap out_vars, + const framework::AttributeMap &attrs) { + InferShapeAndVarType(type, in_vars, &out_vars, attrs); + tape_.emplace_back(type, in_vars, out_vars, attrs); +} + +// Temporary Scope for Operator::Run() +class ScopeWrapper : public framework::Scope { + public: + ScopeWrapper(const VariableHandleMap &in_vars, + const VariableHandleMap &out_vars) { + for (auto &v : in_vars) { + for (auto &vv : v.second) { + if (!vars_.count(vv->Name())) { + vars_[vv->Name()].reset(vv->Var()); + } + } + } + for (auto &v : out_vars) { + for (auto &vv : v.second) { + if (!vars_.count(vv->Name())) { + vars_[vv->Name()].reset(vv->Var()); + } + } + } + } + + ~ScopeWrapper() { + for (auto &pair : vars_) { + pair.second.release(); + } + } +}; + +void Tape::Forward() { + LOG(INFO) << "Starting forward -------------------------"; + PADDLE_ENFORCE(!has_been_backwarded_); + while (current_position_ < tape_.size()) { + OpHandle &op = tape_[current_position_]; + + // Create Output Tensor, this is only necessary for OpWithKernel + for (auto ¶m2var : op.outputs_) { + for (auto &var : param2var.second) { + var->InitializeVariable(); + } + } + + framework::OpDesc op_desc = + CreateOpDesc(op.type_, op.inputs_, op.outputs_, op.attrs_); + ScopeWrapper scope(op.inputs_, op.outputs_); + framework::OpRegistry::CreateOp(op_desc)->Run(scope, platform::CPUPlace()); + current_position_++; + } + + LOG(INFO) << "Finishing forward -------------------------"; +} + +void Tape::Backward(VariableHandle target) { + PADDLE_ENFORCE(!has_been_backwarded_); + + Forward(); + + // TODO(tonyyang-svail): check output of last op is target + backward_tape_.reset(new Tape()); + + framework::AttributeMap attrs; + + // FIXME(tonyyang-svail): Need to infer_data_type + attrs["dtype"] = framework::proto::VarType::Type::VarType_Type_FP32; + attrs["shape"] = std::vector{1}; + attrs["value"] = 1.0f; + backward_tape_->AddOp( + "fill_constant", {}, {{"Out", {target->Grad()}}}, attrs); + + for (auto it = tape_.rbegin(); it != tape_.rend(); ++it) { + framework::OpDesc op_desc = + CreateOpDesc(it->type_, it->inputs_, it->outputs_, it->attrs_); + std::unordered_map grad_to_var; + std::vector> grad_op_descs = + framework::OpInfoMap::Instance() + .Get(op_desc.Type()) + .GradOpMaker()(op_desc, {}, &grad_to_var, {}); + + for (auto &op_desc : grad_op_descs) { + std::unordered_map name2var; + for (auto ¶m2vars : it->inputs_) { + for (auto &a : param2vars.second) { + name2var[a->Name()] = a; + } + } + for (auto ¶m2vars : it->outputs_) { + for (auto &a : param2vars.second) { + name2var[a->Name()] = a; + } + } + + VariableHandleMap in_vars; + VariableHandleMap out_vars; + std::map + loop_over{{&op_desc->Inputs(), &in_vars}, + {&op_desc->Outputs(), &out_vars}}; + for (auto &each : loop_over) { + auto &vmp = *each.first; + auto &vhm = *each.second; + for (auto &p2a : vmp) { + for (auto &argu : p2a.second) { + if (name2var.count(argu)) { + vhm[p2a.first].push_back(name2var[argu]); + } else { + PADDLE_ENFORCE(ends_with(argu, framework::kGradVarSuffix), + argu.c_str()); + std::string name = argu.substr( + 0, argu.size() - std::strlen(framework::kGradVarSuffix)); + PADDLE_ENFORCE(name2var.count(name), name.c_str()); + vhm[p2a.first].push_back(name2var[name]->Grad()); + } + } + } + } + + backward_tape_->AddOp( + op_desc->Type(), in_vars, out_vars, op_desc->GetAttrMap()); + } + + // TODO(tonyyang-svail): how to fill empty grad? + // TODO(tonyyang-svail): Sum var grad is necessary + } + + backward_tape_->Forward(); + has_been_backwarded_ = true; +} + +Tape &get_global_tape() { + static Tape T; + return T; +} + +void reset_global_tape() { get_global_tape() = Tape(); } +} +} diff --git a/paddle/contrib/tape/tape.h b/paddle/contrib/tape/tape.h new file mode 100644 index 0000000000..ed79de17a7 --- /dev/null +++ b/paddle/contrib/tape/tape.h @@ -0,0 +1,64 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#pragma once + +#include +#include +#include +#include + +#include "paddle/contrib/tape/variable.h" + +namespace paddle { +namespace tape { + +using VariableHandleMap = std::map>; + +struct OpHandle { + OpHandle(const std::string &type, + const VariableHandleMap &in_vars, + const VariableHandleMap &out_vars, + const framework::AttributeMap &attrs) + : type_(type), inputs_(in_vars), outputs_(out_vars), attrs_(attrs) {} + + std::string type_; + VariableHandleMap inputs_; + VariableHandleMap outputs_; + framework::AttributeMap attrs_; +}; + +class Tape { + public: + void AddOp(const std::string &type, + const VariableHandleMap &in_vars, + VariableHandleMap out_vars, + const framework::AttributeMap &attrs); + void Forward(); + void Backward(VariableHandle target); + + bool HasBeenBackwarded() { return has_been_backwarded_; } + + private: + bool has_been_backwarded_ = false; + size_t current_position_ = 0; + + std::vector tape_; + std::shared_ptr backward_tape_; +}; + +Tape &get_global_tape(); + +void reset_global_tape(); +} +} diff --git a/paddle/contrib/tape/test_tape.cc b/paddle/contrib/tape/test_tape.cc new file mode 100644 index 0000000000..e9bfd21a71 --- /dev/null +++ b/paddle/contrib/tape/test_tape.cc @@ -0,0 +1,61 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "gtest/gtest.h" +#include "paddle/contrib/tape/function.h" + +using namespace paddle::tape; + +TEST(Tape, TestMLP) { + LOG(INFO) << "TestMLP"; + Linear linear1(3, 3, "relu"); + Linear linear2(3, 3, "relu"); + Mean mean; + + SGD sgd(0.001); + + std::string initializer = "fill_constant"; + paddle::framework::AttributeMap attrs; + attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32; + attrs["shape"] = std::vector{3, 3}; + attrs["value"] = 1.0f; + Fill filler(initializer, attrs); + + for (int i = 0; i < 2; ++i) { + reset_global_tape(); + + VariableHandle input(new Variable("input")); + filler(input); + + auto loss = mean(linear2(linear1(input))); + + get_global_tape().Backward(loss); + + for (auto w : linear1.Params()) { + sgd(w); + } + for (auto w : linear2.Params()) { + sgd(w); + } + } +} + +int main(int argc, char** argv) { + std::vector places; + places.emplace_back(paddle::platform::CPUPlace()); + paddle::platform::DeviceContextPool::Init(places); + + testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/paddle/contrib/tape/variable.cc b/paddle/contrib/tape/variable.cc new file mode 100644 index 0000000000..5ec1612909 --- /dev/null +++ b/paddle/contrib/tape/variable.cc @@ -0,0 +1,33 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/contrib/tape/variable.h" + +namespace paddle { +namespace tape { + +void Variable::InitializeVariable() { + LOG(INFO) << "Initialzing " << desc_.Name() << " as " << desc_.GetType(); + framework::proto::VarType::Type var_type = desc_.GetType(); + if (var_type == framework::proto::VarType::LOD_TENSOR) { + var_.GetMutable(); + } else if (var_type == framework::proto::VarType::SELECTED_ROWS) { + var_.GetMutable(); + } else { + PADDLE_THROW("Variable type %d is not in [LOD_TENSOR, SELECTED_ROWS]", + var_type); + } +} +} +} diff --git a/paddle/contrib/tape/variable.h b/paddle/contrib/tape/variable.h new file mode 100644 index 0000000000..35c328e69c --- /dev/null +++ b/paddle/contrib/tape/variable.h @@ -0,0 +1,85 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#pragma once + +#include + +#include "paddle/fluid/framework/operator.h" // framework::kGradVarSuffix +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/variable.h" + +namespace paddle { +namespace tape { + +class Variable; +using VariableHandle = std::shared_ptr; + +/* + * Combination of + * framework::VarDesc desc_; + * framework::Variable var_; + */ +class Variable { + public: + Variable(const std::string pre_fix) + : desc_(pre_fix + std::to_string(count())) {} + + Variable(const std::string pre_fix, bool is_grad) + : desc_(pre_fix + (is_grad ? framework::kGradVarSuffix + : std::to_string(count()))) {} + + ~Variable() { LOG(INFO) << "Deleting " << Name(); } + + // Instantiate LoDTensor/SelectedRow + void InitializeVariable(); + + VariableHandle Grad() { + if (grad_.expired()) { + VariableHandle new_grad(new Variable(desc_.Name(), true)); + grad_ = new_grad; + return new_grad; + } else { + return VariableHandle(grad_); + } + } + + // Stochastic Gradient Descent with Momentum + // VariableHandle Momentum (); + + // void init(const std::string& initializer, + // const framework::AttributeMap& attrs); + + // void value() {}; + + const framework::VarDesc& Desc() const { return desc_; } + framework::VarDesc* MutableDesc() { return &desc_; } + + // TODO(tonyyang-svail): No need to expose name + std::string Name() const { return desc_.Name(); } + + framework::Variable* Var() { return &var_; } + + private: + int count() { + static int counter = 0; + return counter++; + } + + framework::VarDesc desc_; + framework::Variable var_; + + std::weak_ptr grad_; +}; +} +} diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index ed1e70c646..6286dda4a5 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -83,11 +83,16 @@ cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor) cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog) -cc_library(executor SRCS executor.cc DEPS op_registry device_context scope -framework_proto glog lod_rank_table feed_fetch_method) +if(WITH_DISTRIBUTE) + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr) + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) +else() + cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method) +endif() -cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor) +cc_library(parallel_executor SRCS parallel_executor.cc DEPS ssa_graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor) cc_library(prune SRCS prune.cc DEPS framework_proto) cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context) diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc index fd409ed4c0..f537e4b9e5 100644 --- a/paddle/fluid/framework/block_desc.cc +++ b/paddle/fluid/framework/block_desc.cc @@ -169,17 +169,13 @@ void BlockDesc::Flush() { } if (need_update_) { - auto &op_field = *this->desc_->mutable_ops(); - this->ClearPBOps(); - op_field.Reserve(static_cast(ops_.size())); + this->desc_->mutable_ops()->Clear(); for (auto &op_desc : ops_) { - op_field.AddAllocated(op_desc->Proto()); + this->desc_->mutable_ops()->Add()->CopyFrom(*op_desc->Proto()); } - auto &var_field = *this->desc_->mutable_vars(); - this->ClearPBVars(); - var_field.Reserve(static_cast(vars_.size())); + this->desc_->mutable_vars()->Clear(); for (auto &var_desc : vars_) { - var_field.AddAllocated(var_desc.second->Proto()); + this->desc_->mutable_vars()->Add()->CopyFrom(*var_desc.second->Proto()); } need_update_ = false; } @@ -200,7 +196,7 @@ BlockDesc::BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc) vars_[var_desc.name()].reset(new VarDesc(var_desc)); } for (const proto::OpDesc &op_desc : desc_->ops()) { - ops_.emplace_back(new OpDesc(op_desc, prog, this)); + ops_.emplace_back(new OpDesc(op_desc, this)); } } @@ -209,7 +205,7 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, : prog_(prog), desc_(desc) { need_update_ = true; for (auto &op : other.ops_) { - ops_.emplace_back(new OpDesc(*op->Proto(), prog, this)); + ops_.emplace_back(new OpDesc(*op, this)); } for (auto &it : other.vars_) { auto *var = new VarDesc(*it.second); @@ -217,22 +213,6 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, } } -void BlockDesc::ClearPBOps() { - auto ops = this->desc_->mutable_ops(); - while (!ops->empty()) { - // we do not own the OpDesc, so release the ownership. - ops->ReleaseLast(); - } -} - -void BlockDesc::ClearPBVars() { - auto vars = this->desc_->mutable_vars(); - while (!vars->empty()) { - // we do not own the VarDesc, so release the ownership. - vars->ReleaseLast(); - } -} - void BlockDesc::SetForwardBlockID(int32_t forward_block_id) { PADDLE_ENFORCE(!desc_->has_forward_block_idx(), "Parent block ID has been set to %d. Cannot set to %d", diff --git a/paddle/fluid/framework/block_desc.h b/paddle/fluid/framework/block_desc.h index 600601669c..ce48548418 100644 --- a/paddle/fluid/framework/block_desc.h +++ b/paddle/fluid/framework/block_desc.h @@ -41,11 +41,6 @@ class BlockDesc { BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, ProgramDesc *prog); - ~BlockDesc() { - this->ClearPBVars(); - this->ClearPBOps(); - } - int32_t ID() const { return desc_->idx(); } int32_t Parent() const { return desc_->parent_idx(); } @@ -105,7 +100,7 @@ class BlockDesc { size_t OpSize() const { return ops_.size(); } - OpDesc *Op(int idx) { return ops_.at(idx).get(); } + OpDesc *Op(int idx) const { return ops_.at(idx).get(); } void Flush(); @@ -113,10 +108,6 @@ class BlockDesc { ProgramDesc *Program() const { return this->prog_; } - private: - void ClearPBOps(); - void ClearPBVars(); - private: ProgramDesc *prog_; // not_own proto::BlockDesc *desc_; // not_own diff --git a/paddle/fluid/framework/data_layout.h b/paddle/fluid/framework/data_layout.h index 9c5e2cf7cc..b611bb77b4 100644 --- a/paddle/fluid/framework/data_layout.h +++ b/paddle/fluid/framework/data_layout.h @@ -27,6 +27,7 @@ enum class DataLayout { kNHWC = 0, kNCHW = 1, kAnyLayout = 2, + kMKLDNN = 3, // all layouts supported by MKLDNN internally }; inline DataLayout StringToDataLayout(const std::string& str) { @@ -41,6 +42,8 @@ inline DataLayout StringToDataLayout(const std::string& str) { return DataLayout::kNCHW; } else if (s == "ANYLAYOUT") { return DataLayout::kAnyLayout; + } else if (s == "MKLDNNLAYOUT") { + return DataLayout::kMKLDNN; } else { PADDLE_THROW("Unknown storage order string: %s", s); } @@ -54,8 +57,10 @@ inline std::string DataLayoutToString(const DataLayout& data_layout) { return "NCHW"; case DataLayout::kAnyLayout: return "ANY_LAYOUT"; + case DataLayout::kMKLDNN: + return "MKLDNNLAYOUT"; default: - PADDLE_THROW("unknown DataLayou %d", data_layout); + PADDLE_THROW("unknown DataLayout %d", data_layout); } } diff --git a/paddle/fluid/framework/data_layout_transform.cc b/paddle/fluid/framework/data_layout_transform.cc index 60ec60a427..5b8dfc57ba 100644 --- a/paddle/fluid/framework/data_layout_transform.cc +++ b/paddle/fluid/framework/data_layout_transform.cc @@ -16,6 +16,9 @@ #include #include "paddle/fluid/operators/math/math_function.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif namespace paddle { namespace framework { @@ -88,5 +91,85 @@ void TransDataLayout(const OpKernelType& kernel_type_for_var, out->set_layout(expected_kernel_type.data_layout_); } +#ifdef PADDLE_WITH_MKLDNN +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; + +void* GetDataFromTensor(const Tensor& tensor, mkldnn::memory::data_type type) { + switch (type) { + case mkldnn::memory::data_type::f32: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::s8: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::u8: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::s16: + return platform::to_void_cast(tensor.data()); + case mkldnn::memory::data_type::s32: + return platform::to_void_cast(tensor.data()); + default: + PADDLE_THROW("wrong mkldnn type provided"); + } +} +#endif + +void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var, + const OpKernelType& expected_kernel_type, + const Tensor& in, Tensor* out) { + auto in_layout = kernel_type_for_var.data_layout_; + auto out_layout = expected_kernel_type.data_layout_; + + PADDLE_ENFORCE( + in_layout == DataLayout::kMKLDNN && out_layout != DataLayout::kMKLDNN, + "TransDataLayoutFromMKLDNN only supports transform from MKLDNN to " + "non-MKLDNN"); + +#ifdef PADDLE_WITH_MKLDNN + PADDLE_ENFORCE(in.format() != memory::format::format_undef && + in.format() != memory::format::any, + "Input tensor should have specified memory format"); + + // Set default as NCHW in case not specified + out_layout = + out_layout == DataLayout::kAnyLayout ? DataLayout::kNCHW : out_layout; + + auto& pool = platform::DeviceContextPool::Instance(); + auto* dev_ctx = dynamic_cast( + pool.Get(expected_kernel_type.place_)); + auto& cpu_engine = dev_ctx->GetEngine(); + + std::vector in_tz = paddle::framework::vectorize2int(in.dims()); + std::vector out_tz = in_tz; + + memory::data_type in_type = ToMKLDNNDataType(in.type()); + PADDLE_ENFORCE(in_type != memory::data_type::data_undef, + "Input tensor type is not supported: ", in.type().name()); + memory::data_type out_type = in_type; + + memory::format in_format = + in_tz.size() == 2 ? memory::format::nc : in.format(); + memory::format out_format = + out_tz.size() == 2 ? memory::format::nc : ToMKLDNNFormat(out_layout); + + void* in_data = GetDataFromTensor(in, in_type); + + // output tensor has the same dims as input. Reorder don't change dims + out->Resize(in.dims()); + + auto out_data = out->mutable_data(expected_kernel_type.place_, in.type()); + + auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data); + auto out_memory = + memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data); + + platform::Reorder(in_memory, out_memory); + + out->set_layout(out_layout); + // reset format since the out tensor will be feed to non-MKLDNN OPkernel + out->set_format(memory::format::format_undef); +#endif +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/data_layout_transform.h b/paddle/fluid/framework/data_layout_transform.h index 06b638663d..2ba84ce57f 100644 --- a/paddle/fluid/framework/data_layout_transform.h +++ b/paddle/fluid/framework/data_layout_transform.h @@ -14,6 +14,7 @@ #pragma once +#include #include #include "paddle/fluid/framework/op_kernel_type.h" #include "paddle/fluid/framework/tensor.h" @@ -22,6 +23,50 @@ namespace paddle { namespace framework { +#ifdef PADDLE_WITH_MKLDNN +using MKLDNNFormat = mkldnn::memory::format; +using MKLDNNDataType = mkldnn::memory::data_type; + +inline MKLDNNFormat ToMKLDNNFormat(const DataLayout& layout) { + switch (layout) { + case DataLayout::kNHWC: + return MKLDNNFormat::nhwc; + case DataLayout::kNCHW: + return MKLDNNFormat::nchw; + default: + PADDLE_THROW("Fail to convert layout %s to MKLDNN format", + DataLayoutToString(layout)); + } +} + +inline DataLayout ToPaddleLayout(const MKLDNNFormat& format) { + switch (format) { + case MKLDNNFormat::nhwc: + return DataLayout::kNHWC; + case MKLDNNFormat::nchw: + return DataLayout::kNCHW; + default: + PADDLE_THROW("Fail to convert MKLDNN format to paddle layout"); + } +} + +inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) { + static const std::map dict{ + {std::type_index(typeid(float)), MKLDNNDataType::f32}, // NOLINT + {std::type_index(typeid(char)), MKLDNNDataType::s8}, // NOLINT + {std::type_index(typeid(unsigned char)), MKLDNNDataType::u8}, + {std::type_index(typeid(int16_t)), MKLDNNDataType::s16}, + {std::type_index(typeid(int32_t)), MKLDNNDataType::s32}}; + auto iter = dict.find(type); + if (iter != dict.end()) return iter->second; + return MKLDNNDataType::data_undef; +} +#endif + +void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var, + const OpKernelType& expected_kernel_type, + const Tensor& in, Tensor* out); + std::vector GetAxis(const DataLayout& from, const DataLayout& to); void TransDataLayout(const OpKernelType& kernel_type_for_var, diff --git a/paddle/fluid/framework/data_transform.cc b/paddle/fluid/framework/data_transform.cc index 9c277a27da..b8fcc92697 100644 --- a/paddle/fluid/framework/data_transform.cc +++ b/paddle/fluid/framework/data_transform.cc @@ -33,11 +33,34 @@ void DataTransform(const OpKernelType& expected_kernel_type, Tensor in; in.ShareDataWith(input_tensor); Tensor out; + DataLayout lin = kernel_type_for_var.data_layout_; + DataLayout lout = expected_kernel_type.data_layout_; // do layout transform - if (NeedTransformLayout(expected_kernel_type.data_layout_, - kernel_type_for_var.data_layout_)) { - TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out); + if (NeedTransformLayout(lout, lin)) { + if (lin == DataLayout::kMKLDNN || lout == DataLayout::kMKLDNN) { + PADDLE_ENFORCE( + !(lin == DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN), + "No layout transform needed between two MKLDNN OPKernels"); + + if (lin != DataLayout::kMKLDNN && lout == DataLayout::kMKLDNN) { +#ifdef PADDLE_WITH_MKLDNN + // Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel + // Just set layout/format. No real transform occur + out.ShareDataWith(input_tensor); + out.set_layout(DataLayout::kMKLDNN); + out.set_format(ToMKLDNNFormat(lin)); +#endif + } else { + // Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel + // Do transform via MKLDNN lib + TransDataLayoutFromMKLDNN(kernel_type_for_var, expected_kernel_type, in, + &out); + } + } else { + // Case3 - transfrom between Non-MKLDNN OPKernels + TransDataLayout(kernel_type_for_var, expected_kernel_type, in, &out); + } transformed = true; PassTensorData(&out, &in); } diff --git a/paddle/fluid/framework/data_type.cc b/paddle/fluid/framework/data_type.cc index b6b93cf422..60382faffb 100644 --- a/paddle/fluid/framework/data_type.cc +++ b/paddle/fluid/framework/data_type.cc @@ -28,6 +28,9 @@ struct DataTypeMap { }; static DataTypeMap* InitDataTypeMap(); +// C++11 removes the need for manual locking. Concurrent execution shall wait if +// a static local variable is already being initialized. +// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex static DataTypeMap& gDataTypeMap() { static DataTypeMap* g_data_type_map_ = InitDataTypeMap(); return *g_data_type_map_; diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index 1bcd8412eb..3c73b6cc55 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -7,26 +7,32 @@ cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base) cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph) +cc_library(ssa_graph_printer SRCS ssa_graph_printer.cc DEPS ssa_graph_builder) +cc_library(ssa_graph_checker SRCS ssa_graph_checker.cc DEPS ssa_graph_builder) cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows) if(WITH_GPU) - nv_library(nccl_all_reduce_op_handle SRCS nccl_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory - dynload_cuda) - set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle) + nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory + dynload_cuda variable_visitor) nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda) nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda) else() - set(multi_devices_graph_builder_deps) + cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory + variable_visitor) cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim) cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) endif() cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) +cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle - scale_loss_grad_op_handle rpc_op_handle ${multi_devices_graph_builder_deps} reduce_op_handle broadcast_op_handle) + scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle) + + +cc_library(ssa_graph_builder_factory SRCS ssa_graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer ssa_graph_checker) cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto) cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope @@ -36,5 +42,6 @@ cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_ha device_context broadcast_op_handle) cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory device_context gather_op_handle) +cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_executor.cc DEPS ssa_graph_executor) #cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory # device_context reduce_op_handle ) diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc similarity index 61% rename from paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc rename to paddle/fluid/framework/details/all_reduce_op_handle.cc index 95aa599cd3..b335d3a0d3 100644 --- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc @@ -11,46 +11,65 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. - -#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h" #include + +#include "paddle/fluid/framework/details/all_reduce_op_handle.h" +#include "paddle/fluid/framework/details/container_cast.h" #include "paddle/fluid/framework/details/reduce_and_gather.h" +#include "paddle/fluid/framework/details/variable_visitor.h" namespace paddle { namespace framework { namespace details { -NCCLAllReduceOpHandle::NCCLAllReduceOpHandle( - const std::vector &local_scopes, - const std::vector &places, - const platform::NCCLContextMap &ctxs) + +#ifdef PADDLE_WITH_CUDA +AllReduceOpHandle::AllReduceOpHandle(const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap *ctxs) : local_scopes_(local_scopes), places_(places), nccl_ctxs_(ctxs) { - for (auto &p : places_) { - this->dev_ctxes_[p] = nccl_ctxs_.DevCtx(p); + if (nccl_ctxs_) { + for (auto &p : places_) { + this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p); + } } } +#else +AllReduceOpHandle::AllReduceOpHandle(const std::vector &local_scopes, + const std::vector &places) + : local_scopes_(local_scopes), places_(places) {} +#endif -void NCCLAllReduceOpHandle::RunImpl() { - if (inputs_.size() == 1) { +void AllReduceOpHandle::RunImpl() { + if (NoDummyInputSize() == 1) { return; // No need to all reduce when GPU count = 1; } else { // Wait input done WaitInputVarGenerated(); - - auto &var_name = static_cast(this->inputs_[0])->name_; - int dtype = -1; - size_t numel = 0; + auto in_var_handles = DynamicCast(this->Inputs()); + auto out_var_handles = DynamicCast(this->Outputs()); + PADDLE_ENFORCE_EQ( + in_var_handles.size(), places_.size(), + "The NoDummyInputSize should be equal to the number of places."); + PADDLE_ENFORCE_EQ( + in_var_handles.size(), out_var_handles.size(), + "The NoDummyInputSize and NoDummyOutputSize should be equal."); std::vector lod_tensors; - for (size_t i = 0; i < local_scopes_.size(); ++i) { auto *s = local_scopes_[i]; auto &local_scope = *s->FindVar(kLocalExecScopeName)->Get(); - - auto &lod_tensor = local_scope.FindVar(var_name)->Get(); + auto &lod_tensor = + local_scope.FindVar(in_var_handles[i]->name_)->Get(); lod_tensors.emplace_back(&lod_tensor); + PADDLE_ENFORCE_EQ(in_var_handles[i]->name_, out_var_handles[i]->name_, + "The name of input and output should be equal."); } if (platform::is_gpu_place(lod_tensors[0]->place())) { +#ifdef PADDLE_WITH_CUDA + PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr."); + int dtype = -1; + size_t numel = 0; std::vector> all_reduce_calls; for (size_t i = 0; i < local_scopes_.size(); ++i) { auto &p = places_[i]; @@ -66,7 +85,7 @@ void NCCLAllReduceOpHandle::RunImpl() { } int dev_id = boost::get(p).device; - auto &nccl_ctx = nccl_ctxs_.at(dev_id); + auto &nccl_ctx = nccl_ctxs_->at(dev_id); auto stream = nccl_ctx.stream(); auto comm = nccl_ctx.comm_; all_reduce_calls.emplace_back([=] { @@ -81,22 +100,25 @@ void NCCLAllReduceOpHandle::RunImpl() { call(); } }); +#else + PADDLE_THROW("Not compiled with CUDA"); +#endif } else { // Special handle CPU only Operator's gradient. Like CRF auto &trg = *this->local_scopes_[0] ->FindVar(kLocalExecScopeName) ->Get() - ->Var() + ->FindVar(out_var_handles[0]->name_) ->GetMutable(); // Reduce All Tensor to trg in CPU ReduceLoDTensor func(lod_tensors, &trg); VisitDataType(ToDataType(lod_tensors[0]->type()), func); - for (size_t i = 0; i < local_scopes_.size(); ++i) { + for (size_t i = 1; i < local_scopes_.size(); ++i) { auto &scope = *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get(); auto &p = places_[i]; - auto *var = scope.FindVar(var_name); + auto *var = scope.FindVar(out_var_handles[i]->name_); auto *dev_ctx = dev_ctxes_[p]; RunAndRecordEvent(p, [&trg, var, dev_ctx, p] { @@ -109,7 +131,7 @@ void NCCLAllReduceOpHandle::RunImpl() { } } -std::string NCCLAllReduceOpHandle::Name() const { return "nccl_all_reduce"; } +std::string AllReduceOpHandle::Name() const { return "all_reduce"; } } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h b/paddle/fluid/framework/details/all_reduce_op_handle.h similarity index 68% rename from paddle/fluid/framework/details/nccl_all_reduce_op_handle.h rename to paddle/fluid/framework/details/all_reduce_op_handle.h index a0c321843e..fdd250b0d3 100644 --- a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h +++ b/paddle/fluid/framework/details/all_reduce_op_handle.h @@ -20,17 +20,23 @@ #include "paddle/fluid/framework/details/op_handle_base.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" +#ifdef PADDLE_WITH_CUDA #include "paddle/fluid/platform/nccl_helper.h" +#endif namespace paddle { namespace framework { namespace details { -struct NCCLAllReduceOpHandle : public OpHandleBase { - NCCLAllReduceOpHandle(const std::vector &local_scopes, - const std::vector &places, - const platform::NCCLContextMap &ctxs); - +struct AllReduceOpHandle : public OpHandleBase { +#ifdef PADDLE_WITH_CUDA + AllReduceOpHandle(const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap *ctxs); +#else + AllReduceOpHandle(const std::vector &local_scopes, + const std::vector &places); +#endif std::string Name() const override; // Delay and buffer nccl_all_reduce together can significantly increase @@ -41,9 +47,11 @@ struct NCCLAllReduceOpHandle : public OpHandleBase { void RunImpl() override; private: - const std::vector &local_scopes_; - const std::vector &places_; - const platform::NCCLContextMap &nccl_ctxs_; + std::vector local_scopes_; + std::vector places_; +#ifdef PADDLE_WITH_CUDA + const platform::NCCLContextMap *nccl_ctxs_; +#endif }; } // namespace details diff --git a/paddle/fluid/framework/details/broadcast_op_handle.h b/paddle/fluid/framework/details/broadcast_op_handle.h index 629aa00cb8..8036f756b6 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.h +++ b/paddle/fluid/framework/details/broadcast_op_handle.h @@ -59,8 +59,8 @@ struct BroadcastOpHandle : public OpHandleBase { void RunImpl() override; private: - const std::vector &local_scopes_; - const std::vector &places_; + std::vector local_scopes_; + std::vector places_; #ifdef PADDLE_WITH_CUDA const platform::NCCLContextMap *nccl_ctxs_; #endif diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index 91bdfe6134..64e83acb4d 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -14,6 +14,8 @@ #pragma once +#include + namespace paddle { namespace framework { namespace details { @@ -29,6 +31,8 @@ struct BuildStrategy { ReduceStrategy reduce_{ReduceStrategy::kAllReduce}; GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice}; + + std::string debug_graphviz_path_{""}; }; } // namespace details diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h index e8d510ec95..716d674fa2 100644 --- a/paddle/fluid/framework/details/execution_strategy.h +++ b/paddle/fluid/framework/details/execution_strategy.h @@ -20,8 +20,9 @@ namespace details { struct ExecutionStrategy { size_t num_threads_{0}; - bool use_event_{true}; + bool use_cuda_{true}; bool allow_op_delay_{false}; + size_t num_iteration_per_drop_scope_{100}; }; } // namespace details diff --git a/paddle/fluid/framework/details/fuse_vars_op_handle.cc b/paddle/fluid/framework/details/fuse_vars_op_handle.cc new file mode 100644 index 0000000000..018c9bff71 --- /dev/null +++ b/paddle/fluid/framework/details/fuse_vars_op_handle.cc @@ -0,0 +1,51 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/fuse_vars_op_handle.h" + +namespace paddle { +namespace framework { +namespace details { + +void FuseVarsOpHandle::RunImpl() { + WaitInputVarGenerated(place_); + + auto in_var_handles = DynamicCast(this->Inputs()); + auto out_var_handles = DynamicCast(this->Outputs()); + PADDLE_ENFORCE_EQ(in_var_handles.size(), 0); + PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), ""); + + auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); + + auto out_var_handle = out_var_handles[0]; + auto out_var = scope->Var(out_var_handle->name_); + + auto out_tensor = out_var->GetMutable(); + out_tensor->Resize({total_numel_}).mutable_data(this->place_, type_); + + int64_t s = 0; + for (size_t i = 1; i < out_var_handles.size(); ++i) { + auto out_name = out_var_handles[i]->name_; + auto out_t = scope->Var(out_name)->GetMutable(); + auto numel = this->inputs_numel_.at(out_name); + out_t->ShareDataWith(out_tensor->Slice(s, s + numel)); + s += numel; + } + this->RunAndRecordEvent([] {}); +} + +std::string FuseVarsOpHandle::Name() const { return "fuse vars"; } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fuse_vars_op_handle.h b/paddle/fluid/framework/details/fuse_vars_op_handle.h new file mode 100644 index 0000000000..140fb5bb49 --- /dev/null +++ b/paddle/fluid/framework/details/fuse_vars_op_handle.h @@ -0,0 +1,63 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include + +#include "paddle/fluid/framework/details/container_cast.h" +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace framework { +namespace details { + +struct FuseVarsOpHandle : public OpHandleBase { + public: + FuseVarsOpHandle(Scope *local_scope, const platform::Place &place, + const std::unordered_map &inputs_numel, + const std::type_index &var_type) + : local_scope_(local_scope), + place_(place), + inputs_numel_(inputs_numel), + type_(var_type) { + total_numel_ = 0; + for (auto in_numel : inputs_numel) { + PADDLE_ENFORCE_GT(in_numel.second, 0); + total_numel_ += in_numel.second; + } + } + + std::string Name() const override; + + bool IsMultiDeviceTransfer() override { return false; }; + + protected: + void RunImpl() override; + + private: + Scope *local_scope_; + const platform::Place place_; + const std::unordered_map inputs_numel_; + const std::type_index type_; + int64_t total_numel_; +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc index d8e711994c..78356cb1be 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc @@ -11,28 +11,22 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include #include +#include #include +#include + +#include "paddle/fluid/framework/details/all_reduce_op_handle.h" #include "paddle/fluid/framework/details/broadcast_op_handle.h" #include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" #include "paddle/fluid/framework/details/reduce_op_handle.h" #include "paddle/fluid/framework/details/rpc_op_handle.h" #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h" #include "paddle/fluid/framework/op_info.h" #include "paddle/fluid/framework/scope.h" -#ifdef PADDLE_WITH_CUDA -#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h" -#endif - -#include -#include - -DEFINE_string(ssa_graph_path, "/tmp/ssa_graph.dot", - "the ssa graph path only print with GLOG_v=10," - "default /tmp/graph.dot"); - namespace paddle { namespace framework { namespace details { @@ -92,7 +86,7 @@ std::vector MultiDevSSAGraphBuilder::FindDistTrainSendVars( for (auto *op : program.Block(0).AllOps()) { // TODO(Yancey1989): use a graceful method to find send op, // instead of the the hard code string - if (op->Type() == "send_vars") { + if (op->Type() == "send") { auto op_vars = op->InputArgumentNames(); send_vars.reserve(send_vars.size() + std::distance(op_vars.begin(), op_vars.end())); @@ -148,9 +142,9 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp( std::unique_ptr MultiDevSSAGraphBuilder::Build( const ProgramDesc &program) const { - std::unordered_map var_types; + std::unordered_map all_vars; for (auto *var : program.Block(0).AllVars()) { - var_types[var->Name()] = var->GetType(); + all_vars[var->Name()] = var; } auto graph = new SSAGraph(); @@ -167,12 +161,28 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( auto send_vars = FindDistTrainSendVars(program); auto recv_vars = FindDistTrainRecvVars(program); - size_t cur_device_id = 0; std::vector> var_name_on_devices; std::vector> bcast_var_name_set; var_name_on_devices.resize(places_.size()); bcast_var_name_set.resize(places_.size()); + size_t cur_device_id = 0; + std::vector balance_grads(places_.size(), 0); + + auto get_appropriate_dev = [&](std::string &g_name) -> size_t { + auto var_desc = all_vars.at(g_name); + PADDLE_ENFORCE_NOT_NULL(var_desc); + auto dim = framework::make_ddim(var_desc->GetShape()); + int64_t numel = framework::product(dim); + PADDLE_ENFORCE_GE(numel, 0); + auto smallest = + std::min_element(std::begin(balance_grads), std::end(balance_grads)); + size_t dev_id = + static_cast(std::distance(std::begin(balance_grads), smallest)); + balance_grads[dev_id] += numel; + return dev_id; + }; + bool is_forwarding = true; for (auto *op : program.Block(0).AllOps()) { if (boost::get( @@ -220,17 +230,17 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( switch (strategy_.reduce_) { case BuildStrategy::ReduceStrategy::kReduce: + cur_device_id = get_appropriate_dev(g_name); CreateReduceOp(&result, g_name, cur_device_id); var_name_on_devices[cur_device_id].emplace(g_name); bcast_var_name_set[cur_device_id].emplace(p_name); - cur_device_id = (cur_device_id + 1) % places_.size(); break; case BuildStrategy::ReduceStrategy::kAllReduce: - if (IsSparseGradient(var_types, g_name)) { + if (IsSparseGradient(all_vars, g_name)) { CreateReduceOp(&result, g_name, 0); CreateBroadcastOp(&result, g_name, 0); } else { - InsertNCCLAllReduceOp(&result, g_name); + InsertAllReduceOp(&result, g_name); } break; } @@ -260,24 +270,32 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( */ AddOutputToLeafOps(&result); - if (VLOG_IS_ON(10)) { - std::ofstream fout(FLAGS_ssa_graph_path); - PrintGraphviz(*graph, fout); - } - return std::unique_ptr(graph); } bool MultiDevSSAGraphBuilder::IsSparseGradient( - const std::unordered_map &var_types, + const std::unordered_map &all_vars, const std::string &og) const { - PADDLE_ENFORCE(var_types.count(og) != 0); - if (var_types.at(og) == proto::VarType::SELECTED_ROWS) { + PADDLE_ENFORCE(all_vars.count(og) != 0); + if (all_vars.at(og)->GetType() == proto::VarType::SELECTED_ROWS) { return true; } return false; } +void MultiDevSSAGraphBuilder::SetCommunicationContext( + OpHandleBase *op_handle, const platform::Place &p) const { +#ifdef PADDLE_WITH_CUDA + if (nccl_ctxs_ == nullptr) { + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); + } +#else + op_handle->SetDeviceContext(p, + platform::DeviceContextPool::Instance().Get(p)); +#endif +} + void MultiDevSSAGraphBuilder::CreateBroadcastOp(SSAGraph *result, const std::string &p_name, size_t src_dev_id) const { @@ -292,15 +310,12 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(SSAGraph *result, op_handle->AddInput(in); for (size_t i = 0; i < places_.size(); ++i) { - auto &vars = result->vars_.at(i).at(p_name); auto &p = places_[i]; + SetCommunicationContext(op_handle, p); + auto &vars = result->vars_.at(i).at(p_name); auto *out_var = new VarHandle(vars.size(), i, p_name, p); vars.emplace_back(out_var); op_handle->AddOutput(out_var); -#ifndef ADDLE_WITH_CUDA - op_handle->SetDeviceContext(p, - platform::DeviceContextPool::Instance().Get(p)); -#endif } } @@ -312,15 +327,19 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(SSAGraph *result, CreateOpHandleIOs(result, op, dev_id); } -void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp( - SSAGraph *result, const std::string &og) const { +void MultiDevSSAGraphBuilder::InsertAllReduceOp(SSAGraph *result, + const std::string &og) const { #ifdef PADDLE_WITH_CUDA result->ops_.emplace_back( - new NCCLAllReduceOpHandle(local_scopes_, places_, *nccl_ctxs_)); + new AllReduceOpHandle(local_scopes_, places_, nccl_ctxs_)); +#else + result->ops_.emplace_back(new AllReduceOpHandle(local_scopes_, places_)); +#endif auto *op_handle = result->ops_.back().get(); for (size_t i = 0; i < places_.size(); ++i) { auto &p = places_[i]; + SetCommunicationContext(op_handle, p); auto &vars = result->vars_[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); @@ -330,9 +349,6 @@ void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp( vars.emplace_back(var); op_handle->AddOutput(var); } -#else - PADDLE_ENFORCE("Not implemented"); -#endif } bool MultiDevSSAGraphBuilder::IsParameterGradientOnce( @@ -371,7 +387,9 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const { for (size_t i = 0; i < places_.size(); ++i) { // Insert ScaleCost OpHandle #ifdef PADDLE_WITH_CUDA - auto *communication_dev_ctx = nccl_ctxs_->DevCtx(places_[i]); + auto *communication_dev_ctx = + nccl_ctxs_ ? nccl_ctxs_->DevCtx(places_[i]) + : platform::DeviceContextPool::Instance().Get(places_[i]); #else auto *communication_dev_ctx = platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); @@ -416,12 +434,9 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(SSAGraph *result, auto *op_handle = result->ops_.back().get(); for (size_t i = 0; i < places_.size(); ++i) { - auto &vars = result->vars_[i][og]; -#ifndef PADDLE_WITH_CUDA auto &p = places_[i]; - op_handle->SetDeviceContext(p, - platform::DeviceContextPool::Instance().Get(p)); -#endif + SetCommunicationContext(op_handle, p); + auto &vars = result->vars_[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); op_handle->AddInput(prev_grad.get()); @@ -456,22 +471,21 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result, void MultiDevSSAGraphBuilder::CreateRPCOp(SSAGraph *result, const OpDesc &op) const { - auto &p = places_[0]; - auto *s = local_scopes_[0]; - result->ops_.emplace_back(new RPCOpHandle(op, s, p, op.Type())); + result->ops_.emplace_back( + new RPCOpHandle(op, local_scopes_[0], op.Type(), places_[0])); if (op.Type() == "send_barrier") { - ConnectOp(result, result->ops_.back().get(), "send_vars"); + ConnectOp(result, result->ops_.back().get(), "send"); } else if (op.Type() == "recv") { ConnectOp(result, result->ops_.back().get(), "send_barrier"); } else if (op.Type() == "fetch_barrier") { ConnectOp(result, result->ops_.back().get(), "recv"); - } else if (op.Type() == "send_vars") { + } else if (op.Type() == "send") { // do nothing } else { PADDLE_THROW( "rpc op should be in [" - "send_vars, send_barrier. recv, fetch_barrier]"); + "send, send_barrier. recv, fetch_barrier]"); } // TODO(Yancey1989): schedule rpc op on different place may diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h index e07597dbd8..78581755fe 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.h +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.h @@ -100,17 +100,20 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { const std::vector> &var_name_on_devices, const OpDesc &op) const; - void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const; + void InsertAllReduceOp(SSAGraph *result, const std::string &og) const; void CreateBroadcastOp(SSAGraph *result, const std::string &p_name, size_t src_dev_id) const; bool IsSparseGradient( - const std::unordered_map &var_types, + const std::unordered_map &all_vars, const std::string &og) const; private: BuildStrategy strategy_; + + void SetCommunicationContext(OpHandleBase *op_handle, + const platform::Place &p) const; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index 6b064650b4..f79565fe71 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -39,9 +39,9 @@ OpHandleBase::~OpHandleBase() { #endif } -void OpHandleBase::Run(bool use_event) { +void OpHandleBase::Run(bool use_cuda) { #ifdef PADDLE_WITH_CUDA - if (events_.empty() && use_event) { + if (events_.empty() && use_cuda) { for (auto &p : dev_ctxes_) { int dev_id = boost::get(p.first).device; PADDLE_ENFORCE(cudaSetDevice(dev_id)); @@ -50,7 +50,7 @@ void OpHandleBase::Run(bool use_event) { } } #else - PADDLE_ENFORCE(!use_event); + PADDLE_ENFORCE(!use_cuda); #endif RunImpl(); @@ -104,6 +104,16 @@ void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) { } } +size_t OpHandleBase::NoDummyInputSize() const { + size_t cnt = 0; + for (auto *in : inputs_) { + if (dynamic_cast(in) == nullptr) { + ++cnt; + } + } + return cnt; +} + bool OpHandleBase::NeedWait(VarHandleBase *in_var) { return in_var && in_var->generated_op_; } diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index 8f94206a87..fbd90a3296 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -36,7 +36,7 @@ class OpHandleBase { virtual std::string Name() const = 0; - void Run(bool use_event); + void Run(bool use_cuda); virtual void RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx); @@ -80,6 +80,8 @@ class OpHandleBase { const std::vector &Outputs() const { return outputs_; } + size_t NoDummyInputSize() const; + protected: void RunAndRecordEvent(const std::function &callback); diff --git a/paddle/fluid/framework/details/reduce_and_gather.h b/paddle/fluid/framework/details/reduce_and_gather.h index 2b95a28499..a6ffb37313 100644 --- a/paddle/fluid/framework/details/reduce_and_gather.h +++ b/paddle/fluid/framework/details/reduce_and_gather.h @@ -37,7 +37,9 @@ struct ReduceLoDTensor { PADDLE_ENFORCE_NE(t0.numel(), 0); dst_tensor_.Resize(t0.dims()); T *dst = dst_tensor_.mutable_data(platform::CPUPlace()); - std::copy(t0.data(), t0.data() + t0.numel(), dst); + if (dst != t0.data()) { + std::copy(t0.data(), t0.data() + t0.numel(), dst); + } for (size_t i = 1; i < src_tensors_.size(); ++i) { auto &t = *src_tensors_[i]; diff --git a/paddle/fluid/framework/details/reduce_op_handle.h b/paddle/fluid/framework/details/reduce_op_handle.h index c652a2f4eb..4d14334cdf 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.h +++ b/paddle/fluid/framework/details/reduce_op_handle.h @@ -32,8 +32,8 @@ namespace framework { namespace details { struct ReduceOpHandle : public OpHandleBase { - const std::vector &local_scopes_; - const std::vector &places_; + std::vector local_scopes_; + std::vector places_; #ifdef PADDLE_WITH_CUDA const platform::NCCLContextMap *nccl_ctxs_; diff --git a/paddle/fluid/framework/details/rpc_op_handle.cc b/paddle/fluid/framework/details/rpc_op_handle.cc index 7f4da4c01d..586465f99f 100644 --- a/paddle/fluid/framework/details/rpc_op_handle.cc +++ b/paddle/fluid/framework/details/rpc_op_handle.cc @@ -19,12 +19,12 @@ namespace framework { namespace details { RPCOpHandle::RPCOpHandle(const framework::OpDesc &op_desc, - const Scope *local_scope, const platform::Place &place, - const std::string &name) + const Scope *local_scope, const std::string &name, + const platform::Place &place) : op_(framework::OpRegistry::CreateOp(op_desc)), local_scope_(local_scope), - place_(place), - name_(name) {} + name_(name), + place_(place) {} void RPCOpHandle::RunImpl() { // TODO(wuyi): need further analysis whether wait VarDummyHandle. diff --git a/paddle/fluid/framework/details/rpc_op_handle.h b/paddle/fluid/framework/details/rpc_op_handle.h index d28b772172..ae38c7fe19 100644 --- a/paddle/fluid/framework/details/rpc_op_handle.h +++ b/paddle/fluid/framework/details/rpc_op_handle.h @@ -29,7 +29,7 @@ namespace details { struct RPCOpHandle : public OpHandleBase { RPCOpHandle(const framework::OpDesc& op_desc, const Scope* local_scope, - const platform::Place& place, const std::string& name); + const std::string& name, const platform::Place& place); std::string Name() const override; @@ -43,8 +43,8 @@ struct RPCOpHandle : public OpHandleBase { private: std::unique_ptr op_; const Scope* local_scope_; - const platform::Place& place_; const std::string name_; + platform::Place place_; }; } // namespace details diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc new file mode 100644 index 0000000000..eb4e7ec52f --- /dev/null +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc @@ -0,0 +1,76 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h" +#include +#include +#include "paddle/fluid/framework/executor.h" + +namespace paddle { +namespace framework { +namespace details { +ScopeBufferedSSAGraphExecutor::ScopeBufferedSSAGraphExecutor( + ExecutionStrategy strategy, std::vector local_scopes, + std::vector var_infos, std::vector places, + std::unique_ptr &&underlying_executor) + : strategy_(std::move(strategy)), + underlying_executor_(std::move(underlying_executor)), + local_scopes_(std::move(local_scopes)), + var_infos_(std::move(var_infos)), + places_(std::move(places)) {} + +FeedFetchList ScopeBufferedSSAGraphExecutor::Run( + const std::vector &fetch_tensors) { + if (drop_scope_counter_ == 0) { + // Create local scopes. + for (auto it = local_scopes_.rbegin(); it != local_scopes_.rend(); ++it) { + auto &scope = *it; + Scope &local_scope = scope->NewScope(); + *scope->Var(details::kLocalExecScopeName)->GetMutable() = + &local_scope; + + for (auto &info : var_infos_) { + if (scope->FindVar(info.name_) != nullptr) { + continue; + } + + if (info.persistable_) { // Persistable + InitializeVariable(scope->Var(info.name_), info.type_); + } else { + InitializeVariable(local_scope.Var(info.name_), info.type_); + } + } + } + } + + auto fetch_data = underlying_executor_->Run(fetch_tensors); + drop_scope_counter_ += 1; + if (!fetch_tensors.empty() || + drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) { + drop_scope_counter_ = 0; + // Wait All computational streams + for (auto p : places_) { + platform::DeviceContextPool::Instance().Get(p)->Wait(); + } + for (auto &scope : local_scopes_) { + auto &local_scope = + *scope->Var(details::kLocalExecScopeName)->GetMutable(); + scope->DeleteScope(local_scope); + } + } + return fetch_data; +} +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h new file mode 100644 index 0000000000..20df7a4722 --- /dev/null +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h @@ -0,0 +1,53 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include "paddle/fluid/framework/details/execution_strategy.h" +#include "paddle/fluid/framework/details/ssa_graph_executor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/place.h" +namespace paddle { +namespace framework { +namespace details { + +struct VariableInfo { + std::string name_; + proto::VarType::Type type_; + bool persistable_; +}; + +class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor { + public: + ScopeBufferedSSAGraphExecutor( + ExecutionStrategy strategy, std::vector local_scopes, + std::vector var_infos, std::vector places, + std::unique_ptr&& underlying_executor); + FeedFetchList Run(const std::vector& fetch_tensors) override; + + private: + size_t drop_scope_counter_{0}; + + ExecutionStrategy strategy_; + std::unique_ptr underlying_executor_; + std::vector local_scopes_; + std::vector var_infos_; + std::vector places_; +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_builder.cc b/paddle/fluid/framework/details/ssa_graph_builder.cc index 6a56752755..88a21f4887 100644 --- a/paddle/fluid/framework/details/ssa_graph_builder.cc +++ b/paddle/fluid/framework/details/ssa_graph_builder.cc @@ -11,8 +11,8 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. - #include "paddle/fluid/framework/details/ssa_graph_builder.h" +#include namespace paddle { namespace framework { @@ -73,64 +73,6 @@ void SSAGraphBuilder::CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle, op_handle->AddOutput(var); } -template -void IterAllVar(const SSAGraph &graph, Callback callback) { - for (auto &each : graph.vars_) { - for (auto &pair1 : each) { - for (auto &pair2 : pair1.second) { - callback(*pair2); - } - } - } - - for (auto &var : graph.dep_vars_) { - callback(*var); - } -} - -void SSAGraphBuilder::PrintGraphviz(const SSAGraph &graph, std::ostream &sout) { - size_t var_id = 0; - std::unordered_map vars; - - sout << "digraph G {\n"; - - IterAllVar(graph, [&](const VarHandleBase &var) { - auto *var_ptr = &var; - auto *var_handle_ptr = dynamic_cast(var_ptr); - auto *dummy_ptr = dynamic_cast(var_ptr); - - size_t cur_var_id = var_id++; - vars[var_ptr] = cur_var_id; - - if (var_handle_ptr) { - sout << "var_" << cur_var_id << " [label=\"" << var_handle_ptr->name_ - << "\\n" - << var_handle_ptr->place_ << "\\n" - << var_handle_ptr->version_ << "\"]" << std::endl; - } else if (dummy_ptr) { - sout << "var_" << cur_var_id << " [label=\"dummy\"]" << std::endl; - } - }); - - size_t op_id = 0; - for (auto &op : graph.ops_) { - std::string op_name = "op_" + std::to_string(op_id++); - sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]" - << std::endl; - for (auto in : op->Inputs()) { - std::string var_name = "var_" + std::to_string(vars[in]); - sout << var_name << " -> " << op_name << std::endl; - } - - for (auto out : op->Outputs()) { - std::string var_name = "var_" + std::to_string(vars[out]); - sout << op_name << " -> " << var_name << std::endl; - } - } - - sout << "}\n"; -} - void SSAGraphBuilder::AddOutputToLeafOps(SSAGraph *graph) { for (auto &op : graph->ops_) { if (!op->Outputs().empty()) { diff --git a/paddle/fluid/framework/details/ssa_graph_builder.h b/paddle/fluid/framework/details/ssa_graph_builder.h index 64e5d93081..5fc12a44b5 100644 --- a/paddle/fluid/framework/details/ssa_graph_builder.h +++ b/paddle/fluid/framework/details/ssa_graph_builder.h @@ -55,8 +55,6 @@ class SSAGraphBuilder { const platform::Place &place, size_t place_offset); static void AddOutputToLeafOps(SSAGraph *graph); - - static void PrintGraphviz(const SSAGraph &graph, std::ostream &sout); }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/ssa_graph_builder_factory.cc b/paddle/fluid/framework/details/ssa_graph_builder_factory.cc new file mode 100644 index 0000000000..b4b49d3de6 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_builder_factory.cc @@ -0,0 +1,50 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/ssa_graph_builder_factory.h" +#include +#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include "paddle/fluid/framework/details/ssa_graph_checker.h" +#include "paddle/fluid/framework/details/ssa_graph_printer.h" + +namespace paddle { +namespace framework { +namespace details { +std::unique_ptr SSAGraphBuilderFactory::Create() { + std::unique_ptr res( +#ifdef PADDLE_WITH_CUDA + new MultiDevSSAGraphBuilder(places_, loss_var_name_, param_names_, + local_scopes_, nccl_ctxs_, strategy_) +#else + new MultiDevSSAGraphBuilder(places_, loss_var_name_, param_names_, + local_scopes_, strategy_) +#endif + ); // NOLINT + + if (!strategy_.debug_graphviz_path_.empty()) { + std::unique_ptr fout( + new std::ofstream(strategy_.debug_graphviz_path_)); + PADDLE_ENFORCE(fout->good()); + std::unique_ptr graphviz_printer( + new GraphvizSSAGraphPrinter()); + res.reset(new SSAGraghBuilderWithPrinter( + std::move(fout), std::move(graphviz_printer), std::move(res))); + } + res.reset(new SSAGraghBuilderWithChecker(std::move(res))); + + return res; +} +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_builder_factory.h b/paddle/fluid/framework/details/ssa_graph_builder_factory.h new file mode 100644 index 0000000000..91a119de83 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_builder_factory.h @@ -0,0 +1,71 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include +#include +#include "paddle/fluid/framework/details/build_strategy.h" +#include "paddle/fluid/framework/details/ssa_graph_builder.h" +#include "paddle/fluid/platform/place.h" + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/nccl_helper.h" +#endif + +namespace paddle { +namespace framework { +class Scope; +namespace details { + +class SSAGraphBuilderFactory { + public: + SSAGraphBuilderFactory(const std::vector& places, + const std::string& loss_var_name, + const std::unordered_set& param_names, + const std::vector& local_scopes, + const BuildStrategy& strategy) + : places_(places), + loss_var_name_(loss_var_name), + param_names_(param_names), + local_scopes_(local_scopes), + strategy_(strategy) { +#ifdef PADDLE_WITH_CUDA + nccl_ctxs_ = nullptr; +#endif + } + +#ifdef PADDLE_WITH_CUDA + void SetNCCLContextMap(platform::NCCLContextMap* nccl_ctxs) { + nccl_ctxs_ = nccl_ctxs; + } +#endif + + std::unique_ptr Create(); + + private: + std::vector places_; + std::string loss_var_name_; + std::unordered_set param_names_; + std::vector local_scopes_; + BuildStrategy strategy_; + +#ifdef PADDLE_WITH_CUDA + platform::NCCLContextMap* nccl_ctxs_; +#endif +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_checker.cc b/paddle/fluid/framework/details/ssa_graph_checker.cc new file mode 100644 index 0000000000..da5428946e --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_checker.cc @@ -0,0 +1,87 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/ssa_graph.h" +#include +#include "paddle/fluid/framework/details/ssa_graph_checker.h" + +namespace paddle { +namespace framework { +namespace details { + +bool SSAGraghBuilderWithChecker::IsValidGraph(const SSAGraph *graph) const { + std::unordered_map pending_ops; + std::unordered_set pending_vars; + std::unordered_set ready_vars; + std::unordered_set ready_ops; + + auto insert_pending_var = [&](VarHandleBase *var) { + pending_vars.insert(var); + if (var->generated_op_ == nullptr) { + ready_vars.emplace(var); + } + }; + + for (auto &var_map : graph->vars_) { + for (auto &name_pair : var_map) { + for (auto &version_pair : name_pair.second) { + insert_pending_var(version_pair.get()); + } + } + } + + for (auto &var : graph->dep_vars_) { + insert_pending_var(var.get()); + } + + for (auto &op : graph->ops_) { + if (op->Inputs().empty()) { + ready_ops.insert(op.get()); + } else { + pending_ops.insert({op.get(), op.get()->NoDupInputSize()}); + } + } + + auto run_all_ops = [&](std::unordered_set &set) { + for (auto *op : set) { + for (auto out : op->Outputs()) { + ready_vars.emplace(out); + } + } + set.clear(); + }; + + while (!pending_vars.empty()) { + run_all_ops(ready_ops); + + if (ready_vars.empty()) { + return false; + } + + for (auto ready_var : ready_vars) { + pending_vars.erase(ready_var); + for (auto *op : ready_var->pending_ops_) { + auto &deps = --pending_ops[op]; + if (deps == 0) { + ready_ops.insert(op); + } + } + } + ready_vars.clear(); + } + return true; +} +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_checker.h b/paddle/fluid/framework/details/ssa_graph_checker.h new file mode 100644 index 0000000000..304b221e7e --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_checker.h @@ -0,0 +1,44 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/framework/details/ssa_graph_builder.h" + +namespace paddle { +namespace framework { +namespace details { +struct SSAGraph; + +class SSAGraghBuilderWithChecker : public SSAGraphBuilder { + public: + explicit SSAGraghBuilderWithChecker( + std::unique_ptr&& builder) + : builder_(std::move(builder)) {} + + std::unique_ptr Build(const ProgramDesc& program) const override { + auto graph = builder_->Build(program); + PADDLE_ENFORCE(IsValidGraph(graph.get())); + return graph; + } + + bool IsValidGraph(const SSAGraph* graph) const; + + private: + std::unique_ptr builder_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_executor.cc b/paddle/fluid/framework/details/ssa_graph_executor.cc index 8da6ca889b..09b97bd0d9 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/ssa_graph_executor.cc @@ -17,10 +17,6 @@ namespace paddle { namespace framework { namespace details { - -SSAGraphExecutor::SSAGraphExecutor(std::unique_ptr &&graph) - : graph_(std::move(graph)) {} - SSAGraphExecutor::~SSAGraphExecutor() {} } // namespace details diff --git a/paddle/fluid/framework/details/ssa_graph_executor.h b/paddle/fluid/framework/details/ssa_graph_executor.h index a8833b7388..9580860336 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.h +++ b/paddle/fluid/framework/details/ssa_graph_executor.h @@ -28,15 +28,11 @@ class SSAGraphExecutor { DISABLE_COPY_AND_ASSIGN(SSAGraphExecutor); public: - // Steal graph inside - explicit SSAGraphExecutor(std::unique_ptr &&graph); + SSAGraphExecutor() {} virtual ~SSAGraphExecutor(); virtual FeedFetchList Run(const std::vector &fetch_tensors) = 0; - - protected: - std::unique_ptr graph_; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/ssa_graph_printer.cc b/paddle/fluid/framework/details/ssa_graph_printer.cc new file mode 100644 index 0000000000..22a40ca4b2 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_printer.cc @@ -0,0 +1,83 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/details/ssa_graph_printer.h" +#include +#include "paddle/fluid/framework/details/ssa_graph.h" + +namespace paddle { +namespace framework { +namespace details { + +template +static inline void IterAllVar(const SSAGraph &graph, Callback callback) { + for (auto &each : graph.vars_) { + for (auto &pair1 : each) { + for (auto &pair2 : pair1.second) { + callback(*pair2); + } + } + } + + for (auto &var : graph.dep_vars_) { + callback(*var); + } +} + +void GraphvizSSAGraphPrinter::Print(const SSAGraph &graph, + std::ostream &sout) const { + size_t var_id = 0; + std::unordered_map vars; + + sout << "digraph G {\n"; + + IterAllVar(graph, [&](const VarHandleBase &var) { + auto *var_ptr = &var; + auto *var_handle_ptr = dynamic_cast(var_ptr); + auto *dummy_ptr = dynamic_cast(var_ptr); + + size_t cur_var_id = var_id++; + vars[var_ptr] = cur_var_id; + + if (var_handle_ptr) { + sout << "var_" << cur_var_id << " [label=\"" << var_handle_ptr->name_ + << "\\n" + << var_handle_ptr->place_ << "\\n" + << var_handle_ptr->version_ << "\"]" << std::endl; + } else if (dummy_ptr) { + sout << "var_" << cur_var_id << " [label=\"dummy\"]" << std::endl; + } + }); + + size_t op_id = 0; + for (auto &op : graph.ops_) { + std::string op_name = "op_" + std::to_string(op_id++); + sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]" + << std::endl; + for (auto in : op->Inputs()) { + std::string var_name = "var_" + std::to_string(vars[in]); + sout << var_name << " -> " << op_name << std::endl; + } + + for (auto out : op->Outputs()) { + std::string var_name = "var_" + std::to_string(vars[out]); + sout << op_name << " -> " << var_name << std::endl; + } + } + + sout << "}\n"; +} +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_printer.h b/paddle/fluid/framework/details/ssa_graph_printer.h new file mode 100644 index 0000000000..b4c9001378 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_printer.h @@ -0,0 +1,67 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include "paddle/fluid/framework/details/ssa_graph_builder.h" + +namespace paddle { +namespace framework { +namespace details { +struct SSAGraph; +class SSAGraphPrinter { + public: + virtual ~SSAGraphPrinter() {} + virtual void Print(const SSAGraph& graph, std::ostream& sout) const = 0; +}; + +class GraphvizSSAGraphPrinter : public SSAGraphPrinter { + public: + void Print(const SSAGraph& graph, std::ostream& sout) const override; +}; + +class SSAGraghBuilderWithPrinter : public SSAGraphBuilder { + public: + SSAGraghBuilderWithPrinter(std::ostream& sout, + std::unique_ptr&& printer, + std::unique_ptr&& builder) + : printer_(std::move(printer)), + builder_(std::move(builder)), + stream_ref_(sout) {} + + SSAGraghBuilderWithPrinter(std::unique_ptr&& sout, + std::unique_ptr&& printer, + std::unique_ptr&& builder) + : printer_(std::move(printer)), + builder_(std::move(builder)), + stream_ptr_(std::move(sout)), + stream_ref_(*stream_ptr_) {} + + std::unique_ptr Build(const ProgramDesc& program) const override { + auto graph = builder_->Build(program); + printer_->Print(*graph, stream_ref_); + return graph; + } + + private: + std::unique_ptr printer_; + std::unique_ptr builder_; + std::unique_ptr stream_ptr_; + std::ostream& stream_ref_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 815f739371..6c5098ce85 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -21,7 +21,7 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, std::unique_ptr &&graph) - : SSAGraphExecutor(std::move(graph)), + : graph_(std::move(graph)), pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_) : nullptr), local_scopes_(local_scopes), @@ -185,12 +185,15 @@ void ThreadedSSAGraphExecutor::InsertPendingVar( ready_vars->Push(var); } } + void ThreadedSSAGraphExecutor::RunOp( BlockingQueue *ready_var_q, details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { - VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); - op->Run(strategy_.use_event_); + if (VLOG_IS_ON(10)) { + VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); + } + op->Run(strategy_.use_cuda_); VLOG(10) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index 1f7f88d752..4a2075f1cc 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -51,6 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { details::OpHandleBase *op); private: + std::unique_ptr graph_; std::unique_ptr<::ThreadPool> pool_; std::vector local_scopes_; std::vector places_; diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 863053c32b..e15232a77b 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -20,10 +20,14 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" +#ifdef PADDLE_WITH_DISTRIBUTE +#include "paddle/fluid/operators/detail/grpc_client.h" +#endif #include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/profiler.h" DECLARE_bool(benchmark); +DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run"); namespace paddle { namespace framework { @@ -43,6 +47,14 @@ ExecutorPrepareContext::~ExecutorPrepareContext() { Executor::Executor(const platform::Place& place) : place_(place) {} +#ifdef PADDLE_WITH_DISTRIBUTE +void Executor::Complete() { + ::paddle::operators::detail::RPCClient::GetInstance< + ::paddle::operators::detail::GRPCClient>() + ->SendComplete(); +} +#endif + void InitializeVariable(Variable* var, proto::VarType::Type var_type) { if (var_type == proto::VarType::LOD_TENSOR) { var->GetMutable(); @@ -115,6 +127,7 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id, bool create_local_scope, bool create_vars) { platform::RecordBlock b(block_id); + if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc); auto ctx = Prepare(pdesc, block_id); RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars); } @@ -214,16 +227,18 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, const std::string& feed_holder_name, const std::string& fetch_holder_name) { platform::RecordBlock b(kProgramId); + if (FLAGS_use_mkldnn) EnableMKLDNN(program); bool has_feed_ops = has_feed_operators(program.Block(0), *feed_targets, feed_holder_name); bool has_fetch_ops = has_fetch_operators(program.Block(0), *fetch_targets, fetch_holder_name); ProgramDesc* copy_program = const_cast(&program); + std::unique_ptr unique_ptr_of_copy_program; if (!has_feed_ops || !has_fetch_ops) { - copy_program = std::unique_ptr(new ProgramDesc(program)).get(); + unique_ptr_of_copy_program.reset(new ProgramDesc(program)); + copy_program = unique_ptr_of_copy_program.get(); } - auto* global_block = copy_program->MutableBlock(0); if (!has_feed_ops) { @@ -315,8 +330,12 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } for (auto& op : ctx->ops_) { - VLOG(3) << place_ << " " << op->DebugStringEx(local_scope); + VLOG(4) << place_ << " " << op->DebugStringEx(local_scope); op->Run(*local_scope, place_); + // NOTE! Please do not delete this line, it's usefull because the debug + // string before and after op.run are different, after run the output + // will have right shape which is usefull for debug. + VLOG(3) << place_ << " " << op->DebugStringEx(local_scope); if (FLAGS_benchmark) { VLOG(2) << "Memory used after operator " + op->Type() + " running: " @@ -376,5 +395,19 @@ void Executor::RunPreparedContext( } } +void Executor::EnableMKLDNN(const ProgramDesc& program) { +#ifdef PADDLE_WITH_MKLDNN + VLOG(3) << "use_mkldnn=True"; + for (size_t bid = 0; bid < program.Size(); ++bid) { + auto* block = const_cast(program).MutableBlock(bid); + for (auto* op : block->AllOps()) { + if (op->HasAttr("use_mkldnn")) { + op->SetAttr("use_mkldnn", true); + } + } + } +#endif +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 0c3c23611d..67a0761dac 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -44,6 +44,13 @@ class Executor { explicit Executor(const platform::Place& place); +#ifdef PADDLE_WITH_DISTRIBUTE + /* + * Sending signal to pserver to mark current trainer stop. + */ + void Complete(); +#endif + /* @Brief * Runtime evaluation of the given ProgramDesc under certain Scope * @@ -81,6 +88,8 @@ class Executor { const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); + void EnableMKLDNN(const ProgramDesc& program); + private: const platform::Place place_; }; diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto index d35125fe8c..68fcc104d4 100644 --- a/paddle/fluid/framework/framework.proto +++ b/paddle/fluid/framework/framework.proto @@ -71,6 +71,7 @@ message OpProto { optional bool duplicable = 3 [ default = false ]; optional bool intermediate = 4 [ default = false ]; optional bool dispensable = 5 [ default = false ]; + optional string reuse = 6; } // AttrProto describes the C++ type Attribute. diff --git a/paddle/fluid/framework/init.cc b/paddle/fluid/framework/init.cc index 85beae775b..a1094976f6 100644 --- a/paddle/fluid/framework/init.cc +++ b/paddle/fluid/framework/init.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/string/piece.h" @@ -113,6 +114,9 @@ void InitDevices(bool init_p2p, const std::vector devices) { } places.emplace_back(platform::CPUPlace()); platform::DeviceContextPool::Init(places); +#ifndef PADDLE_WITH_MKLDNN + operators::math::SetNumThreads(1); +#endif } void InitGLOG(const std::string &prog_name) { diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index 09b67e5a17..f92769192c 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -103,7 +103,7 @@ void OpDesc::CopyFrom(const OpDesc &op_desc) { need_update_ = true; } -OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block) +OpDesc::OpDesc(const proto::OpDesc &desc, BlockDesc *block) : desc_(desc), need_update_(false) { // restore inputs_ int input_size = desc_.inputs_size(); diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index 1a330db7cc..a02d3e2691 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -33,13 +33,14 @@ class OpDesc { OpDesc(const std::string &type, const VariableNameMap &inputs, const VariableNameMap &outputs, const AttributeMap &attrs); - OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block); + OpDesc(const proto::OpDesc &desc, BlockDesc *block); explicit OpDesc(BlockDesc *block) : block_(block) {} OpDesc(const OpDesc &other, BlockDesc *block) { *this = other; block_ = block; + need_update_ = true; } void CopyFrom(const OpDesc &op_desc); diff --git a/paddle/fluid/framework/op_info.cc b/paddle/fluid/framework/op_info.cc index b99e82f8c4..f1261dee03 100644 --- a/paddle/fluid/framework/op_info.cc +++ b/paddle/fluid/framework/op_info.cc @@ -17,12 +17,11 @@ limitations under the License. */ namespace paddle { namespace framework { -static OpInfoMap* g_op_info_map = nullptr; - +// C++11 removes the need for manual locking. Concurrent execution shall wait if +// a static local variable is already being initialized. +// https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex OpInfoMap& OpInfoMap::Instance() { - if (g_op_info_map == nullptr) { - g_op_info_map = new OpInfoMap(); - } + static OpInfoMap* g_op_info_map = new OpInfoMap(); return *g_op_info_map; } } // namespace framework diff --git a/paddle/fluid/framework/op_kernel_type.h b/paddle/fluid/framework/op_kernel_type.h index fab20d75f5..f51a184e7b 100644 --- a/paddle/fluid/framework/op_kernel_type.h +++ b/paddle/fluid/framework/op_kernel_type.h @@ -87,7 +87,14 @@ inline std::string KernelTypeToString(const OpKernelType& kernel_key) { } inline bool NeedTransformLayout(const DataLayout& l, const DataLayout& r) { - return l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r; + bool ret = + (l != DataLayout::kAnyLayout && r != DataLayout::kAnyLayout && l != r); +#ifdef PADDLE_WITH_MKLDNN + // Layout transform needed for either non-MKLDNN to MKLDNN or vice versa + ret |= (l != DataLayout::kMKLDNN && r == DataLayout::kMKLDNN); + ret |= (l == DataLayout::kMKLDNN && r != DataLayout::kMKLDNN); +#endif + return ret; } inline bool TransFromNeeded(const OpKernelType& l, const OpKernelType& r) { diff --git a/paddle/fluid/framework/op_proto_maker.cc b/paddle/fluid/framework/op_proto_maker.cc index ae9f4efd44..001b5cb5a8 100644 --- a/paddle/fluid/framework/op_proto_maker.cc +++ b/paddle/fluid/framework/op_proto_maker.cc @@ -21,6 +21,7 @@ namespace framework { void OpProtoAndCheckerMaker::Validate() { validated_ = true; CheckNoDuplicatedInOutAttrs(); + CheckReuseVars(); } OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput( @@ -56,6 +57,24 @@ void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { } } +void OpProtoAndCheckerMaker::CheckReuseVars() { + std::unordered_set names; + for (auto& input : proto_->inputs()) { + names.insert(input.name()); + } + auto checker = [&](const std::string& name, const std::string& reused) { + PADDLE_ENFORCE( + names.count(reused), + "Output [%s] reuse Input [%s], but the input is not registered.", name, + reused); + }; + for (auto& output : proto_->outputs()) { + if (output.has_reuse()) { + checker(output.name(), output.reuse()); + } + } +} + void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto, OpAttrChecker* attr_checker) { proto_ = proto; diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h index 8493b9d8b3..92f86bb5de 100644 --- a/paddle/fluid/framework/op_proto_maker.h +++ b/paddle/fluid/framework/op_proto_maker.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once #include +#include + #include "glog/logging.h" #include "paddle/fluid/framework/attribute.h" #include "paddle/fluid/framework/framework.pb.h" @@ -64,6 +66,11 @@ class OpProtoAndCheckerMaker { var_->set_dispensable(true); return *this; } + + VariableBuilder &Reuse(const std::string &name) { + var_->set_reuse(name); + return *this; + } }; VariableBuilder AddInput(const std::string &name, const std::string &comment); @@ -89,6 +96,8 @@ class OpProtoAndCheckerMaker { void CheckNoDuplicatedInOutAttrs(); void Validate(); + void CheckReuseVars(); + proto::OpProto *proto_; OpAttrChecker *op_checker_; bool validated_{false}; diff --git a/paddle/fluid/framework/op_proto_maker_test.cc b/paddle/fluid/framework/op_proto_maker_test.cc index a8030d377f..58f70cb39c 100644 --- a/paddle/fluid/framework/op_proto_maker_test.cc +++ b/paddle/fluid/framework/op_proto_maker_test.cc @@ -47,3 +47,23 @@ TEST(ProtoMaker, DuplicatedInOut) { ASSERT_THROW(proto_maker(&op_proto, &op_checker), paddle::platform::EnforceNotMet); } + +class TestInplaceProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { + public: + void Make() { + AddInput("X", "input of test op"); + AddOutput("XOut", "output of test op").Reuse("X"); + AddOutput("NoOut", "output of test op").Reuse("NotExists"); + } +}; + +TEST(ProtoMaker, InplaceOutput) { + paddle::framework::proto::OpProto op_proto; + paddle::framework::OpAttrChecker op_checker; + TestInplaceProtoMaker proto_maker; + ASSERT_THROW(proto_maker(&op_proto, &op_checker), + paddle::platform::EnforceNotMet); + // proto_maker(&op_proto, &op_checker); + // proto_maker.Make(); + // ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); +} diff --git a/paddle/fluid/framework/op_registry.h b/paddle/fluid/framework/op_registry.h index 748317438b..43ab227a94 100644 --- a/paddle/fluid/framework/op_registry.h +++ b/paddle/fluid/framework/op_registry.h @@ -83,8 +83,14 @@ struct OpKernelRegistrarFunctor { void operator()(const char* op_type, const char* library_type) const { using T = typename KERNEL_TYPE::ELEMENT_TYPE; + std::string library(library_type); + std::string data_layout = "ANYLAYOUT"; + if (library == "MKLDNN") { + data_layout = "MKLDNNLAYOUT"; + } OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(), - DataLayout::kAnyLayout, StringToLibraryType(library_type)); + StringToDataLayout(data_layout), + StringToLibraryType(library_type)); OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KERNEL_TYPE); constexpr auto size = std::tuple_size>::value; @@ -99,7 +105,8 @@ struct OpKernelRegistrarFunctor { void operator()(const char* op_type, const char* library_type) const {} }; -// User can register many kernel in one place. The data type could be different. +// User can register many kernel in one place. The data type could be +// different. template class OpKernelRegistrar : public Registrar { public: @@ -149,15 +156,15 @@ class OpKernelRegistrar : public Registrar { /** * Macro to register OperatorKernel. */ -#define REGISTER_OP_KERNEL(op_type, LIBRARY_TYPE, place_class, ...) \ +#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...) \ STATIC_ASSERT_GLOBAL_NAMESPACE( \ - __reg_op_kernel_##op_type##_##LIBRARY_TYPE##__, \ + __reg_op_kernel_##op_type##_##library_type##__, \ "REGISTER_OP_KERNEL must be called in global namespace"); \ static ::paddle::framework::OpKernelRegistrar \ - __op_kernel_registrar_##op_type##_##LIBRARY_TYPE##__(#op_type, \ - #LIBRARY_TYPE); \ - int TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE() { \ - __op_kernel_registrar_##op_type##_##LIBRARY_TYPE##__.Touch(); \ + __op_kernel_registrar_##op_type##_##library_type##__(#op_type, \ + #library_type); \ + int TouchOpKernelRegistrar_##op_type##_##library_type() { \ + __op_kernel_registrar_##op_type##_##library_type##__.Touch(); \ return 0; \ } diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index f87d552149..122ee1dab3 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -69,6 +69,19 @@ static DDim GetDims(const Scope& scope, const std::string& name, } } +static int GetRowSize(const Scope& scope, const std::string& name) { + Variable* var = scope.FindVar(name); + if (var == nullptr) { + return -1; + } + + if (var->IsType()) { + return var->Get().rows().size(); + } + + return -1; +} + static LoD GetLoD(const Scope& scope, const std::string& name) { Variable* var = scope.FindVar(name); auto default_lod = LoD({{}}); @@ -85,6 +98,7 @@ static LoD GetLoD(const Scope& scope, const std::string& name) { } void OperatorBase::Run(const Scope& scope, const platform::Place& place) { + VLOG(10) << "- " << DebugStringEx(&scope); if (platform::is_gpu_place(place)) { #ifndef PADDLE_WITH_CUDA PADDLE_THROW("Cannot run operator on place %s", place); @@ -94,6 +108,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { #endif } RunImpl(scope, place); + VLOG(10) << "+ " << DebugStringEx(&scope); } bool OperatorBase::HasInputs(const std::string& name) const { @@ -153,6 +168,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { for (size_t i = 0; i < input.second.size(); ++i) { ss << input.second[i]; if (scope) { + int row_size = GetRowSize(*scope, input.second[i]); + if (row_size >= 0) { + ss << "[row_size=" << row_size << "]"; + } ss << "[" << GetDims(*scope, input.second[i], true) << "]"; ss << "(" << GetLoD(*scope, input.second[i]) << ")"; } @@ -173,6 +192,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { for (size_t i = 0; i < output.second.size(); ++i) { ss << output.second[i]; if (scope) { + int row_size = GetRowSize(*scope, output.second[i]); + if (row_size >= 0) { + ss << "[row_size=" << row_size << "]"; + } ss << "[" << GetDims(*scope, output.second[i], true) << "]"; ss << "(" << GetLoD(*scope, output.second[i]) << ")"; } @@ -293,6 +316,38 @@ static Tensor* GetMutableTensorFromVar(Variable* var) { } } +bool ExecutionContext::HasInput(const std::string& name) const { + if (!op_.HasInputs(name)) { + return false; + } + auto& ins = Inputs(name); + size_t length = ins.size(); + if (length == 0) { + return false; + } + PADDLE_ENFORCE_EQ(length, 1UL, + "Input %s should not have more than one inputs", name); + auto arg = ins[0]; + auto* var = arg == kEmptyVarName ? nullptr : scope_.FindVar(arg); + return var != nullptr; +} + +bool ExecutionContext::HasOutput(const std::string& name) const { + if (!op_.HasOutputs(name)) { + return false; + } + auto& outs = Outputs(name); + size_t length = outs.size(); + if (length == 0) { + return false; + } + PADDLE_ENFORCE_EQ(length, 1UL, + "Output %s should not have more than one inputs", name); + auto arg = outs[0]; + auto* var = arg == kEmptyVarName ? nullptr : scope_.FindVar(arg); + return var != nullptr; +} + template <> const Tensor* ExecutionContext::Input(const std::string& name) const { auto* var = InputVar(name); @@ -444,10 +499,25 @@ class RuntimeInferShapeContext : public InferShapeContext { auto* out_tensor = out_var->GetMutable(); out_tensor->set_lod(in_tensor.lod()); - // TODO(dzhwinter) : reuse ShareLoD in most operators. - // Need to call ShareLayout explicitly in sequence related ops. - // Shall we have a better method to shared info between in/out Tensor? - out_tensor->set_layout(in_tensor.layout()); +// TODO(dzhwinter) : reuse ShareLoD in most operators. +// Need to call ShareLayout explicitly in sequence related ops. +// Shall we have a better method to shared info between in/out Tensor? +#ifdef PADDLE_WITH_MKLDNN + // Fix me: ugly workaround below + // Correct solution: + // set_layout() should NOT be called here (i.e. ShareLoD). Instead, + // layout of output tensor should be set "manually" in Compute() + // of each OPKernel. The reason layout should NOT be shared between + // input and output "automatically" (now by InferShape()->ShareLoD()) + // is that layout transform may occur after InferShape(). + // Workaround: + // Skip set_layout() when input layout is kMKLDNN + // This is to avoid kMKLDNN is populated wrongly into a non-MKLDNN + // OPKernel. In all MKLDNN OPkernel, set_layout(kMKLDNN) should be called + // in Compute() + if (in_tensor.layout() != DataLayout::kMKLDNN) +#endif + out_tensor->set_layout(in_tensor.layout()); } void ShareLayout(const std::string& in, const std::string& out, size_t i = 0, @@ -646,8 +716,10 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType( } if (t != nullptr) { int tmp = static_cast(ToDataType(t->type())); - PADDLE_ENFORCE(tmp == data_type || data_type == -1, - "DataType of Paddle Op %s must be the same.", Type()); + PADDLE_ENFORCE( + tmp == data_type || data_type == -1, + "DataType of Paddle Op %s must be the same. Get %d != %d", Type(), + data_type, tmp); data_type = tmp; } } @@ -665,7 +737,8 @@ OpKernelType OperatorWithKernel::GetExpectedKernelType( OpKernelType OperatorWithKernel::GetKernelTypeForVar( const std::string& var_name, const Tensor& tensor, const OpKernelType& expected_kernel_type) const { - return OpKernelType(expected_kernel_type.data_type_, tensor.place()); + return OpKernelType(expected_kernel_type.data_type_, tensor.place(), + tensor.layout()); } } // namespace framework diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 2f480e00c1..b1d75d0d0f 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -191,9 +191,9 @@ class ExecutionContext { return op_.Attr(name); } - bool HasInput(const std::string& name) const { return op_.HasInputs(name); } + bool HasInput(const std::string& name) const; - bool HasOutput(const std::string& name) const { return op_.HasOutputs(name); } + bool HasOutput(const std::string& name) const; size_t InputSize(const std::string& name) const { return op_.Inputs(name).size(); diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 50c3468d55..9406c6155d 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -22,7 +22,8 @@ limitations under the License. */ #include "paddle/fluid/platform/nccl_helper.h" #endif -#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h" +#include "paddle/fluid/framework/details/ssa_graph_builder_factory.h" #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" #include "paddle/fluid/platform/profiler.h" @@ -42,9 +43,8 @@ class ParallelExecutorPrivate { #ifdef PADDLE_WITH_CUDA std::unique_ptr nccl_ctxs_; #endif - - std::vector> var_types_; - bool own_local_scope; + bool own_local_scope_; + bool use_cuda_; }; std::vector &ParallelExecutor::GetLocalScopes() { @@ -61,65 +61,78 @@ ParallelExecutor::ParallelExecutor( size_t num_trainers, size_t trainer_id) : member_(new ParallelExecutorPrivate(places)) { member_->global_scope_ = scope; + member_->use_cuda_ = exec_strategy.use_cuda_; // Step 1. Bcast the params to devs. // Create local scopes if (local_scopes.empty()) { - member_->own_local_scope = true; + member_->own_local_scope_ = true; member_->local_scopes_.emplace_back(member_->global_scope_); for (size_t i = 1; i < member_->places_.size(); ++i) { member_->local_scopes_.emplace_back(&scope->NewScope()); } } else { - member_->own_local_scope = false; + member_->own_local_scope_ = false; PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size()); for (size_t i = 0; i < member_->places_.size(); ++i) { member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope()); } } + if (member_->use_cuda_) { // Bcast Parameters to all GPUs #ifdef PADDLE_WITH_CUDA - auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME); - ncclUniqueId *nccl_id = nullptr; - if (nccl_id_var != nullptr) { - nccl_id = nccl_id_var->GetMutable(); - } - member_->nccl_ctxs_.reset(new platform::NCCLContextMap( - member_->places_, nccl_id, num_trainers, trainer_id)); + auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME); + ncclUniqueId *nccl_id = nullptr; + if (nccl_id_var != nullptr) { + nccl_id = nccl_id_var->GetMutable(); + } + member_->nccl_ctxs_.reset(new platform::NCCLContextMap( + member_->places_, nccl_id, num_trainers, trainer_id)); +#else + PADDLE_THROW("Not compiled with CUDA"); #endif - if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 && - local_scopes.empty()) { // Is CUDA + } + + if (member_->local_scopes_.size() != 1 && local_scopes.empty()) { BCastParamsToGPUs(bcast_vars); } -// Startup Program has been run. All local scopes has correct parameters. + // Startup Program has been run. All local scopes has correct parameters. -// Step 2. Convert main_program to SSA form and dependency graph. Also, insert -// ncclOp -#ifdef PADDLE_WITH_CUDA - details::MultiDevSSAGraphBuilder builder( + // Step 2. Create vars in each scope; + std::vector var_infos; + for (auto *var : main_program.Block(0).AllVars()) { + var_infos.emplace_back(); + var_infos.back().name_ = var->Name(); + var_infos.back().type_ = var->GetType(); + var_infos.back().persistable_ = var->Persistable(); + } + + // Step 3. Convert main_program to SSA form and dependency graph. Also, insert + // ncclOp + + details::SSAGraphBuilderFactory builder_factory( member_->places_, loss_var_name, params, member_->local_scopes_, - member_->nccl_ctxs_.get(), build_strategy); + build_strategy); + if (member_->use_cuda_) { +#ifdef PADDLE_WITH_CUDA + builder_factory.SetNCCLContextMap(member_->nccl_ctxs_.get()); #else - details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name, - params, member_->local_scopes_, - build_strategy); + PADDLE_THROW("Not compiled with CUDA"); #endif - auto graph = builder.Build(main_program); + } member_->executor_.reset(new details::ThreadedSSAGraphExecutor( - exec_strategy, member_->local_scopes_, places, std::move(graph))); + exec_strategy, member_->local_scopes_, places, + builder_factory.Create()->Build(main_program))); - // Step 3. Create vars in each scope; - for (auto *var : main_program.Block(0).AllVars()) { - member_->var_types_.emplace_back(var->Name(), var->GetType(), - var->Persistable()); - } + member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor( + exec_strategy, member_->local_scopes_, std::move(var_infos), + member_->places_, std::move(member_->executor_))); } void ParallelExecutor::BCastParamsToGPUs( const std::unordered_set &vars) const { -#ifdef PADDLE_WITH_CUDA auto *main_scope = member_->local_scopes_[0]; for (auto &var : vars) { @@ -131,9 +144,10 @@ void ParallelExecutor::BCastParamsToGPUs( auto &main_tensor = main_var->Get(); auto &dims = main_tensor.dims(); if (paddle::platform::is_gpu_place(main_tensor.place())) { +#ifdef PADDLE_WITH_CUDA + std::vector buffers; size_t numel = main_tensor.numel(); ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type()); - platform::NCCLGroupGuard guard; for (size_t i = 0; i < member_->places_.size(); ++i) { auto place = member_->places_[i]; void *buffer; @@ -145,10 +159,24 @@ void ParallelExecutor::BCastParamsToGPUs( t->Resize(dims); buffer = t->mutable_data(place, main_tensor.type()); } - auto &nccl_ctx = member_->nccl_ctxs_->at(place); - platform::dynload::ncclBcast(buffer, numel, data_type, 0, - nccl_ctx.comm_, nccl_ctx.stream()); + buffers.push_back(buffer); } + + PADDLE_ENFORCE_EQ(member_->places_.size(), buffers.size(), + "variables' buffer size to bcast NOT equal to places"); + { + platform::NCCLGroupGuard guard; + for (size_t i = 0; i < member_->places_.size(); ++i) { + auto &nccl_ctx = member_->nccl_ctxs_->at(member_->places_[i]); + platform::dynload::ncclBcast(buffers[i], numel, data_type, 0, + nccl_ctx.comm_, nccl_ctx.stream()); + } + member_->nccl_ctxs_->WaitAll(); + } + +#else + PADDLE_THROW("Not compiled with CUDA"); +#endif } else { platform::CPUPlace cpu; for (size_t i = 1; i < member_->places_.size(); ++i) { @@ -159,52 +187,15 @@ void ParallelExecutor::BCastParamsToGPUs( paddle::framework::TensorCopy(main_tensor, cpu, t); } } - member_->nccl_ctxs_->WaitAll(); } -#else - PADDLE_THROW("Not compiled with CUDA"); -#endif } void ParallelExecutor::Run(const std::vector &fetch_tensors, const std::string &fetched_var_name) { platform::RecordBlock b(0); - // Create local scopes. - for (auto it = member_->local_scopes_.rbegin(); - it != member_->local_scopes_.rend(); ++it) { - auto &scope = *it; - Scope &local_scope = scope->NewScope(); - *scope->Var(details::kLocalExecScopeName)->GetMutable() = - &local_scope; - - for (auto &name_type_pair : member_->var_types_) { - if (scope->FindVar(std::get<0>(name_type_pair)) != nullptr) { - continue; - } - - if (std::get<2>(name_type_pair)) { // Persistable - InitializeVariable(scope->Var(std::get<0>(name_type_pair)), - std::get<1>(name_type_pair)); - } else { - InitializeVariable(local_scope.Var(std::get<0>(name_type_pair)), - std::get<1>(name_type_pair)); - } - } - } - auto fetch_data = member_->executor_->Run(fetch_tensors); *member_->global_scope_->Var(fetched_var_name)->GetMutable() = fetch_data; - - // Wait All computational streams - for (auto p : member_->places_) { - platform::DeviceContextPool::Instance().Get(p)->Wait(); - } - for (auto &scope : member_->local_scopes_) { - auto &local_scope = - *scope->Var(details::kLocalExecScopeName)->GetMutable(); - scope->DeleteScope(local_scope); - } } void ParallelExecutor::FeedTensorsIntoLocalScopes( @@ -242,7 +233,7 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes( } ParallelExecutor::~ParallelExecutor() { - if (member_->own_local_scope) { + if (member_->own_local_scope_) { for (size_t i = 1; i < member_->local_scopes_.size(); ++i) { member_->global_scope_->DeleteScope(member_->local_scopes_[i]); } diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc index 64fb028f83..1e01a6e900 100644 --- a/paddle/fluid/framework/program_desc.cc +++ b/paddle/fluid/framework/program_desc.cc @@ -51,12 +51,15 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) { auto *block = desc_.mutable_blocks(i); blocks_.emplace_back(new BlockDesc(*o.blocks_[i], block, this)); } - for (auto &block : blocks_) { - for (auto *op : block->AllOps()) { - for (const auto &attr : op->Proto()->attrs()) { - if (attr.type() == proto::AttrType::BLOCK) { - size_t blk_idx = attr.block_idx(); - op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); + for (size_t block_id = 0; block_id < blocks_.size(); ++block_id) { + auto all_ops = blocks_[block_id]->AllOps(); + for (size_t op_id = 0; op_id < all_ops.size(); ++op_id) { + auto &op = all_ops[op_id]; + for (const std::string &attr_name : op->AttrNames()) { + if (op->GetAttrType(attr_name) == proto::AttrType::BLOCK) { + int sub_block_id = + o.Block(block_id).Op(op_id)->GetBlockAttr(attr_name); + op->SetBlockAttr(attr_name, MutableBlock(sub_block_id)); } } } @@ -86,6 +89,16 @@ ProgramDesc::ProgramDesc(const std::string &binary_str) { for (auto &block_desc : *desc_.mutable_blocks()) { blocks_.emplace_back(new BlockDesc(this, &block_desc)); } + for (auto &block : blocks_) { + for (auto *op : block->AllOps()) { + for (const auto &attr : op->Proto()->attrs()) { + if (attr.type() == proto::AttrType::BLOCK) { + size_t blk_idx = attr.block_idx(); + op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); + } + } + } + } } const std::vector ProgramDesc::GetFeedTargetNames() { diff --git a/paddle/fluid/framework/reader.cc b/paddle/fluid/framework/reader.cc index 76126f3dc6..0b36f1116d 100644 --- a/paddle/fluid/framework/reader.cc +++ b/paddle/fluid/framework/reader.cc @@ -25,8 +25,10 @@ void FileReader::ReadNext(std::vector *out) { if (out->empty()) { return; } + + PADDLE_ENFORCE_EQ(out->size(), dims_.size()); for (size_t i = 0; i < dims_.size(); ++i) { - auto &actual = out->at(i).dims(); + auto &actual = (*out)[i].dims(); auto &expect = dims_[i]; PADDLE_ENFORCE_EQ(actual.size(), expect.size()); diff --git a/paddle/fluid/framework/reader.h b/paddle/fluid/framework/reader.h index 3a413941df..64d4ceab62 100644 --- a/paddle/fluid/framework/reader.h +++ b/paddle/fluid/framework/reader.h @@ -35,14 +35,15 @@ class ReaderBase { class DecoratedReader : public ReaderBase { public: - explicit DecoratedReader(ReaderBase* reader) : ReaderBase(), reader_(reader) { + explicit DecoratedReader(const std::shared_ptr& reader) + : ReaderBase(), reader_(reader) { PADDLE_ENFORCE_NOT_NULL(reader_); } void ReInit() override { reader_->ReInit(); } protected: - ReaderBase* reader_; + std::shared_ptr reader_; }; class FileReader : public ReaderBase { @@ -64,7 +65,7 @@ class ReaderHolder { public: void Reset(ReaderBase* reader) { reader_.reset(reader); } - ReaderBase* Get() const { return reader_.get(); } + std::shared_ptr Get() const { return reader_; } void ReadNext(std::vector* out) { PADDLE_ENFORCE_NOT_NULL(reader_); @@ -76,7 +77,7 @@ class ReaderHolder { } private: - std::unique_ptr reader_; + std::shared_ptr reader_; }; } // namespace framework diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index 9091713158..bb2d866c82 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -34,13 +34,7 @@ DEFINE_bool( namespace paddle { namespace framework { -Scope::~Scope() { - DropKids(); - for (auto& kv : vars_) { - VLOG(3) << "Destroy variable " << kv.first; - delete kv.second; - } -} +Scope::~Scope() { DropKids(); } Scope& Scope::NewScope() const { std::unique_lock lock(mutex_); @@ -49,10 +43,13 @@ Scope& Scope::NewScope() const { } Variable* Scope::Var(const std::string& name) { + // acquire the lock when new var under this scope + std::unique_lock lock(mutex_); auto* v = FindVarLocally(name); if (v != nullptr) return v; + v = new Variable(); - vars_[name] = v; + vars_[name].reset(v); VLOG(3) << "Create variable " << name; v->name_ = &(vars_.find(name)->first); return v; @@ -67,22 +64,29 @@ Variable* Scope::Var(std::string* name) { } Variable* Scope::FindVar(const std::string& name) const { + // acquire the lock when find var + std::unique_lock lock(mutex_); + return FindVarInternal(name); +} + +Variable* Scope::FindVarInternal(const std::string& name) const { auto var = FindVarLocally(name); if (var != nullptr) { return var; } - return (parent_ == nullptr) ? nullptr : parent_->FindVar(name); + return (parent_ == nullptr) ? nullptr : parent_->FindVarInternal(name); } const Scope* Scope::FindScope(const Variable* var) const { for (auto& kv : vars_) { - if (kv.second == var) { + if (kv.second.get() == var) { return this; } } return (parent_ == nullptr) ? nullptr : parent_->FindScope(var); } void Scope::DropKids() { + std::unique_lock lock(mutex_); for (Scope* s : kids_) delete s; kids_.clear(); } @@ -110,10 +114,10 @@ void Scope::DeleteScope(Scope* scope) const { } void Scope::EraseVars(const std::vector& var_names) { + std::unique_lock lock(mutex_); std::set var_set(var_names.begin(), var_names.end()); for (auto it = vars_.begin(); it != vars_.end();) { if (var_set.find(it->first) != var_set.end()) { - delete it->second; it = vars_.erase(it); } else { ++it; @@ -129,7 +133,7 @@ void Scope::Rename(const std::string& origin_name, auto new_it = vars_.find(new_name); PADDLE_ENFORCE(new_it == vars_.end(), "The variable with name %s is already in the scope", new_name); - vars_[new_name] = origin_it->second; + vars_[new_name].reset(origin_it->second.release()); vars_.erase(origin_it); } @@ -141,7 +145,7 @@ std::string Scope::Rename(const std::string& origin_name) const { Variable* Scope::FindVarLocally(const std::string& name) const { auto it = vars_.find(name); - if (it != vars_.end()) return it->second; + if (it != vars_.end()) return it->second.get(); return nullptr; } diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h index abc82e452d..95b4f7c5f6 100644 --- a/paddle/fluid/framework/scope.h +++ b/paddle/fluid/framework/scope.h @@ -47,15 +47,18 @@ class Scope { Scope& NewScope() const; /// Create a variable with given name if it doesn't exist. + /// Caller doesn't own the returned Variable. Variable* Var(const std::string& name); /// Create a variable with a scope-unique name. + /// Caller doesn't own the returned Variable. Variable* Var(std::string* name = nullptr); void EraseVars(const std::vector& var_names); /// Find a variable in the scope or any of its ancestors. Returns /// nullptr if cannot find. + /// Caller doesn't own the returned Variable. Variable* FindVar(const std::string& name) const; const Scope* parent() const { return parent_; } @@ -78,13 +81,22 @@ class Scope { // Rename variable to a new name and return the new name std::string Rename(const std::string& origin_name) const; - Variable* FindVarLocally(const std::string& name) const; + protected: + mutable std::unordered_map> vars_; private: // Call Scope::NewScope for a sub-scope. explicit Scope(Scope const* parent) : parent_(parent) {} - mutable std::unordered_map vars_; + // Called by FindVar recursively. + // Caller doesn't own the returned Variable. + Variable* FindVarInternal(const std::string& name) const; + + // Called by FindVarInternal and Var. + // Caller doesn't own the returned Variable. + Variable* FindVarLocally(const std::string& name) const; + + // Scope in `kids_` are owned by this class. mutable std::list kids_; Scope const* parent_{nullptr}; diff --git a/paddle/fluid/framework/tensor.cc b/paddle/fluid/framework/tensor.cc index e97ada06f0..c7286dacf0 100644 --- a/paddle/fluid/framework/tensor.cc +++ b/paddle/fluid/framework/tensor.cc @@ -15,5 +15,102 @@ limitations under the License. */ #include "paddle/fluid/framework/tensor.h" namespace paddle { -namespace framework {} +namespace framework { +extern size_t SizeOfType(std::type_index type); +void Tensor::check_memory_size() const { + PADDLE_ENFORCE_NOT_NULL( + holder_, "Tensor holds no memory. Call Tensor::mutable_data first."); + PADDLE_ENFORCE_LE( + numel() * SizeOfType(type()), memory_size(), + "Tensor's dims_ is out of bound. Call Tensor::mutable_data " + "first to re-allocate memory.\n" + "or maybe the required data-type mismatches the data already stored."); +} + +size_t Tensor::memory_size() const { + return holder_ == nullptr ? 0UL : holder_->size() - offset_; +} + +void* Tensor::mutable_data(platform::Place place, std::type_index type) { + if (holder_ != nullptr) { + holder_->set_type(type); + } + PADDLE_ENFORCE_GE(numel(), 0, + "When calling this method, the Tensor's numel must be " + "equal or larger than zero. " + "Please check Tensor::Resize has been called first."); + int64_t size = numel() * SizeOfType(type); + /* some versions of boost::variant don't have operator!= */ + if (holder_ == nullptr || !(holder_->place() == place) || + holder_->size() < size + offset_) { + if (platform::is_cpu_place(place)) { + holder_.reset(new PlaceholderImpl( + boost::get(place), size, type)); + } else if (platform::is_gpu_place(place) || + platform::is_cuda_pinned_place(place)) { +#ifndef PADDLE_WITH_CUDA + PADDLE_THROW( + "CUDAPlace or CUDAPinnedPlace is not supported in CPU-only mode."); + } +#else + if (platform::is_gpu_place(place)) { + holder_.reset(new PlaceholderImpl( + boost::get(place), size, type)); + } else if (platform::is_cuda_pinned_place(place)) { + holder_.reset(new PlaceholderImpl( + boost::get(place), size, type)); + } + } +#endif + offset_ = 0; + } + return reinterpret_cast(reinterpret_cast(holder_->ptr()) + + offset_); +} + +void* Tensor::mutable_data(platform::Place place) { + PADDLE_ENFORCE(this->holder_ != nullptr, + "Cannot invoke mutable data if current hold nothing."); + return mutable_data(place, holder_->type()); +} + +Tensor& Tensor::ShareDataWith(const Tensor& src) { + src.check_memory_size(); + *this = src; + return *this; +} + +Tensor Tensor::Slice(int begin_idx, int end_idx) const { + check_memory_size(); + PADDLE_ENFORCE_GE(begin_idx, 0, + "The start row index must be greater than 0."); + PADDLE_ENFORCE_LE(end_idx, dims_[0], "The end row index is out of bound."); + PADDLE_ENFORCE_LT( + begin_idx, end_idx, + "The start row index must be lesser than the end row index."); + + if (dims_[0] == 1) { + return *this; + } else { + size_t base = numel() / dims_[0]; + Tensor dst; + dst.holder_ = holder_; + dst.set_layout(layout_); + DDim dst_dims = dims_; + dst_dims[0] = end_idx - begin_idx; + dst.Resize(dst_dims); + dst.offset_ = offset_ + begin_idx * base * SizeOfType(type()); + return dst; + } +} + +Tensor& Tensor::Resize(const DDim& dims) { + dims_ = dims; + return *this; +} + +const DDim& Tensor::dims() const { return dims_; } + +int64_t Tensor::numel() const { return product(dims_); } +} // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/tensor.h b/paddle/fluid/framework/tensor.h index 6f878541e6..ef224d68f1 100644 --- a/paddle/fluid/framework/tensor.h +++ b/paddle/fluid/framework/tensor.h @@ -34,6 +34,28 @@ namespace framework { class LoDTensor; class Tensor { +#ifdef PADDLE_WITH_MKLDNN + + public: + inline mkldnn::memory::format format() const { return format_; } + + inline void set_format(const mkldnn::memory::format format) { + format_ = format; + } + + protected: + /** + * @brief the detail format of memory block which have layout as kMKLDNN + * + * @note MKLDNN lib support various memory format like nchw, nhwc, nChw8C, + * nChw16c, etc. For a MKLDNN memory block, layout will be set as + * DataLayout::kMKLDNN meanwhile detail memory format will be kept in + * this field. + */ + + mkldnn::memory::format format_ = mkldnn::memory::format::format_undef; +#endif + public: template friend struct EigenTensor; @@ -54,26 +76,24 @@ class Tensor { /*! Return a pointer to mutable memory block. */ template - inline T* data(); + T* data(); /*! Return a pointer to constant memory block. */ template - inline const T* data() const; + const T* data() const; - inline bool IsInitialized() const; - - inline void switch_place(platform::Place new_place); + bool IsInitialized() const; /** * @brief Return a pointer to mutable memory block. * @note If not exist, then allocation. */ template - inline T* mutable_data(platform::Place place); + T* mutable_data(platform::Place place); - inline void* mutable_data(platform::Place place, std::type_index type); + void* mutable_data(platform::Place place, std::type_index type); - inline void* mutable_data(platform::Place place); + void* mutable_data(platform::Place place); /** * @brief Return a pointer to mutable memory block. @@ -84,19 +104,19 @@ class Tensor { * @note If not exist, then allocation. */ template - inline T* mutable_data(DDim dims, platform::Place place); + T* mutable_data(DDim dims, platform::Place place); /*! Return the dimensions of the memory block. */ - inline const DDim& dims() const; + const DDim& dims() const; /*! Return the numel of the memory block. */ - inline int64_t numel() const; + int64_t numel() const; /*! Resize the dimensions of the memory block. */ - inline Tensor& Resize(const DDim& dims); + Tensor& Resize(const DDim& dims); /*! The internal of two tensors share the same memory block. */ - inline Tensor& ShareDataWith(const Tensor& src); + Tensor& ShareDataWith(const Tensor& src); /** * @brief Return a sub-tensor of the given tensor. @@ -106,7 +126,7 @@ class Tensor { * @param[in] end_idx The index of the end row(exclusive) to slice. * The index number begins from 0. */ - inline Tensor Slice(int begin_idx, int end_idx) const; + Tensor Slice(int begin_idx, int end_idx) const; platform::Place place() const { PADDLE_ENFORCE_NOT_NULL( @@ -123,11 +143,11 @@ class Tensor { // memory size returns the holding memory size in byte. size_t memory_size() const; - inline void check_memory_size() const; + void check_memory_size() const; - inline DataLayout layout() const { return layout_; } + DataLayout layout() const { return layout_; } - inline void set_layout(const DataLayout layout) { layout_ = layout; } + void set_layout(const DataLayout layout) { layout_ = layout; } private: /** @@ -197,8 +217,10 @@ class Tensor { * N,C,H,W for respectively the batch size, the number of * feature maps, the height. */ - - DataLayout layout_ = DataLayout::kNHWC; + // Fix me: here just change the default layout to kNCHW + // it doesn't fix the real issue, i.e. feeder should set up tensor layout + // according to actual input data + DataLayout layout_ = DataLayout::kNCHW; /** * @brief A PlaceHolder may be shared by more than one tensor. @@ -210,15 +232,6 @@ class Tensor { size_t offset_; }; -inline void Tensor::switch_place(platform::Place new_place) { - if (holder_->place() == new_place) { - return; - } - - // TODO(tonyyang-svail): do memcpy here. - PADDLE_THROW("Not Implemented"); -} - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h index 0a1db7758b..96114678a9 100644 --- a/paddle/fluid/framework/tensor_impl.h +++ b/paddle/fluid/framework/tensor_impl.h @@ -20,26 +20,11 @@ limitations under the License. */ namespace paddle { namespace framework { -extern size_t SizeOfType(std::type_index type); -inline void Tensor::check_memory_size() const { - PADDLE_ENFORCE_NOT_NULL( - holder_, "Tensor holds no memory. Call Tensor::mutable_data first."); - PADDLE_ENFORCE_LE( - numel() * SizeOfType(type()), memory_size(), - "Tensor's dims_ is out of bound. Call Tensor::mutable_data " - "first to re-allocate memory.\n" - "or maybe the required data-type mismatches the data already stored."); -} - -inline size_t Tensor::memory_size() const { - return holder_ == nullptr ? 0UL : holder_->size() - offset_; -} - template inline const T* Tensor::data() const { check_memory_size(); PADDLE_ENFORCE(std::is_same::value || - holder_->type().hash_code() == typeid(T).hash_code(), + holder_->type() == std::type_index(typeid(T)), "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); @@ -53,7 +38,7 @@ template inline T* Tensor::data() { check_memory_size(); PADDLE_ENFORCE(std::is_same::value || - holder_->type().hash_code() == typeid(T).hash_code(), + holder_->type() == std::type_index(typeid(T)), "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); return reinterpret_cast(reinterpret_cast(holder_->ptr()) + @@ -73,88 +58,6 @@ inline T* Tensor::mutable_data(platform::Place place) { return reinterpret_cast(mutable_data(place, typeid(T))); } -inline void* Tensor::mutable_data(platform::Place place, std::type_index type) { - if (holder_ != nullptr) { - holder_->set_type(type); - } - PADDLE_ENFORCE_GE(numel(), 0, - "When calling this method, the Tensor's numel must be " - "equal or larger than zero. " - "Please check Tensor::Resize has been called first."); - int64_t size = numel() * SizeOfType(type); - /* some versions of boost::variant don't have operator!= */ - if (holder_ == nullptr || !(holder_->place() == place) || - holder_->size() < size + offset_) { - if (platform::is_cpu_place(place)) { - holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); - } else if (platform::is_gpu_place(place) || - platform::is_cuda_pinned_place(place)) { -#ifndef PADDLE_WITH_CUDA - PADDLE_THROW( - "CUDAPlace or CUDAPinnedPlace is not supported in CPU-only mode."); - } -#else - if (platform::is_gpu_place(place)) { - holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); - } else if (platform::is_cuda_pinned_place(place)) { - holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); - } - } -#endif - offset_ = 0; - } - return reinterpret_cast(reinterpret_cast(holder_->ptr()) + - offset_); -} - -inline void* Tensor::mutable_data(platform::Place place) { - PADDLE_ENFORCE(this->holder_ != nullptr, - "Cannot invoke mutable data if current hold nothing."); - return mutable_data(place, holder_->type()); -} - -inline Tensor& Tensor::ShareDataWith(const Tensor& src) { - src.check_memory_size(); - *this = src; - return *this; -} - -inline Tensor Tensor::Slice(int begin_idx, int end_idx) const { - check_memory_size(); - PADDLE_ENFORCE_GE(begin_idx, 0, - "The start row index must be greater than 0."); - PADDLE_ENFORCE_LE(end_idx, dims_[0], "The end row index is out of bound."); - PADDLE_ENFORCE_LT( - begin_idx, end_idx, - "The start row index must be lesser than the end row index."); - - if (dims_[0] == 1) { - return *this; - } else { - size_t base = numel() / dims_[0]; - Tensor dst; - dst.holder_ = holder_; - dst.set_layout(layout_); - DDim dst_dims = dims_; - dst_dims[0] = end_idx - begin_idx; - dst.Resize(dst_dims); - dst.offset_ = offset_ + begin_idx * base * SizeOfType(type()); - return dst; - } -} - -inline Tensor& Tensor::Resize(const DDim& dims) { - dims_ = dims; - return *this; -} - -inline const DDim& Tensor::dims() const { return dims_; } - -inline int64_t Tensor::numel() const { return product(dims_); } - inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) { Tensor res; res.ShareDataWith(src); diff --git a/paddle/fluid/framework/tensor_test.cc b/paddle/fluid/framework/tensor_test.cc index e1012de2ec..0a1cb6d570 100644 --- a/paddle/fluid/framework/tensor_test.cc +++ b/paddle/fluid/framework/tensor_test.cc @@ -209,7 +209,7 @@ TEST(Tensor, ReshapeToMatrix) { TEST(Tensor, Layout) { framework::Tensor src; - ASSERT_EQ(src.layout(), framework::DataLayout::kNHWC); + ASSERT_EQ(src.layout(), framework::DataLayout::kNCHW); src.set_layout(framework::DataLayout::kAnyLayout); ASSERT_EQ(src.layout(), framework::DataLayout::kAnyLayout); } diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index cc4a725dfb..ec16a1c600 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -5,14 +5,19 @@ cc_library(paddle_fluid_api SRCS io.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) -# Create static library get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) -cc_library(paddle_fluid DEPS ${fluid_modules}) +if(WITH_CONTRIB) + set(fluid_modules "${fluid_modules}" paddle_inference_api) +endif() + +# Create static library +cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api) # Create shared library cc_library(paddle_fluid_shared SHARED SRCS io.cc - DEPS ${fluid_modules}) + DEPS ${fluid_modules} paddle_fluid_api) + set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) if(NOT APPLE) # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac. diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt index 9faf5bb303..5083578444 100644 --- a/paddle/fluid/inference/analysis/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -15,3 +15,9 @@ cc_test(test_subgraph_splitter DEPS analysis paddle_fluid tensor ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model) set_tests_properties(test_subgraph_splitter PROPERTIES DEPENDS test_word2vec) + +cc_test(test_dfg_graphviz_draw_pass + SRCS dfg_graphviz_draw_pass_tester.cc + DEPS analysis + ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model) +set_tests_properties(test_dfg_graphviz_draw_pass PROPERTIES DEPENDS test_word2vec) diff --git a/paddle/fluid/inference/analysis/data_flow_graph.h b/paddle/fluid/inference/analysis/data_flow_graph.h index 9f6ce40ede..913e344d37 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph.h +++ b/paddle/fluid/inference/analysis/data_flow_graph.h @@ -21,7 +21,10 @@ limitations under the License. */ #include #include +#include #include +#include +#include #include "paddle/fluid/inference/analysis/graph_traits.h" #include "paddle/fluid/inference/analysis/node.h" diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc index 60f159da91..dcee75cee5 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc @@ -44,6 +44,6 @@ TEST_F(DFG_Tester, Test) { LOG(INFO) << graph.nodes.size(); } -} // analysis -} // inference -} // paddle +}; // namespace analysis +}; // namespace inference +}; // namespace paddle diff --git a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h new file mode 100644 index 0000000000..41d4475382 --- /dev/null +++ b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h @@ -0,0 +1,68 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +/* + * This file create an DFG_GraphvizDrawPass which helps to draw a data flow + * graph's structure using graphviz. + */ + +#pragma once + +#include +#include +#include "paddle/fluid/inference/analysis/pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +/* + * Output a dot file and write to some place. + */ +class DFG_GraphvizDrawPass : public DataFlowGraphPass { + public: + DFG_GraphvizDrawPass(const std::string& dir, const std::string& id) + : dir_(dir), id_(id) {} + + bool Initialize() override { return Pass::Initialize(); } + void Run(DataFlowGraph* graph) override { + auto content = Draw(graph); + std::ofstream file(GenDotPath()); + file.write(content.c_str(), content.size()); + file.close(); + LOG(INFO) << "draw dot to " << GenDotPath(); + } + + bool Finalize() override { return Pass::Finalize(); } + + Pass* CreatePrinterPass(std::ostream& os, + const std::string& banner) const override { + return nullptr; + } + + private: + // Path of the dot file to output. + std::string GenDotPath() const { + return dir_ + "/" + "graph_" + id_ + ".dot"; + } + + std::string Draw(DataFlowGraph* graph) { return graph->DotString(); } + + std::string dir_; + std::string id_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc new file mode 100644 index 0000000000..3fc1cc18b8 --- /dev/null +++ b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc @@ -0,0 +1,46 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" + +#include +#include +#include +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST_F(DFG_Tester, dfg_graphviz_draw_pass_tester) { + auto dfg = ProgramDescToDFG(desc); + DFG_GraphvizDrawPass pass("./", "test"); + pass.Initialize(); + pass.Run(&dfg); + + // test content + std::ifstream file("./graph_test.dot"); + ASSERT_TRUE(file.is_open()); + + std::string line; + int no{0}; + while (std::getline(file, line)) { + no++; + } + ASSERT_EQ(no, 82); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc index f848a7d1ad..9f67c989cc 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc @@ -12,9 +12,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" +#include #include +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" + namespace paddle { namespace inference { namespace analysis { diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h index cd0d4fabaa..33517e57be 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h @@ -19,6 +19,8 @@ #pragma once +#include + #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/pass.h" diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc index 851c98bef3..817d32c92c 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc @@ -32,6 +32,6 @@ TEST_F(DFG_Tester, Init) { LOG(INFO) << '\n' << graph.DotString(); } -} // analysis -} // inference -} // paddle +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/helper.h b/paddle/fluid/inference/analysis/helper.h index 24ea9a4bae..58eb0e715c 100644 --- a/paddle/fluid/inference/analysis/helper.h +++ b/paddle/fluid/inference/analysis/helper.h @@ -18,6 +18,8 @@ limitations under the License. */ #include #include +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/variable.h" #include "paddle/fluid/platform/enforce.h" namespace paddle { @@ -50,7 +52,7 @@ struct DataTypeNamer { return dic_.at(x); } - const std::string &repr(size_t &hash) const { + const std::string &repr(size_t &hash) const { // NOLINT PADDLE_ENFORCE(dic_.count(hash), "unknown type for representation"); return dic_.at(hash); } @@ -62,7 +64,9 @@ struct DataTypeNamer { SET_TYPE(float); } - std::unordered_map dic_; + std::unordered_map + dic_; }; #undef SET_TYPE @@ -105,6 +109,13 @@ class OrderedRegistry { std::vector> data_; }; +template +T &GetFromScope(const framework::Scope &scope, const std::string &name) { + framework::Variable *var = scope.FindVar(name); + PADDLE_ENFORCE(var != nullptr); + return *var->GetMutable(); +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/pass.h b/paddle/fluid/inference/analysis/pass.h index 5c89b1304d..aa0e8667b5 100644 --- a/paddle/fluid/inference/analysis/pass.h +++ b/paddle/fluid/inference/analysis/pass.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include +#include #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.h b/paddle/fluid/inference/analysis/subgraph_splitter.h index ed90a0dcf3..a31afbe693 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter.h +++ b/paddle/fluid/inference/analysis/subgraph_splitter.h @@ -18,6 +18,8 @@ limitations under the License. */ #pragma once +#include + #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/node.h" diff --git a/paddle/fluid/inference/analysis/ut_helper.h b/paddle/fluid/inference/analysis/ut_helper.h index c86083d121..722fa99a48 100644 --- a/paddle/fluid/inference/analysis/ut_helper.h +++ b/paddle/fluid/inference/analysis/ut_helper.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include #include +#include #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" diff --git a/paddle/fluid/inference/io.cc b/paddle/fluid/inference/io.cc index 65db7c7b50..6b03ac7119 100644 --- a/paddle/fluid/inference/io.cc +++ b/paddle/fluid/inference/io.cc @@ -20,16 +20,20 @@ limitations under the License. */ #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/feed_fetch_type.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/pybind/pybind.h" DEFINE_string(devices, "", "The devices to be used which is joined by comma."); DEFINE_bool(init_p2p, false, "Whether to init p2p."); +DEFINE_int32(math_num_threads, 1, + "Number of threads used to run math functions."); namespace paddle { namespace inference { void Init(const std::vector argv) { framework::InitGflags(argv); + operators::math::SetNumThreads(FLAGS_math_num_threads); // init devices std::vector devices; std::string token; diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 5ada1d6312..748f5a084e 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,10 +1,15 @@ # Add TRT tests -nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine) -# This test is not stable -# See https://paddleci.ngrok.io/viewLog.html?tab=buildLog&buildTypeId=Paddle_PrCi2&buildId=36834&_focus=8828 -#nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc -# DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine -# SERIAL) +nv_library(tensorrt_converter + SRCS mul_op.cc conv2d_op.cc fc_op.cc + DEPS tensorrt_engine mul_op) + +nv_test(test_op_converter SRCS test_op_converter.cc DEPS + ${FLUID_CORE_MODULES} tensorrt_engine tensorrt_converter) + nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL) +nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL) +nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine activation_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc index 6297051e5a..e1cace9cc1 100644 --- a/paddle/fluid/inference/tensorrt/convert/activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" namespace paddle { @@ -21,10 +22,11 @@ namespace tensorrt { class ReluOpConverter : public OpConverter { public: ReluOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { // Here the two nullptr looks strange, that's because the // framework::OpDesc's constructor is strange. - framework::OpDesc op_desc(op, nullptr, nullptr); + framework::OpDesc op_desc(op, nullptr); LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose " "type is Relu"; const nvinfer1::ITensor* input_tensor = @@ -32,12 +34,17 @@ class ReluOpConverter : public OpConverter { nvinfer1::IActivationLayer* layer = TRT_ENGINE_ADD_LAYER( engine_, Activation, *const_cast(input_tensor), nvinfer1::ActivationType::kRELU); - engine_->SetITensor(op_desc.Output("Out")[0], layer->getOutput(0)); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } } }; -REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter); - } // namespace tensorrt } // namespace inference } // namespace paddle + +REGISTER_TRT_OP_CONVERTER(relu, ReluOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index 209936c3ba..8e7e23377d 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -21,14 +21,15 @@ namespace tensorrt { class Conv2dOpConverter : public OpConverter { public: Conv2dOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { LOG(INFO) << "convert a fluid conv2d op to tensorrt conv layer without bias"; } }; -REGISTER_TRT_OP_CONVERTER(conv2d, Conv2dOpConverter); - } // namespace tensorrt } // namespace inference } // namespace paddle + +REGISTER_TRT_OP_CONVERTER(conv2d, Conv2dOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc new file mode 100644 index 0000000000..bb603efaf3 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/fc_op.cc @@ -0,0 +1,121 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +// Reorder the elements from istrides to ostrides, borrowed from TRT convert in +// tensorflow. +// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/tensorrt/convert/convert_nodes.cc#L318 +template +void Reorder2(nvinfer1::DimsHW shape, const T* idata, nvinfer1::DimsHW istrides, + T* odata, nvinfer1::DimsHW ostrides) { + for (int h = 0; h < shape.h(); ++h) { + for (int w = 0; w < shape.w(); ++w) { + odata[h * ostrides.h() + w * ostrides.w()] = + idata[h * ostrides.h() + w * ostrides.w()]; + } + } +} + +// Reorder the data layout from CK to KC. +void ReorderCKtoKC(TensorRTEngine::Weight& iweights, + TensorRTEngine::Weight* oweights) { + int c = iweights.dims[0]; + int k = iweights.dims[1]; + oweights->dims.assign({k, c}); + nvinfer1::DimsHW istrides = {1, k}; + nvinfer1::DimsHW ostrides = {c, 1}; + Reorder2({k, c}, static_cast(iweights.get().values), istrides, + static_cast(const_cast(oweights->get().values)), + ostrides); +} + +/* + * FC converter convert a MUL op in Fluid to a FC layer in TRT. + */ +class FcOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) << "convert a fluid fc op to tensorrt fc layer without bias"; + + framework::OpDesc op_desc(op, nullptr); + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); + + // Declare inputs + auto* X = engine_->GetITensor(op_desc.Input("X").front()); + + // Declare weights + auto* Y_v = scope.FindVar(op_desc.Input("Y").front()); + PADDLE_ENFORCE_NOT_NULL(Y_v); + auto* Y_t = Y_v->GetMutable(); + // This may trigger a GPU->CPU copy, because TRT's weight can only be + // assigned from CPU memory, that can't be avoided. + auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); + PADDLE_ENFORCE_EQ(Y_t->dims().size(), 2UL); // a matrix + size_t n_output = Y_t->dims()[1]; + + framework::LoDTensor tmp; + tmp.Resize(Y_t->dims()); + memcpy(tmp.mutable_data(platform::CPUPlace()), Y_t->data(), + Y_t->dims()[0] * Y_t->dims()[1]); + + TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, + static_cast(weight_data), + Y_t->memory_size() / sizeof(float)}; + TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT, + static_cast(tmp.data()), + Y_t->memory_size() / sizeof(float)); + weight.dims.assign({Y_t->dims()[0], Y_t->dims()[1]}); + tmp_weight.dims = weight.dims; + + // The data layout of TRT FC layer's weight is different from fluid's FC, + // need to reorder the elements. + ReorderCKtoKC(tmp_weight, &weight); + + // Currently, the framework can only handle one fluid op -> one TRT layer, + // but fc fuses `mul` and `bias` (2 fluid ops), so here is a trick, just + // handle `mul`, leave `add` as another layer. + // DEBUG + TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT, nullptr, 0}; + + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, + *const_cast(X), + n_output, weight.get(), bias.get()); + + auto output_name = op_desc.Output("Out").front(); + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(fc, FcOpConverter); +USE_OP(mul); diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc index ed09f54bde..3c34295736 100644 --- a/paddle/fluid/inference/tensorrt/convert/mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/mul_op.cc @@ -23,11 +23,11 @@ namespace tensorrt { */ class MulOpConverter : public OpConverter { public: - MulOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { - VLOG(4) << "convert a fluid mul op to tensorrt fc layer without bias"; + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias"; - framework::OpDesc op_desc(op, nullptr, nullptr); + framework::OpDesc op_desc(op, nullptr); // Declare inputs auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); auto* input2 = engine_->GetITensor(op_desc.Input("Y")[0]); @@ -36,12 +36,18 @@ class MulOpConverter : public OpConverter { engine_, MatrixMultiply, *const_cast(input1), false, *const_cast(input2), false); - engine_->DeclareOutput(layer, 0, op_desc.Output("Out")[0]); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } } }; -REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter); - } // namespace tensorrt } // namespace inference } // namespace paddle + +USE_OP(mul); +REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index 1cd3ed9a00..c7a5a49dd0 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -17,6 +17,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/inference/utils/singleton.h" @@ -31,27 +32,45 @@ namespace tensorrt { class OpConverter { public: OpConverter() {} - virtual void operator()(const framework::proto::OpDesc& op) {} - void Run(const framework::proto::OpDesc& op, TensorRTEngine* engine) { - std::string type = op.type(); - auto* it = Registry::Lookup(type); - PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type); + // Converter logic for an op. + virtual void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, + bool test_mode = false) {} + + // Convert a single fluid operator and add the corresponding layer to TRT. + // test_mode: whether the instance executes in an unit test. + void ConvertOp(const framework::proto::OpDesc& op, + const std::unordered_set& parameters, + const framework::Scope& scope, TensorRTEngine* engine, + bool test_mode = false) { + framework::OpDesc op_desc(op, nullptr); + + OpConverter* it{nullptr}; + + if (op_desc.Type() == "mul") { + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1UL); + std::string Y = op_desc.Input("Y")[0]; + if (parameters.count(Y)) { + it = Registry::Lookup("fc"); + } + } + if (!it) { + it = Registry::Lookup(op_desc.Type()); + } + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", + op_desc.Type()); it->SetEngine(engine); - (*it)(op); - } - - // convert fluid op to tensorrt layer - void ConvertOp(const framework::proto::OpDesc& op, TensorRTEngine* engine) { - OpConverter::Run(op, engine); + (*it)(op, scope, test_mode); } // convert fluid block to tensorrt network void ConvertBlock(const framework::proto::BlockDesc& block, - TensorRTEngine* engine) { + const std::unordered_set& parameters, + const framework::Scope& scope, TensorRTEngine* engine) { for (int i = 0; i < block.ops_size(); i++) { const auto& op = block.ops(i); - OpConverter::Run(op, engine); + ConvertOp(op, parameters, scope, engine); } } @@ -62,6 +81,9 @@ class OpConverter { // TensorRT engine TensorRTEngine* engine_{nullptr}; + protected: + bool test_mode_; + private: // registered op converter map, whose key is the fluid op type, and value is // the pointer position of corresponding OpConverter class. @@ -70,13 +92,24 @@ class OpConverter { framework::Scope* scope_{nullptr}; }; -#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ - struct trt_##op_type__##_converter { \ - trt_##op_type__##_converter() { \ - Registry::Register(#op_type__); \ - } \ - }; \ - trt_##op_type__##_converter trt_##op_type__##_converter__; +#define REGISTER_TRT_OP_CONVERTER(op_type__, Converter__) \ + struct trt_##op_type__##_converter : public ::paddle::framework::Registrar { \ + trt_##op_type__##_converter() { \ + ::paddle::inference:: \ + Registry::Register< \ + ::paddle::inference::tensorrt::Converter__>(#op_type__); \ + } \ + }; \ + trt_##op_type__##_converter trt_##op_type__##_converter__; \ + int TouchConverterRegister_##op_type__() { \ + trt_##op_type__##_converter__.Touch(); \ + return 0; \ + } + +#define USE_TRT_CONVERTER(op_type__) \ + extern int TouchConverterRegister_##op_type__(); \ + static int use_op_converter_trt_##op_type__ __attribute__((unused)) = \ + TouchConverterRegister_##op_type__(); } // namespace tensorrt } // namespace inference diff --git a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc index 86ca2ca08e..0a02a7bebf 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_activation_op.cc @@ -1,106 +1,47 @@ /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 + http://www.apache.org/licenses/LICENSE-2.0 -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ #include -#include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/inference/tensorrt/convert/io_converter.h" -#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" -#include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/place.h" - -USE_OP(relu); +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" namespace paddle { namespace inference { namespace tensorrt { -void Compare(const std::string op_type, float input, float expect) { +TEST(ReluOpConverter, main) { framework::Scope scope; - platform::CUDAPlace place; - platform::CUDADeviceContext ctx(place); - - // init fluid op and variable - auto x_var = scope.Var("X"); - auto x_tensor = x_var->GetMutable(); - x_tensor->Resize({1, 1}); - x_tensor->mutable_data(place); - std::vector init; - init.push_back(input); - framework::TensorFromVector(init, ctx, x_tensor); - - auto out_var = scope.Var("Out"); - auto out_tensor = out_var->GetMutable(); - out_tensor->Resize({1, 1}); - out_tensor->mutable_data(place); - - framework::OpDesc op_desc; - op_desc.SetType(op_type); - op_desc.SetInput("X", {"X"}); - op_desc.SetOutput("Out", {"Out"}); - - auto op = framework::OpRegistry::CreateOp(*op_desc.Proto()); - - // run fluid op - op->Run(scope, place); - // get fluid output - std::vector out1; - framework::TensorToVector(*out_tensor, ctx, &out1); - - // init tensorrt op - cudaStream_t stream; - ASSERT_EQ(0, cudaStreamCreate(&stream)); - TensorRTEngine* engine = new TensorRTEngine(1, 1 << 10, &stream); - engine->InitNetwork(); - engine->DeclareInput("X", nvinfer1::DataType::kFLOAT, - nvinfer1::DimsCHW{1, 1, 1}); - // convert op - OpConverter op_converter; - op_converter.ConvertOp(*op_desc.Proto(), engine); - - engine->DeclareOutput("Out"); - engine->FreezeNetwork(); - - // convert LoDTensor to ITensor - size_t size = x_tensor->memory_size(); - EngineIOConverter::ConvertInput(op_type, *x_tensor, - engine->buffer("X").buffer, size, &stream); - // run tensorrt Outp - engine->Execute(1); - // convert ITensor to LoDTensor - EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer, - out_tensor, size, &stream); - // get tensorrt output - std::vector out2; - framework::TensorToVector(*out_tensor, ctx, &out2); - - // compare - ASSERT_EQ(out1[0], out2[0]); - ASSERT_EQ(out1[0], expect); - - delete engine; - cudaStreamDestroy(stream); -} - -TEST(OpConverter, ConvertRelu) { - Compare("relu", 1, 1); // relu(1) = 1 - Compare("relu", -5, 0); // relu(-5) = 0 + std::unordered_set parameters; + TRTConvertValidation validator(10, parameters, scope, 1000); + validator.DeclInputVar("relu-X", nvinfer1::Dims2(10, 6)); + validator.DeclOutputVar("relu-Out", nvinfer1::Dims2(10, 6)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("relu"); + desc.SetInput("X", {"relu-X"}); + desc.SetOutput("Out", {"relu-Out"}); + + LOG(INFO) << "set OP"; + validator.SetOp(*desc.Proto()); + LOG(INFO) << "execute"; + + validator.Execute(10); } } // namespace tensorrt } // namespace inference } // namespace paddle -USE_OP(activation); +USE_OP(relu); diff --git a/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc new file mode 100644 index 0000000000..a30253072a --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc @@ -0,0 +1,46 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(fc_op, test) { + std::unordered_set parameters({"mul-Y"}); + framework::Scope scope; + TRTConvertValidation validator(20, parameters, scope, 1000); + + validator.DeclInputVar("mul-X", nvinfer1::Dims4(8, 3, 1, 1)); + validator.DeclParamVar("mul-Y", nvinfer1::Dims2(3, 2)); + validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(8, 2)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("mul"); + desc.SetInput("X", {"mul-X"}); + desc.SetInput("Y", {"mul-Y"}); + desc.SetOutput("Out", {"mul-Out"}); + + validator.SetOp(*desc.Proto()); + + validator.Execute(10); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc index d8b61d5f08..1ce1130e5d 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc @@ -21,7 +21,9 @@ namespace inference { namespace tensorrt { TEST(MulOpConverter, main) { - TRTConvertValidation validator(10, 1000); + framework::Scope scope; + std::unordered_set parameters; + TRTConvertValidation validator(10, parameters, scope, 1000); validator.DeclInputVar("mul-X", nvinfer1::Dims2(10, 6)); validator.DeclInputVar("mul-Y", nvinfer1::Dims2(6, 10)); validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(10, 10)); diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index 9ae7de9cbf..9b79f86b0e 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -12,9 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + #include #include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" namespace paddle { namespace inference { @@ -27,9 +28,13 @@ TEST(OpConverter, ConvertBlock) { conv2d_op->SetType("conv2d"); OpConverter converter; - converter.ConvertBlock(*block->Proto(), nullptr /*TensorRTEngine*/); + framework::Scope scope; + converter.ConvertBlock(*block->Proto(), {}, scope, + nullptr /*TensorRTEngine*/); } } // namespace tensorrt } // namespace inference } // namespace paddle + +USE_TRT_CONVERTER(conv2d) diff --git a/paddle/fluid/inference/tensorrt/convert/ut_helper.h b/paddle/fluid/inference/tensorrt/convert/ut_helper.h index 37fcb5c503..3b1f531adc 100644 --- a/paddle/fluid/inference/tensorrt/convert/ut_helper.h +++ b/paddle/fluid/inference/tensorrt/convert/ut_helper.h @@ -19,11 +19,15 @@ limitations under the License. */ #pragma once +#include +#include + #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/inference/utils/singleton.h" namespace paddle { namespace inference { @@ -58,7 +62,11 @@ class TRTConvertValidation { public: TRTConvertValidation() = delete; - TRTConvertValidation(int batch_size, int workspace_size = 1 << 10) { + TRTConvertValidation(int batch_size, + const std::unordered_set& parameters, + framework::Scope& scope, // NOLINT + int workspace_size = 1 << 10) + : parameters_(parameters), scope_(scope) { // create engine. engine_.reset(new TensorRTEngine(10, 1 << 10, &stream_)); engine_->InitNetwork(); @@ -73,19 +81,22 @@ class TRTConvertValidation { engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, dims); } + // Declare a parameter varaible in the scope. + void DeclParamVar(const std::string& name, const nvinfer1::Dims& dims) { + DeclVar(name, dims); + } + void DeclOutputVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims); } + // Declare a variable in a fluid Scope. void DeclVar(const std::string& name, const nvinfer1::Dims& dims) { platform::CPUPlace place; platform::CPUDeviceContext ctx(place); // Init Fluid tensor. - std::vector dim_vec(dims.nbDims); - for (int i = 0; i < dims.nbDims; i++) { - dim_vec[i] = dims.d[i]; - } + std::vector dim_vec(dims.d, dims.d + dims.nbDims); auto* x = scope_.Var(name); auto* x_tensor = x->GetMutable(); x_tensor->Resize(framework::make_ddim(dim_vec)); @@ -95,21 +106,23 @@ class TRTConvertValidation { void SetOp(const framework::proto::OpDesc& desc) { op_ = framework::OpRegistry::CreateOp(desc); - OpConverter op_converter; - op_converter.ConvertOp(desc, engine_.get()); + Singleton::Global().ConvertOp( + desc, parameters_, scope_, engine_.get(), true /*test_mode*/); engine_->FreezeNetwork(); // Declare outputs. - op_desc_.reset(new framework::OpDesc(desc, nullptr, nullptr)); + op_desc_.reset(new framework::OpDesc(desc, nullptr)); // Set Inputs. for (const auto& input : op_desc_->InputArgumentNames()) { + if (parameters_.count(input)) continue; auto* var = scope_.FindVar(input); PADDLE_ENFORCE(var); auto tensor = var->GetMutable(); + engine_->SetInputFromCPU( - input, static_cast(tensor->data()), + input, static_cast(tensor->data()), sizeof(float) * analysis::AccuDims(tensor->dims(), tensor->dims().size())); } @@ -117,18 +130,21 @@ class TRTConvertValidation { void Execute(int batch_size) { // Execute Fluid Op - // Execute TRT platform::CPUPlace place; platform::CPUDeviceContext ctx(place); - engine_->Execute(batch_size); - op_->Run(scope_, place); + // Execute TRT. + engine_->Execute(batch_size); + cudaStreamSynchronize(*engine_->stream()); ASSERT_FALSE(op_desc_->OutputArgumentNames().empty()); + const size_t output_space_size = 200; for (const auto& output : op_desc_->OutputArgumentNames()) { std::vector fluid_out; - std::vector trt_out(200); - engine_->GetOutputInCPU(output, &trt_out[0], 200 * sizeof(float)); + std::vector trt_out(output_space_size); + engine_->GetOutputInCPU(output, &trt_out[0], + output_space_size * sizeof(float)); + cudaStreamSynchronize(*engine_->stream()); auto* var = scope_.FindVar(output); auto tensor = var->GetMutable(); @@ -136,7 +152,8 @@ class TRTConvertValidation { // Compare two output ASSERT_FALSE(fluid_out.empty()); for (size_t i = 0; i < fluid_out.size(); i++) { - EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 0.001); + // Loose the threshold for CI in different machine model. + EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 2e-5); } } } @@ -146,9 +163,10 @@ class TRTConvertValidation { private: std::unique_ptr engine_; cudaStream_t stream_; - framework::Scope scope_; std::unique_ptr op_; std::unique_ptr op_desc_; + const std::unordered_set& parameters_; + framework::Scope& scope_; }; } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index a88236ae98..596e0fe9da 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -43,9 +43,10 @@ void TensorRTEngine::Execute(int batch_size) { } TensorRTEngine::~TensorRTEngine() { + cudaStreamSynchronize(*stream_); // clean buffer for (auto& buf : buffers_) { - if (buf.buffer != nullptr) { + if (buf.device == DeviceType::GPU && buf.buffer != nullptr) { PADDLE_ENFORCE_EQ(0, cudaFree(buf.buffer)); buf.buffer = nullptr; buf.max_size = 0; @@ -80,6 +81,8 @@ void TensorRTEngine::FreezeNetwork() { auto& buf = buffer(item.first); CHECK(buf.buffer == nullptr); // buffer should be allocated only once. PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second)); + VLOG(4) << "buffer malloc " << item.first << " " << item.second << " " + << buf.buffer; buf.size = buf.max_size = item.second; buf.device = DeviceType::GPU; } @@ -96,6 +99,7 @@ nvinfer1::ITensor* TensorRTEngine::DeclareInput(const std::string& name, PADDLE_ENFORCE(input, "infer network add input %s failed", name); buffer_sizes_[name] = kDataTypeSize[static_cast(dtype)] * analysis::AccuDims(dims.d, dims.nbDims); + PADDLE_ENFORCE(input->isNetworkInput()); TensorRTEngine::SetITensor(name, input); return input; } @@ -106,9 +110,12 @@ void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer* layer, int offset, name); auto* output = layer->getOutput(offset); + SetITensor(name, output); PADDLE_ENFORCE(output != nullptr); output->setName(name.c_str()); + PADDLE_ENFORCE(!output->isNetworkInput()); infer_network_->markOutput(*output); + PADDLE_ENFORCE(output->isNetworkOutput()); // output buffers' size can only be decided latter, set zero here to mark this // and will reset latter. buffer_sizes_[name] = 0; @@ -121,6 +128,7 @@ void TensorRTEngine::DeclareOutput(const std::string& name) { auto* output = TensorRTEngine::GetITensor(name); PADDLE_ENFORCE(output != nullptr); output->setName(name.c_str()); + PADDLE_ENFORCE(!output->isNetworkInput()); infer_network_->markOutput(*output); // output buffers' size can only be decided latter, set zero here to mark this // and will reset latter. diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index d9d3163b66..b60f00de9f 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -21,6 +21,7 @@ limitations under the License. */ #include #include "paddle/fluid/inference/engine.h" #include "paddle/fluid/inference/tensorrt/helper.h" +#include "paddle/fluid/inference/utils/singleton.h" namespace paddle { namespace inference { @@ -37,13 +38,15 @@ class TensorRTEngine : public EngineBase { // Weight is model parameter. class Weight { public: - Weight(nvinfer1::DataType dtype, void* value, int num_elem) { + Weight(nvinfer1::DataType dtype, void* value, size_t num_elem) { w_.type = dtype; w_.values = value; w_.count = num_elem; } const nvinfer1::Weights& get() { return w_; } + std::vector dims; + private: nvinfer1::Weights w_; }; @@ -129,7 +132,11 @@ class TensorRTEngine : public EngineBase { // TensorRT related internal members template struct Destroyer { - void operator()(T* x) { x->destroy(); } + void operator()(T* x) { + if (x) { + x->destroy(); + } + } }; template using infer_ptr = std::unique_ptr>; @@ -153,6 +160,27 @@ class TensorRTEngine : public EngineBase { #define TRT_ENGINE_ADD_LAYER(engine__, layer__, ARGS...) \ engine__->network()->add##layer__(ARGS); +/* + * Helper to control the TensorRT engine's creation and deletion. + */ +class TRT_EngineManager { + public: + TensorRTEngine* Create(int max_batch, int max_workspace, + cudaStream_t* stream) { + engines_.emplace_back(new TensorRTEngine(max_batch, max_workspace, stream)); + return engines_.back().get(); + } + + void DeleteALl() { + for (auto& ptr : engines_) { + ptr.reset(nullptr); + } + } + + private: + std::vector> engines_; +}; + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/book/CMakeLists.txt b/paddle/fluid/inference/tests/book/CMakeLists.txt index dbb81462b8..2fa5a9540b 100644 --- a/paddle/fluid/inference/tests/book/CMakeLists.txt +++ b/paddle/fluid/inference/tests/book/CMakeLists.txt @@ -38,3 +38,11 @@ inference_test(recommender_system) #inference_test(rnn_encoder_decoder) #inference_test(understand_sentiment ARGS conv) inference_test(word2vec) + +# This is an unly work around to make this test run +# TODO(TJ): clean me up +cc_test(test_inference_nlp + SRCS test_inference_nlp.cc + DEPS paddle_fluid + ARGS + --model_path=${PADDLE_BINARY_DIR}/python/paddle/fluid/tests/book/recognize_digits_mlp.inference.model) diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc index 987da18116..60c761c528 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -21,7 +21,6 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model."); DEFINE_int32(batch_size, 1, "Batch size of input data"); DEFINE_int32(repeat, 1, "Running the inference program repeat times"); DEFINE_bool(skip_cpu, false, "Skip the cpu test"); -DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference"); TEST(inference, image_classification) { if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { @@ -59,10 +58,8 @@ TEST(inference, image_classification) { // Run inference on CPU LOG(INFO) << "--- CPU Runs: ---"; LOG(INFO) << "Batch size is " << FLAGS_batch_size; - LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn; TestInference( - dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined, - FLAGS_use_mkldnn); + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined); LOG(INFO) << output1.dims(); } diff --git a/paddle/fluid/inference/tests/book/test_inference_nlp.cc b/paddle/fluid/inference/tests/book/test_inference_nlp.cc new file mode 100644 index 0000000000..9dcd79c3bb --- /dev/null +++ b/paddle/fluid/inference/tests/book/test_inference_nlp.cc @@ -0,0 +1,227 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include // NOLINT +#include "gflags/gflags.h" +#include "gtest/gtest.h" +#include "paddle/fluid/inference/tests/test_helper.h" +#ifdef PADDLE_WITH_MKLML +#include +#include +#endif + +DEFINE_string(model_path, "", "Directory of the inference model."); +DEFINE_string(data_file, "", "File of input index data."); +DEFINE_int32(repeat, 100, "Running the inference program repeat times"); +DEFINE_bool(prepare_vars, true, "Prepare variables before executor"); +DEFINE_int32(num_threads, 1, "Number of threads should be used"); + +inline double GetCurrentMs() { + struct timeval time; + gettimeofday(&time, NULL); + return 1e+3 * time.tv_sec + 1e-3 * time.tv_usec; +} + +// This function just give dummy data for recognize_digits model. +size_t DummyData(std::vector* out) { + paddle::framework::LoDTensor input; + SetupTensor(&input, {1, 1, 28, 28}, -1.f, 1.f); + out->emplace_back(input); + return 1; +} + +// Load the input word index data from file and save into LodTensor. +// Return the size of words. +size_t LoadData(std::vector* out, + const std::string& filename) { + if (filename.empty()) { + return DummyData(out); + } + + size_t sz = 0; + std::fstream fin(filename); + std::string line; + out->clear(); + while (getline(fin, line)) { + std::istringstream iss(line); + std::vector ids; + std::string field; + while (getline(iss, field, ' ')) { + ids.push_back(stoi(field)); + } + if (ids.size() >= 1024) { + // Synced with NLP guys, they will ignore input larger then 1024 + continue; + } + + paddle::framework::LoDTensor words; + paddle::framework::LoD lod{{0, ids.size()}}; + words.set_lod(lod); + int64_t* pdata = words.mutable_data( + {static_cast(ids.size()), 1}, paddle::platform::CPUPlace()); + memcpy(pdata, ids.data(), words.numel() * sizeof(int64_t)); + out->emplace_back(words); + sz += ids.size(); + } + return sz; +} + +// Split input data samples into small pieces jobs as balanced as possible, +// according to the number of threads. +void SplitData( + const std::vector& datasets, + std::vector>* jobs, + const int num_threads) { + size_t s = 0; + jobs->resize(num_threads); + while (s < datasets.size()) { + for (auto it = jobs->begin(); it != jobs->end(); it++) { + it->emplace_back(&datasets[s]); + s++; + if (s >= datasets.size()) { + break; + } + } + } +} + +void ThreadRunInfer( + const int tid, paddle::framework::Scope* scope, + const std::vector>& jobs) { + // maybe framework:ProgramDesc is not thread-safe + auto& sub_scope = scope->NewScope(); + auto place = paddle::platform::CPUPlace(); + auto executor = paddle::framework::Executor(place); + auto inference_program = + paddle::inference::Load(&executor, scope, FLAGS_model_path); + + auto ctx = executor.Prepare(*inference_program, /*block_id*/ 0); + executor.CreateVariables(*inference_program, &sub_scope, /*block_id*/ 0); + + const std::vector& feed_target_names = + inference_program->GetFeedTargetNames(); + const std::vector& fetch_target_names = + inference_program->GetFetchTargetNames(); + + PADDLE_ENFORCE_EQ(fetch_target_names.size(), 1UL); + std::map fetch_targets; + paddle::framework::LoDTensor outtensor; + fetch_targets[fetch_target_names[0]] = &outtensor; + + std::map feed_targets; + PADDLE_ENFORCE_EQ(feed_target_names.size(), 1UL); + + auto& inputs = jobs[tid]; + auto start_ms = GetCurrentMs(); + for (size_t i = 0; i < inputs.size(); ++i) { + feed_targets[feed_target_names[0]] = inputs[i]; + executor.RunPreparedContext(ctx.get(), &sub_scope, &feed_targets, + &fetch_targets, false /*create_local_scope*/); + } + auto stop_ms = GetCurrentMs(); + scope->DeleteScope(&sub_scope); + LOG(INFO) << "Tid: " << tid << ", process " << inputs.size() + << " samples, avg time per sample: " + << (stop_ms - start_ms) / inputs.size() << " ms"; +} + +TEST(inference, nlp) { + if (FLAGS_model_path.empty()) { + LOG(FATAL) << "Usage: ./example --model_path=path/to/your/model"; + } + if (FLAGS_data_file.empty()) { + LOG(WARNING) << "No data file provided, will use dummy data!" + << "Note: if you use nlp model, please provide data file."; + } + LOG(INFO) << "Model Path: " << FLAGS_model_path; + LOG(INFO) << "Data File: " << FLAGS_data_file; + + std::vector datasets; + size_t num_total_words = LoadData(&datasets, FLAGS_data_file); + LOG(INFO) << "Number of samples (seq_len<1024): " << datasets.size(); + LOG(INFO) << "Total number of words: " << num_total_words; + + // 0. Call `paddle::framework::InitDevices()` initialize all the devices + std::unique_ptr scope( + new paddle::framework::Scope()); + +#ifdef PADDLE_WITH_MKLML + // only use 1 thread number per std::thread + omp_set_dynamic(0); + omp_set_num_threads(1); + mkl_set_num_threads(1); +#endif + + double start_ms = 0, stop_ms = 0; + if (FLAGS_num_threads > 1) { + std::vector> jobs; + SplitData(datasets, &jobs, FLAGS_num_threads); + std::vector> threads; + start_ms = GetCurrentMs(); + for (int i = 0; i < FLAGS_num_threads; ++i) { + threads.emplace_back( + new std::thread(ThreadRunInfer, i, scope.get(), std::ref(jobs))); + } + for (int i = 0; i < FLAGS_num_threads; ++i) { + threads[i]->join(); + } + stop_ms = GetCurrentMs(); + } else { + // 1. Define place, executor, scope + auto place = paddle::platform::CPUPlace(); + auto executor = paddle::framework::Executor(place); + + // 2. Initialize the inference_program and load parameters + std::unique_ptr inference_program; + inference_program = InitProgram(&executor, scope.get(), FLAGS_model_path, + /*model combined*/ false); + // always prepare context + std::unique_ptr ctx; + ctx = executor.Prepare(*inference_program, 0); + if (FLAGS_prepare_vars) { + executor.CreateVariables(*inference_program, scope.get(), 0); + } + // preapre fetch + const std::vector& fetch_target_names = + inference_program->GetFetchTargetNames(); + PADDLE_ENFORCE_EQ(fetch_target_names.size(), 1UL); + std::map fetch_targets; + paddle::framework::LoDTensor outtensor; + fetch_targets[fetch_target_names[0]] = &outtensor; + + // prepare feed + const std::vector& feed_target_names = + inference_program->GetFeedTargetNames(); + PADDLE_ENFORCE_EQ(feed_target_names.size(), 1UL); + std::map feed_targets; + + // feed data and run + start_ms = GetCurrentMs(); + for (size_t i = 0; i < datasets.size(); ++i) { + feed_targets[feed_target_names[0]] = &(datasets[i]); + executor.RunPreparedContext(ctx.get(), scope.get(), &feed_targets, + &fetch_targets, !FLAGS_prepare_vars); + } + stop_ms = GetCurrentMs(); + LOG(INFO) << "Tid: 0, process " << datasets.size() + << " samples, avg time per sample: " + << (stop_ms - start_ms) / datasets.size() << " ms"; + } + LOG(INFO) << "Total inference time with " << FLAGS_num_threads + << " threads : " << (stop_ms - start_ms) / 1000.0 + << " sec, QPS: " << datasets.size() / ((stop_ms - start_ms) / 1000); +} diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 01b8dc0be6..44c36b1683 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -22,6 +22,8 @@ limitations under the License. */ #include "paddle/fluid/inference/io.h" #include "paddle/fluid/platform/profiler.h" +DECLARE_bool(use_mkldnn); + template void SetupTensor(paddle::framework::LoDTensor* input, paddle::framework::DDim dims, T lower, T upper) { @@ -133,24 +135,11 @@ std::vector> GetFeedTargetShapes( return feed_target_shapes; } -void EnableMKLDNN( - const std::unique_ptr& program) { - for (size_t bid = 0; bid < program->Size(); ++bid) { - auto* block = program->MutableBlock(bid); - for (auto* op : block->AllOps()) { - if (op->HasAttr("use_mkldnn")) { - op->SetAttr("use_mkldnn", true); - } - } - } -} - template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, const std::vector& cpu_fetchs, - const int repeat = 1, const bool is_combined = false, - const bool use_mkldnn = false) { + const int repeat = 1, const bool is_combined = false) { // 1. Define place, executor, scope auto place = Place(); auto executor = paddle::framework::Executor(place); @@ -182,9 +171,6 @@ void TestInference(const std::string& dirname, "init_program", paddle::platform::DeviceContextPool::Instance().Get(place)); inference_program = InitProgram(&executor, scope, dirname, is_combined); - if (use_mkldnn) { - EnableMKLDNN(inference_program); - } } // Disable the profiler and print the timing information paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, @@ -210,7 +196,10 @@ void TestInference(const std::string& dirname, fetch_targets[fetch_target_names[i]] = cpu_fetchs[i]; } - // 6. Run the inference program + // 6. If export Flags_use_mkldnn=True, use mkldnn related ops. + if (FLAGS_use_mkldnn) executor.EnableMKLDNN(*inference_program); + + // 7. Run the inference program { if (!CreateVars) { // If users don't want to create and destroy variables every time they diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index b4eca5bd9c..e4bb04295b 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -166,8 +166,6 @@ function(op_library TARGET) # NOTE(*): activation use macro to regist the kernels, set use_op manually. if(${TARGET} STREQUAL "activation") file(APPEND ${pybind_file} "USE_OP(relu);\n") - elseif(${TARGET} STREQUAL "reduce") - file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n") elseif(${TARGET} STREQUAL "fake_dequantize") file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n") else() @@ -188,19 +186,23 @@ endif() add_subdirectory(detail) if(WITH_DISTRIBUTE) - - set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf) + + set(DISTRIBUTE_DEPS "") + if(WITH_GRPC) + set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf) + else() + set(DISTRIBUTE_DEPS sendrecvop_brpc brpc leveldb snappystream snappy protobuf ssl crypto zlib) + endif() + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - op_library(send_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) op_library(prefetch_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(prefetch_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) op_library(recv_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) op_library(listen_and_serv_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(listen_and_serv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - op_library(send_vars_op DEPS ${DISTRIBUTE_DEPS}) - set_source_files_properties(send_vars_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + op_library(send_op DEPS ${DISTRIBUTE_DEPS}) + set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS}) op_library(fetch_barrier_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) @@ -210,15 +212,18 @@ if(WITH_DISTRIBUTE) # listen_and_serv_op sum_op executor SERIAL) if(WITH_GPU) set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op - listen_and_serv_op executor SERIAL) - op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc) + cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS listen_and_serv_op executor SERIAL) + if(WITH_GRPC) + op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc) + else() + op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_brpc) + endif() set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) else() set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op) endif() else() - set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op fetch_barrier_op gen_nccl_id_op) + set(DEPS_OPS ${DEPS_OPS} prefetch_op recv_op listen_and_serv_op send_op send_barrier_op fetch_barrier_op gen_nccl_id_op) endif() op_library(cross_entropy_op DEPS cross_entropy) @@ -227,6 +232,8 @@ op_library(softmax_op DEPS softmax) op_library(sequence_softmax_op DEPS softmax) if (WITH_GPU AND TENSORRT_FOUND) op_library(tensorrt_engine_op DEPS tensorrt_engine) + nv_test(test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc + DEPS tensorrt_engine_op tensorrt_engine tensorrt_converter) else() set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op) endif() diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc index b892ac77d9..46ed99bcf2 100644 --- a/paddle/fluid/operators/activation_mkldnn_op.cc +++ b/paddle/fluid/operators/activation_mkldnn_op.cc @@ -222,35 +222,35 @@ struct MKLDNNActivationGradFunc : public BaseActivationFunctor { }; template -using ReluMkldnnFunctor = +using ReluMKLDNNFunctor = MKLDNNActivationFunc; template -using TanhMkldnnFunctor = +using TanhMKLDNNFunctor = MKLDNNActivationFunc; template -using SqrtMkldnnFunctor = +using SqrtMKLDNNFunctor = MKLDNNActivationFunc; template -using AbsMkldnnFunctor = +using AbsMKLDNNFunctor = MKLDNNActivationFunc; template -using ReluMkldnnGradFunctor = +using ReluMKLDNNGradFunctor = MKLDNNActivationGradFunc; template -using TanhMkldnnGradFunctor = +using TanhMKLDNNGradFunctor = MKLDNNActivationGradFunc; template -using SqrtMkldnnGradFunctor = +using SqrtMKLDNNGradFunctor = MKLDNNActivationGradFunc; template -using AbsMkldnnGradFunctor = +using AbsMKLDNNGradFunctor = MKLDNNActivationGradFunc; } // namespace operators } // namespace paddle @@ -265,9 +265,9 @@ namespace ops = paddle::operators; ops::MKLDNNActivationGradKernel>); #define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro) \ - __macro(relu, ReluMkldnnFunctor, ReluMkldnnGradFunctor); \ - __macro(tanh, TanhMkldnnFunctor, TanhMkldnnGradFunctor); \ - __macro(sqrt, SqrtMkldnnFunctor, SqrtMkldnnGradFunctor); \ - __macro(abs, AbsMkldnnFunctor, AbsMkldnnGradFunctor); + __macro(relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor); \ + __macro(tanh, TanhMKLDNNFunctor, TanhMKLDNNGradFunctor); \ + __macro(sqrt, SqrtMKLDNNFunctor, SqrtMKLDNNGradFunctor); \ + __macro(abs, AbsMKLDNNFunctor, AbsMKLDNNGradFunctor); FOR_EACH_MKLDNN_KERNEL_FUNCTOR(REGISTER_ACTIVATION_MKLDNN_KERNEL); diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index dd71c66a75..af1d85047e 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -24,12 +24,12 @@ namespace operators { : public ::paddle::framework::OpProtoAndCheckerMaker { \ public: \ void Make() override { \ - AddInput("X", "Input of " #OP_NAME "operator"); \ - AddOutput("Out", "Output of" #OP_NAME "operator"); \ + AddInput("X", "Input of " #OP_NAME " operator"); \ + AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X"); \ AddAttr("use_mkldnn", \ "(bool, default false) Only used in mkldnn kernel") \ .SetDefault(false); \ - AddComment(#OP_COMMENT); \ + AddComment(OP_COMMENT); \ } \ } @@ -58,14 +58,16 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx, const framework::OperatorWithKernel& oper, const std::string& name) { framework::LibraryType library{framework::LibraryType::kPlain}; + + framework::DataLayout layout = framework::DataLayout::kAnyLayout; #ifdef PADDLE_WITH_MKLDNN auto it = oper.Attrs().find("use_mkldnn"); if (library == framework::LibraryType::kPlain && it != oper.Attrs().end() && platform::CanMKLDNNBeUsed(ctx)) { library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif - framework::DataLayout layout = framework::DataLayout::kAnyLayout; return framework::OpKernelType( framework::ToDataType(ctx.Input(name)->type()), ctx.GetPlace(), layout, library); diff --git a/paddle/fluid/operators/adam_op.cc b/paddle/fluid/operators/adam_op.cc index 99b0239855..6ee73c3000 100644 --- a/paddle/fluid/operators/adam_op.cc +++ b/paddle/fluid/operators/adam_op.cc @@ -89,9 +89,9 @@ class AdamOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Beta1Pow", "(Tensor) Input beta1 power accumulator"); AddInput("Beta2Pow", "(Tensor) Input beta2 power accumulator"); - AddOutput("ParamOut", "(Tensor) Output parameter"); - AddOutput("Moment1Out", "(Tensor) Output first moment"); - AddOutput("Moment2Out", "(Tensor) Output second moment"); + AddOutput("ParamOut", "(Tensor) Output parameter").Reuse("Param"); + AddOutput("Moment1Out", "(Tensor) Output first moment").Reuse("Moment1"); + AddOutput("Moment2Out", "(Tensor) Output second moment").Reuse("Moment2"); AddAttr("beta1", "(float, default 0.9) " diff --git a/paddle/fluid/operators/arg_max_op.cc b/paddle/fluid/operators/arg_max_op.cc new file mode 100644 index 0000000000..8174d37358 --- /dev/null +++ b/paddle/fluid/operators/arg_max_op.cc @@ -0,0 +1,33 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OPERATOR(arg_max, paddle::operators::ArgMinMaxOp, + paddle::operators::ArgMaxOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL( + arg_max, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel); diff --git a/paddle/fluid/operators/arg_max_op.cu b/paddle/fluid/operators/arg_max_op.cu new file mode 100644 index 0000000000..a147d77a9e --- /dev/null +++ b/paddle/fluid/operators/arg_max_op.cu @@ -0,0 +1,31 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OP_CUDA_KERNEL( + arg_max, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel, + paddle::operators::ArgMaxKernel); diff --git a/paddle/fluid/operators/arg_min_max_op_base.h b/paddle/fluid/operators/arg_min_max_op_base.h new file mode 100644 index 0000000000..6cbdaefeda --- /dev/null +++ b/paddle/fluid/operators/arg_min_max_op_base.h @@ -0,0 +1,160 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include +#include "paddle/fluid/framework/ddim.h" +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/string/printf.h" + +namespace paddle { +namespace operators { + +enum ArgMinMaxType { kArgMin, kArgMax }; + +template +struct ArgMinMaxFunctor {}; + +#define DECLARE_ARG_MIN_MAX_FUNCTOR(eigen_op_type, enum_argminmax_value) \ + template \ + struct ArgMinMaxFunctor { \ + void operator()(const DeviceContext& ctx, const framework::LoDTensor& in, \ + framework::LoDTensor* out, int64_t axis) { \ + auto in_eigen = framework::EigenTensor::From(in); \ + auto out_eigen = framework::EigenTensor::From(*out); \ + out_eigen.device(*(ctx.eigen_device())) = \ + in_eigen.eigen_op_type(axis).template cast(); \ + } \ + } + +DECLARE_ARG_MIN_MAX_FUNCTOR(argmin, ArgMinMaxType::kArgMin); +DECLARE_ARG_MIN_MAX_FUNCTOR(argmax, ArgMinMaxType::kArgMax); + +template +class ArgMinMaxKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& x = *(ctx.Input("X")); + auto& out = *(ctx.Output("Out")); + out.mutable_data(ctx.GetPlace()); + auto axis = ctx.Attr("axis"); + auto& dev_ctx = ctx.template device_context(); + +#define CALL_ARG_MINMAX_FUNCTOR(rank) \ + ArgMinMaxFunctor \ + functor##rank; \ + functor##rank(dev_ctx, x, &out, axis) + + switch (x.dims().size()) { + case 1: + CALL_ARG_MINMAX_FUNCTOR(1); + break; + case 2: + CALL_ARG_MINMAX_FUNCTOR(2); + break; + case 3: + CALL_ARG_MINMAX_FUNCTOR(3); + break; + case 4: + CALL_ARG_MINMAX_FUNCTOR(4); + break; + case 5: + CALL_ARG_MINMAX_FUNCTOR(5); + break; + case 6: + CALL_ARG_MINMAX_FUNCTOR(6); + break; + default: + PADDLE_THROW( + "%s operator doesn't supports tensors whose ranks are greater " + "than 6.", + (EnumArgMinMaxValue == kArgMin ? "argmin" : "argmax")); + break; +#undef CALL_ARG_MINMAX_FUNCTOR + } + } +}; + +template +using ArgMinKernel = + ArgMinMaxKernel; + +template +using ArgMaxKernel = + ArgMinMaxKernel; + +class ArgMinMaxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + const auto& x_dims = ctx->GetInputDim("X"); + int64_t axis = ctx->Attrs().Get("axis"); + PADDLE_ENFORCE(axis >= -x_dims.size() && axis < x_dims.size(), + "'axis' must be inside [-Rank(X), Rank(X))"); + + auto x_rank = x_dims.size(); + if (axis < 0) axis += x_rank; + + std::vector vec; + for (int64_t i = 0; i < axis; i++) vec.push_back(x_dims[i]); + for (int64_t i = axis + 1; i < x_rank; i++) vec.push_back(x_dims[i]); + ctx->SetOutputDim("Out", framework::make_ddim(vec)); + } +}; + +class BaseArgMinMaxOpMaker : public framework::OpProtoAndCheckerMaker { + protected: + virtual const char* OpName() const = 0; + virtual const char* Name() const = 0; + + public: + void Make() override { + AddInput("X", "Input tensor."); + AddOutput("Out", "Output tensor."); + AddAttr("axis", "The axis in which to compute the arg indics."); + AddComment(string::Sprintf(R"DOC( + %s Operator. + + Computes the indices of the %s elements of the input tensor's element + along the provided axis. +)DOC", + OpName(), Name())); + } +}; + +class ArgMinOpMaker : public BaseArgMinMaxOpMaker { + protected: + const char* OpName() const override { return "ArgMin"; } + const char* Name() const override { return "min"; } +}; + +class ArgMaxOpMaker : public BaseArgMinMaxOpMaker { + protected: + const char* OpName() const override { return "ArgMax"; } + const char* Name() const override { return "max"; } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/arg_min_op.cc b/paddle/fluid/operators/arg_min_op.cc new file mode 100644 index 0000000000..41f188029f --- /dev/null +++ b/paddle/fluid/operators/arg_min_op.cc @@ -0,0 +1,33 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OPERATOR(arg_min, paddle::operators::ArgMinMaxOp, + paddle::operators::ArgMinOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL( + arg_min, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel); diff --git a/paddle/fluid/operators/arg_min_op.cu b/paddle/fluid/operators/arg_min_op.cu new file mode 100644 index 0000000000..4d02050850 --- /dev/null +++ b/paddle/fluid/operators/arg_min_op.cu @@ -0,0 +1,31 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/arg_min_max_op_base.h" + +REGISTER_OP_CUDA_KERNEL( + arg_min, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel, + paddle::operators::ArgMinKernel); diff --git a/paddle/fluid/operators/batch_norm_mkldnn_op.cc b/paddle/fluid/operators/batch_norm_mkldnn_op.cc index 0e4a56d4a4..8206cc9890 100644 --- a/paddle/fluid/operators/batch_norm_mkldnn_op.cc +++ b/paddle/fluid/operators/batch_norm_mkldnn_op.cc @@ -19,10 +19,17 @@ limitations under the License. */ namespace paddle { namespace operators { -using Tensor = framework::Tensor; +using batch_norm_bwd = mkldnn::batch_normalization_backward; +using batch_norm_fwd = mkldnn::batch_normalization_forward; +using framework::DataLayout; +using framework::Tensor; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; +using mkldnn::stream; using paddle::platform::MKLDNNDeviceContext; using paddle::platform::MKLDNNMemDesc; -using mkldnn::memory; +using platform::to_void_cast; template using EigenArrayMap = @@ -64,21 +71,12 @@ void run_batch_norm_op(Args &&... args) { mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } -template -inline void *cast_const_to_void(const T *t) { - return static_cast(const_cast(t)); -} } // namespace template class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { - auto data_layout_str = ctx.Attr("data_layout"); - auto data_layout = framework::StringToDataLayout(data_layout_str); - PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, - "MKLDNN batch normalization handles only NCHW data layout"); - const float epsilon = ctx.Attr("epsilon"); const float momentum = ctx.Attr("momentum"); const bool is_test = ctx.Attr("is_test"); @@ -99,41 +97,53 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { const auto *scale = ctx.Input("Scale"); const auto *shift = ctx.Input("Bias"); - y->mutable_data(ctx.GetPlace()); - mean_out->mutable_data(ctx.GetPlace()); - variance_out->mutable_data(ctx.GetPlace()); + PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN && + x->format() != memory::format::format_undef, + "Wrong layout/format set for Input x tensor"); + + const T *x_data = x->data(); + const T *mean_data = mean->data(); + const T *variance_data = variance->data(); + T *y_data = y->mutable_data(ctx.GetPlace()); + T *mean_out_data = mean_out->mutable_data(ctx.GetPlace()); + T *variance_out_data = variance_out->mutable_data(ctx.GetPlace()); + T *batch_mean_data = nullptr; + T *batch_variance_data = nullptr; if (!is_test) { - batch_mean->mutable_data(ctx.GetPlace()); - batch_variance->mutable_data(ctx.GetPlace()); + batch_mean_data = batch_mean->mutable_data(ctx.GetPlace()); + batch_variance_data = batch_variance->mutable_data(ctx.GetPlace()); } auto propagation = is_test == true ? mkldnn::prop_kind::forward_scoring : mkldnn::prop_kind::forward_training; - auto dims = paddle::framework::vectorize2int(x->dims()); - - auto src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - - auto src_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine}; - auto dst_pd = mkldnn::memory::primitive_desc{dst_md, mkldnn_engine}; - - auto src = mkldnn::memory{src_pd, cast_const_to_void(x->data())}; - auto dst = mkldnn::memory{dst_pd, y->data()}; + auto src_tz = paddle::framework::vectorize2int(x->dims()); + auto scale_tz = paddle::framework::vectorize2int(scale->dims()); + PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1"); + const unsigned int ic = scale_tz[0]; unsigned flags = mkldnn::use_scale_shift; if (is_test) flags |= mkldnn::use_global_stats; + // create mkldnn memory from input x tensor + auto src_memory = + memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine}, + to_void_cast(x_data)); + + // create primitive descriptor for batch norm forward using bn_fwd_types = bn_type_traits; - auto batch_norm_fwd_desc = - bn_fwd_types::op_desc{propagation, src_md, epsilon, flags}; - auto batch_norm_fwd_pd = - bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ + propagation, src_memory.get_primitive_desc().desc(), epsilon, flags}; + std::shared_ptr batch_norm_fwd_pd = + std::shared_ptr( + new batch_norm_fwd::primitive_desc(batch_norm_fwd_desc, + mkldnn_engine)); - const unsigned int ic = dims[1]; + // Save the pd to be used in backward pass + const std::string key = ctx.op().Output("SavedMean"); + const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd"; + dev_ctx.SetBlob(key_batch_norm_fwd_pd, batch_norm_fwd_pd); // MKLDNN requires a single piece of memory for scale and shift/bias data const size_t scaleshift_size = 2 * ic; @@ -143,73 +153,58 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel { copy_to_weights(scale->data(), scale->data() + ic, shift->data(), shift->data() + ic, &scaleshift_data); - auto scaleshift_memory = mkldnn::memory{ - batch_norm_fwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + // crate mkldnn memory for weights(scale/shift) + auto scaleshift_memory = memory(batch_norm_fwd_pd->weights_primitive_desc(), + scaleshift_data.data()); - if (is_test) { - auto mean_memory = mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), - cast_const_to_void(mean->data())}; + // create mkldnn memory for output y tensor + auto dst_memory = memory(batch_norm_fwd_pd->dst_primitive_desc(), y_data); + if (is_test) { + // create mkldnn memory for stats (as input) + auto mean_memory = memory(batch_norm_fwd_pd->mean_primitive_desc(), + to_void_cast(mean_data)); auto variance_memory = - mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), - cast_const_to_void(variance->data())}; + memory(batch_norm_fwd_pd->variance_primitive_desc(), + to_void_cast(variance_data)); run_batch_norm_op( - batch_norm_fwd_pd, src, (const mkldnn::primitive::at &)mean_memory, + *batch_norm_fwd_pd, src_memory, + (const mkldnn::primitive::at &)mean_memory, (const mkldnn::primitive::at &)variance_memory, scaleshift_memory, - dst); + dst_memory); } else { + // create mkldnn memory for stats (as output) auto mean_memory = - mkldnn::memory{batch_norm_fwd_pd.mean_primitive_desc(), - cast_const_to_void(batch_mean->data())}; - - auto variance_memory = - mkldnn::memory{batch_norm_fwd_pd.variance_primitive_desc(), - cast_const_to_void(batch_variance->data())}; + memory(batch_norm_fwd_pd->mean_primitive_desc(), batch_mean_data); + auto variance_memory = memory( + batch_norm_fwd_pd->variance_primitive_desc(), batch_variance_data); - run_batch_norm_op(batch_norm_fwd_pd, src, - scaleshift_memory, dst, + run_batch_norm_op(*batch_norm_fwd_pd, src_memory, + scaleshift_memory, dst_memory, mean_memory, variance_memory); } if (!is_test) { - const unsigned int in = dims[0]; - const unsigned int sample_size = x->numel() / in / ic; - - // saved_xx is use just in this batch of data - EigenVectorArrayMap saved_mean_e( - batch_mean->mutable_data(ctx.GetPlace()), ic); - EigenVectorArrayMap saved_variance_e( - batch_variance->mutable_data(ctx.GetPlace()), ic); - saved_mean_e.setZero(); - saved_variance_e.setZero(); - - const unsigned int x_arr_size = in * ic; - ConstEigenArrayMap x_arr(x->data(), sample_size, x_arr_size); - for (unsigned int nc = 0; nc < x_arr_size; ++nc) { - saved_mean_e(nc % ic) += x_arr.col(nc).sum(); - } - saved_mean_e /= in * sample_size; - for (unsigned int nc = 0; nc < x_arr_size; ++nc) { - saved_variance_e(nc % ic) += - (x_arr.col(nc) - saved_mean_e(nc % ic)).matrix().squaredNorm(); - } - saved_variance_e /= in * sample_size; - - ConstEigenVectorArrayMap mean_arr{mean->data(), ic}; - ConstEigenVectorArrayMap variance_arr{variance->data(), ic}; - - EigenVectorArrayMap running_mean_arr( - mean_out->mutable_data(ctx.GetPlace()), ic); - EigenVectorArrayMap running_var_arr( - variance_out->mutable_data(ctx.GetPlace()), ic); + // mkldnn only compute stats for current batch + // so we need compute momentum stats via Eigen lib + EigenVectorArrayMap batch_mean_e(batch_mean_data, ic); + EigenVectorArrayMap batch_variance_e(batch_variance_data, ic); + ConstEigenVectorArrayMap mean_e(mean_data, ic); + ConstEigenVectorArrayMap variance_e{variance_data, ic}; + + EigenVectorArrayMap running_mean_e(mean_out_data, ic); + EigenVectorArrayMap running_variance_e(variance_out_data, ic); auto one_minus_momentum = 1. - momentum; - running_mean_arr = - mean_arr * momentum + saved_mean_e * one_minus_momentum; - running_var_arr = - variance_arr * momentum + saved_variance_e * one_minus_momentum; + running_mean_e = mean_e * momentum + batch_mean_e * one_minus_momentum; + running_variance_e = + variance_e * momentum + batch_variance_e * one_minus_momentum; } + + y->set_layout(DataLayout::kMKLDNN); + y->set_format( + (memory::format)dst_memory.get_primitive_desc().desc().data.format); } }; @@ -217,11 +212,6 @@ template class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { public: void Compute(const paddle::framework::ExecutionContext &ctx) const override { - auto data_layout_str = ctx.Attr("data_layout"); - auto data_layout = framework::StringToDataLayout(data_layout_str); - PADDLE_ENFORCE(data_layout == framework::DataLayout::kNCHW, - "MKLDNN batch normalization handles only NCHW data layout"); - auto &dev_ctx = ctx.template device_context(); auto mkldnn_engine = dev_ctx.GetEngine(); @@ -238,88 +228,132 @@ class BatchNormMKLDNNGradOpKernel : public paddle::framework::OpKernel { auto *diff_scale = ctx.Output(framework::GradVarName("Scale")); auto *diff_shift = ctx.Output(framework::GradVarName("Bias")); - diff_x->mutable_data(ctx.GetPlace()); - diff_scale->mutable_data(ctx.GetPlace()); - diff_shift->mutable_data(ctx.GetPlace()); + PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN && + diff_y->format() != memory::format::format_undef, + "Wrong layout/format set for Input diff_y tensor"); + + const T *x_data = x->data(); + const T *diff_y_data = diff_y->data(); + const T *batch_mean_data = batch_mean->data(); + const T *batch_variance_data = batch_variance->data(); + const T *scale_data = scale->data(); + const T *shift_data = shift->data(); + T *diff_x_data = diff_x->mutable_data(ctx.GetPlace()); + T *diff_scale_data = diff_scale->mutable_data(ctx.GetPlace()); + T *diff_shift_data = diff_shift->mutable_data(ctx.GetPlace()); + + auto src_tz = paddle::framework::vectorize2int(x->dims()); + auto diff_src_tz = src_tz; + auto dst_tz = src_tz; + auto diff_dst_tz = dst_tz; + auto scale_tz = paddle::framework::vectorize2int(scale->dims()); + PADDLE_ENFORCE(scale_tz.size() == 1, "Dims of scale tensor is NOT 1"); + + const unsigned int ic = scale_tz[0]; + + // Retrieve bn_fwd_pd from device context + const std::string key = ctx.op().Input("SavedMean"); + const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd"; + auto batch_norm_fwd_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_batch_norm_fwd_pd)); + PADDLE_ENFORCE(batch_norm_fwd_pd != nullptr, + "Fail to find batch_norm_fwd_pd in device context"); - auto dims = paddle::framework::vectorize2int(x->dims()); - unsigned flags = mkldnn::use_scale_shift | !mkldnn::use_global_stats; + using bn_bwd_types = bn_type_traits; - auto src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto diff_src_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); - auto diff_dst_md = - MKLDNNMemDesc(dims, memory::data_type::f32, memory::format::nchw); + // create mkldnn memory from input diff_y tensor + auto user_diff_dst_memory = + memory({{{diff_dst_tz}, memory::data_type::f32, diff_y->format()}, + mkldnn_engine}, + to_void_cast(diff_y_data)); - using bn_bwd_types = bn_type_traits; - using bn_fwd_types = bn_type_traits; + // create mkldnn memory from input x tensor + auto src_memory = + memory({{{src_tz}, memory::data_type::f32, x->format()}, mkldnn_engine}, + to_void_cast(x_data)); - auto batch_norm_fwd_desc = bn_fwd_types::op_desc{ - mkldnn::prop_kind::forward_training, src_md, epsilon, flags}; - auto batch_norm_fwd_pd = - bn_fwd_types::op_prim{batch_norm_fwd_desc, mkldnn_engine}; + // for diff_dst, try to use same format as dst in forward pass + auto diff_dst_pd = batch_norm_fwd_pd.get()->dst_primitive_desc(); + auto diff_dst_md = diff_dst_pd.desc(); + // create primitive descriptor for batch norm backward + unsigned flags = mkldnn::use_scale_shift; auto batch_norm_bwd_desc = bn_bwd_types::op_desc{ - mkldnn::prop_kind::backward, diff_dst_md, dst_md, epsilon, flags}; + mkldnn::prop_kind::backward, diff_dst_md, + src_memory.get_primitive_desc().desc(), epsilon, flags}; auto batch_norm_bwd_pd = bn_bwd_types::op_prim{ - batch_norm_bwd_desc, mkldnn_engine, batch_norm_fwd_pd}; - - auto src = mkldnn::memory{{src_md, mkldnn_engine}, - cast_const_to_void(x->data())}; - - auto mean = mkldnn::memory{batch_norm_bwd_pd.mean_primitive_desc(), - cast_const_to_void(batch_mean->data())}; - - auto variance = - mkldnn::memory{batch_norm_bwd_pd.variance_primitive_desc(), - cast_const_to_void(batch_variance->data())}; - - auto diff_dst = mkldnn::memory{{diff_dst_md, mkldnn_engine}, - cast_const_to_void(diff_y->data())}; + batch_norm_bwd_desc, mkldnn_engine, *batch_norm_fwd_pd}; + + // reorder user_diff_dst if it's not in preferred format + auto diff_dst_memory = user_diff_dst_memory; + primitive reorder_diff_dst; + bool is_diff_dst_reordered = false; + if (diff_dst_pd != user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory = memory(diff_dst_pd); + reorder_diff_dst = reorder(user_diff_dst_memory, diff_dst_memory); + is_diff_dst_reordered = true; + } - const unsigned int ic = dims[1]; + // create mkldnn memory for input tensors (src/mean/variance) + auto mean_memory = memory(batch_norm_bwd_pd.mean_primitive_desc(), + to_void_cast(batch_mean_data)); + auto variance_memory = memory(batch_norm_bwd_pd.variance_primitive_desc(), + to_void_cast(batch_variance_data)); + // MKLDNN requires a single piece of memory for scale and shift/bias data const size_t scaleshift_size = 2 * ic; std::vector scaleshift_data; scaleshift_data.reserve(scaleshift_size); - copy_to_weights(scale->data(), scale->data() + ic, shift->data(), - shift->data() + ic, &scaleshift_data); + copy_to_weights(scale_data, scale_data + ic, shift_data, shift_data + ic, + &scaleshift_data); - auto scaleshift_memory = mkldnn::memory{ - batch_norm_bwd_pd.weights_primitive_desc(), scaleshift_data.data()}; + // create mkldnn memory for input tensors (scale/shift) + auto scaleshift_memory = memory(batch_norm_bwd_pd.weights_primitive_desc(), + scaleshift_data.data()); + // create mkldnn memory for output diff weights (combined scale/shift) std::vector diff_scaleshift_data; diff_scaleshift_data.reserve(scaleshift_size); - copy_to_weights(diff_scale->data(), diff_scale->data() + ic, - diff_shift->data(), diff_shift->data() + ic, - &diff_scaleshift_data); - auto diff_scaleshift_memory = - mkldnn::memory{batch_norm_bwd_pd.diff_weights_primitive_desc(), - diff_scaleshift_data.data()}; - - auto diff_src = mkldnn::memory{{diff_src_md, mkldnn_engine}, - static_cast(diff_x->data())}; - - run_batch_norm_op( - batch_norm_bwd_pd, src, mean, variance, diff_dst, scaleshift_memory, - diff_src, diff_scaleshift_memory); - + memory(batch_norm_bwd_pd.diff_weights_primitive_desc(), + diff_scaleshift_data.data()); + + // here assume diff_src is in the same format of src + auto diff_src_memory = memory(src_memory.get_primitive_desc(), diff_x_data); + + // finally create batch_norm backward primitive + auto batch_norm_bwd_prim = + batch_norm_bwd(batch_norm_bwd_pd, src_memory, mean_memory, + variance_memory, diff_dst_memory, scaleshift_memory, + diff_src_memory, diff_scaleshift_memory); + + // execute optional reorder and batch_norm backward primitive + std::vector pipeline; + if (is_diff_dst_reordered) pipeline.push_back(reorder_diff_dst); + pipeline.push_back(batch_norm_bwd_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + // copy back diff sacle/shift to output tensors (diff scale/shift) + diff_scaleshift_data.resize(scaleshift_size); auto it = std::begin(diff_scaleshift_data); - std::copy(it, std::next(it, ic), diff_scale->data()); + std::copy(it, std::next(it, ic), diff_scale_data); std::copy(std::next(it, ic), std::end(diff_scaleshift_data), - diff_shift->data()); + diff_shift_data); + + // set layout/format of output tensors + diff_x->set_layout(DataLayout::kMKLDNN); + diff_x->set_format((memory::format)diff_src_memory.get_primitive_desc() + .desc() + .data.format); } }; } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_KERNEL(batch_norm, MKLDNN, paddle::platform::CPUPlace, +REGISTER_OP_KERNEL(batch_norm, MKLDNN, ::paddle::platform::CPUPlace, ops::BatchNormMKLDNNOpKernel); -REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, paddle::platform::CPUPlace, +REGISTER_OP_KERNEL(batch_norm_grad, MKLDNN, ::paddle::platform::CPUPlace, ops::BatchNormMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 6ec8c9d18b..625ca2d7c4 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -110,17 +110,19 @@ class BatchNormOp : public framework::OperatorWithKernel { ctx.Input("Variance")->type()), "Variance input should be of float type"); - framework::LibraryType library_{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; #ifdef PADDLE_WITH_MKLDNN - if (library_ == framework::LibraryType::kPlain && + if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kMKLDNN; + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, - library_); + library); } }; @@ -149,13 +151,15 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Variance", "The global variance (for training) " "or estimated Variance (for testing)"); - AddOutput("Y", "result after normalization"); + AddOutput("Y", "result after normalization").Reuse("X"); AddOutput("MeanOut", "Share memory with Mean. " - "Store the global mean when training"); + "Store the global mean when training") + .Reuse("Mean"); AddOutput("VarianceOut", "Share memory with Variance. " - "Store the global Variance when training"); + "Store the global Variance when training") + .Reuse("Variance"); AddOutput("SavedMean", "Mean of the current mini batch, " "will apply to output when training") @@ -366,18 +370,21 @@ class BatchNormGradOp : public framework::OperatorWithKernel { PADDLE_THROW("can't find Y@GRAD"); } - framework::LibraryType library_{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::LibraryType library = framework::LibraryType::kPlain; + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + #ifdef PADDLE_WITH_MKLDNN - if (library_ == framework::LibraryType::kPlain && + if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kMKLDNN; + library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - layout, library_); + layout, library); } }; diff --git a/paddle/fluid/operators/batch_size_like.h b/paddle/fluid/operators/batch_size_like.h index 483c9f8c21..fc15d56891 100644 --- a/paddle/fluid/operators/batch_size_like.h +++ b/paddle/fluid/operators/batch_size_like.h @@ -54,18 +54,18 @@ class BatchSizeLikeOp : public framework::OperatorWithKernel { class BatchSizeLikeOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() final { - AddInput("Input", - "(Tensor) Tensor " - "whose input_dim_idx'th dimension specifies the batch_size"); + AddInput( + "Input", + "Tensor whose input_dim_idx'th dimension specifies the batch_size"); AddOutput("Out", - "(Tensor) Tensor of specified shape will be filled " + "Tensor of specified shape will be filled " "with the specified value"); - AddAttr>("shape", "(vector) The shape of the output"); + AddAttr>("shape", "The shape of the output"); AddAttr("input_dim_idx", - "(int, default 0) The index of input's batch size dimension") + "default 0. The index of input's batch size dimension") .SetDefault(0); AddAttr("output_dim_idx", - "(int, default 0) The index of output's batch size dimension") + "default 0. The index of output's batch size dimension") .SetDefault(0); Apply(); } diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc index d46fda54e7..2572e813d6 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cc +++ b/paddle/fluid/operators/bilinear_interp_op.cc @@ -34,22 +34,38 @@ class BilinearInterpOp : public framework::OperatorWithKernel { int out_w = ctx->Attrs().Get("out_w"); PADDLE_ENFORCE_EQ(dim_x.size(), 4, "X's dimension must be 4"); + if (ctx->HasInput("OutSize")) { + auto out_size_dim = ctx->GetInputDim("OutSize"); + PADDLE_ENFORCE_EQ(out_size_dim.size(), 1, + "OutSize's dimension size must be 1"); + PADDLE_ENFORCE_EQ(out_size_dim[0], 2, "OutSize's dim[0] must be 2"); + } std::vector dim_out({dim_x[0], dim_x[1], out_h, out_w}); ctx->SetOutputDim("Out", framework::make_ddim(dim_out)); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace()); + } }; class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "(Tensor) The input tensor of bilinear interpolation, " + "The input tensor of bilinear interpolation, " "This is a 4-D tensor with shape of (N x C x h x w)"); - AddOutput("Out", - "(Tensor) The dimension of output is (N x C x out_h x out_w]"); + AddInput("OutSize", + "This is a 1-D tensor with two number. " + "The first number is height and the second number is width.") + .AsDispensable(); + AddOutput("Out", "The dimension of output is (N x C x out_h x out_w)"); - AddAttr("out_h", "(int) output height of bilinear interpolation op."); - AddAttr("out_w", "(int) output width of bilinear interpolation op."); + AddAttr("out_h", "output height of bilinear interpolation op."); + AddAttr("out_w", "output width of bilinear interpolation op."); AddComment(R"DOC( Bilinear interpolation is an extension of linear interpolation for interpolating functions of two variables (e.g. H-direction and @@ -78,6 +94,12 @@ class BilinearInterpOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), dim_x); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace()); + } }; } // namespace operators diff --git a/paddle/fluid/operators/bilinear_interp_op.cu b/paddle/fluid/operators/bilinear_interp_op.cu index 510190f1aa..4c19715384 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cu +++ b/paddle/fluid/operators/bilinear_interp_op.cu @@ -102,10 +102,21 @@ class BilinearInterpOpCUDAKernel : public framework::OpKernel { auto* input_t = ctx.Input("X"); // float tensor auto* output_t = ctx.Output("Out"); // float tensor auto* input = input_t->data(); - auto* output = output_t->mutable_data(ctx.GetPlace()); int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + auto out_dims = output_t->dims(); + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + auto* output = output_t->mutable_data( + {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); + int batch_size = input_t->dims()[0]; int channels = input_t->dims()[1]; int in_h = input_t->dims()[2]; @@ -139,8 +150,8 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* d_input_t = ctx.Output(framework::GradVarName("X")); auto* d_output_t = ctx.Input(framework::GradVarName("Out")); - auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto* d_output = d_output_t->data(); + auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto& device_ctx = ctx.template device_context(); @@ -149,6 +160,16 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + int batch_size = d_input_t->dims()[0]; int channels = d_input_t->dims()[1]; int in_h = d_input_t->dims()[2]; diff --git a/paddle/fluid/operators/bilinear_interp_op.h b/paddle/fluid/operators/bilinear_interp_op.h index f6cd77e4d4..8b03cd5a06 100644 --- a/paddle/fluid/operators/bilinear_interp_op.h +++ b/paddle/fluid/operators/bilinear_interp_op.h @@ -24,11 +24,18 @@ class BilinearInterpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* input_t = ctx.Input("X"); // float tensor auto* output_t = ctx.Output("Out"); // float tensor + auto out_dims = output_t->dims(); auto* input = input_t->data(); - auto* output = output_t->mutable_data(ctx.GetPlace()); - int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + auto out_size_data = out_size_t->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + auto* output = output_t->mutable_data( + {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); int batch_size = input_t->dims()[0]; int channels = input_t->dims()[1]; int in_h = input_t->dims()[2]; @@ -83,9 +90,8 @@ class BilinearInterpGradKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* d_input_t = ctx.Output(framework::GradVarName("X")); auto* d_output_t = ctx.Input(framework::GradVarName("Out")); - auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto* d_output = d_output_t->data(); - + auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto& device_ctx = ctx.template device_context(); math::SetConstant zero; @@ -93,6 +99,14 @@ class BilinearInterpGradKernel : public framework::OpKernel { int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + auto out_size_data = out_size_t->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + int batch_size = d_input_t->dims()[0]; int channels = d_input_t->dims()[1]; int in_h = d_input_t->dims()[2]; diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc index 38337f9aa5..c724055937 100644 --- a/paddle/fluid/operators/concat_op.cc +++ b/paddle/fluid/operators/concat_op.cc @@ -107,7 +107,13 @@ REGISTER_OPERATOR(concat, ops::ConcatOp, ops::ConcatOpMaker, false> /* set false to disable empty grad */); REGISTER_OPERATOR(concat_grad, ops::ConcatOpGrad); REGISTER_OP_CPU_KERNEL( - concat, ops::ConcatKernel); + concat, ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel); REGISTER_OP_CPU_KERNEL( concat_grad, - ops::ConcatGradKernel); + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel); diff --git a/paddle/fluid/operators/concat_op.cu.cc b/paddle/fluid/operators/concat_op.cu.cc index 590eca9d06..8e38e5231f 100644 --- a/paddle/fluid/operators/concat_op.cu.cc +++ b/paddle/fluid/operators/concat_op.cu.cc @@ -15,7 +15,13 @@ limitations under the License. */ #include "paddle/fluid/operators/concat_op.h" namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( - concat, ops::ConcatKernel); + concat, ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel, + ops::ConcatKernel); REGISTER_OP_CUDA_KERNEL( concat_grad, - ops::ConcatGradKernel); + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel, + ops::ConcatGradKernel); diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index 7a7b8b76e4..1828be57b5 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -20,7 +20,7 @@ limitations under the License. */ #include "paddle/fluid/platform/cudnn_helper.h" #include "paddle/fluid/platform/float16.h" -DEFINE_bool(cudnn_algo_use_autotune, true, +DEFINE_bool(cudnn_deterministic, true, "Whether allow using an autotuning algorithm for convolution " "operator. The autotuning algorithm may be non-deterministic. If " "false, the algorithm is deterministic."); @@ -272,7 +272,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { auto& dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); if (input_grad) { - if (FLAGS_cudnn_algo_use_autotune) { + if (FLAGS_cudnn_deterministic) { PADDLE_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( handle, cudnn_filter_desc, @@ -297,7 +297,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { } if (filter_grad) { - if (FLAGS_cudnn_algo_use_autotune) { + if (FLAGS_cudnn_deterministic) { PADDLE_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm( handle, cudnn_input_desc, cudnn_output_grad_desc, diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index 63d371310d..6b06913d1c 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -18,6 +18,17 @@ namespace paddle { namespace operators { +using conv_bwd_data = mkldnn::convolution_backward_data; +using conv_bwd_weights = mkldnn::convolution_backward_weights; +using conv_fwd = mkldnn::convolution_forward; +using framework::DataLayout; +using mkldnn::memory; +using mkldnn::primitive; +using mkldnn::reorder; +using mkldnn::stream; +using platform::to_void_cast; +using platform::GetMKLDNNFormat; + template class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { public: @@ -25,6 +36,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace."); + // Get unique name for index + const std::string key = ctx.op().Output("Output"); + const std::string key_conv_pd = key + "@conv_pd"; + auto& dev_ctx = ctx.template device_context(); const auto& mkldnn_engine = dev_ctx.GetEngine(); @@ -33,10 +48,12 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto* filter = ctx.Input("Filter"); auto* output = ctx.Output("Output"); - // Get an unique name from "argument" name of "Output" variable - // This name will be used as key when saving info into device context - const std::string key = ctx.op().Output("Output"); - const std::string key_conv_pd = key + "@conv_pd"; + PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && + input->format() != memory::format::format_undef, + "Wrong layout/format set for Input tensor"); + PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN && + filter->format() != memory::format::format_undef, + "Wrong layout/format set for Filter tensor"); std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); @@ -63,60 +80,86 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { paddle::framework::vectorize2int(filter->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); - // TODO(pzelazko-intel): support more formats - auto src_md = platform::MKLDNNMemDesc( - src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - auto weights_md = - platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, - mkldnn::memory::format::oihw); - auto dst_md = platform::MKLDNNMemDesc( - dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - - auto src_memory = - mkldnn::memory({src_md, mkldnn_engine}, - reinterpret_cast(const_cast(input_data))); - auto weights_memory = - mkldnn::memory({weights_md, mkldnn_engine}, - reinterpret_cast(const_cast(filter_data))); - auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, output_data); - - std::shared_ptr conv_pd = - ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings, - mkldnn_engine); - - // save conv_pd into global device context to be referred in backward path - dev_ctx.SetBlob(key_conv_pd, conv_pd); + // create mkldnn memory from input tensors (data/weights) + auto user_src_memory = memory( + {{{src_tz}, memory::data_type::f32, input->format()}, mkldnn_engine}, + to_void_cast(input_data)); + auto user_weights_memory = + memory({{{weights_tz}, memory::data_type::f32, filter->format()}, + mkldnn_engine}, + to_void_cast(filter_data)); + + /* create memory descriptor for convolution without specified format + * ('any') which lets a primitive (convolution in this case) choose + * the memory format preferred for best performance + */ + auto src_md = platform::MKLDNNMemDesc(src_tz, memory::data_type::f32, + memory::format::any); + auto weights_md = platform::MKLDNNMemDesc( + weights_tz, memory::data_type::f32, memory::format::any); + auto dst_md = platform::MKLDNNMemDesc(dst_tz, memory::data_type::f32, + memory::format::any); + + // create a conv primitive descriptor and save it for usage in backward + std::shared_ptr conv_pd = ConvFwdPrimitiveDesc( + src_md, weights_md, dst_md, strides, paddings, mkldnn_engine); + + // create reorder primitive if the input format is not the preferred one + auto src_memory = user_src_memory; + primitive reorder_src; + bool is_src_reordered = false; + if (memory::primitive_desc(conv_pd->src_primitive_desc()) != + user_src_memory.get_primitive_desc()) { + src_memory = memory(conv_pd->src_primitive_desc()); + reorder_src = reorder(user_src_memory, src_memory); + is_src_reordered = true; + } + auto weights_memory = user_weights_memory; + primitive reorder_weights; + bool is_weights_reordered = false; + if (memory::primitive_desc(conv_pd->weights_primitive_desc()) != + user_weights_memory.get_primitive_desc()) { + weights_memory = memory(conv_pd->weights_primitive_desc()); + reorder_weights = reorder(user_weights_memory, weights_memory); + is_weights_reordered = true; + } + + // create memory primitive for conv dst + auto dst_memory = memory(conv_pd->dst_primitive_desc(), output_data); // create convolution op primitive - auto conv_prim = mkldnn::convolution_forward(*conv_pd, src_memory, - weights_memory, dst_memory); + auto conv_prim = conv_fwd(*conv_pd, src_memory, weights_memory, dst_memory); // push primitive to stream and wait until it's executed - std::vector pipeline{conv_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + std::vector pipeline; + if (is_src_reordered) pipeline.push_back(reorder_src); + if (is_weights_reordered) pipeline.push_back(reorder_weights); + pipeline.push_back(conv_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + // Save conv_pd/src_memory/weights_memory for backward pass + dev_ctx.SetBlob(key_conv_pd, conv_pd); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(GetMKLDNNFormat(dst_memory)); } private: - std::unique_ptr - ConvFwdPrimitiveDesc(const mkldnn::memory::desc& src, - const mkldnn::memory::desc& weights, - const mkldnn::memory::desc& dst, - const std::vector& strides, - const std::vector& paddings, - const mkldnn::engine& engine) const { - mkldnn::memory::dims stride_dims = {strides[0], strides[1]}; - mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]}; - - auto conv_desc = mkldnn::convolution_forward::desc( - mkldnn::prop_kind::forward, mkldnn::convolution_direct, src, weights, - dst, stride_dims, padding_dims, padding_dims, - mkldnn::padding_kind::zero); - - auto p_conv_pd = - new mkldnn::convolution_forward::primitive_desc(conv_desc, engine); - - return std::unique_ptr( - p_conv_pd); + std::unique_ptr ConvFwdPrimitiveDesc( + const memory::desc& src, const memory::desc& weights, + const memory::desc& dst, const std::vector& strides, + const std::vector& paddings, const mkldnn::engine& engine) const { + memory::dims stride_dims = {strides[0], strides[1]}; + memory::dims padding_dims = {paddings[0], paddings[1]}; + + auto conv_desc = + conv_fwd::desc(mkldnn::prop_kind::forward, mkldnn::convolution_direct, + src, weights, dst, stride_dims, padding_dims, + padding_dims, mkldnn::padding_kind::zero); + + auto p_conv_pd = new conv_fwd::primitive_desc(conv_desc, engine); + + return std::unique_ptr(p_conv_pd); } }; @@ -139,6 +182,19 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { Tensor* input_grad = ctx.Output(framework::GradVarName("Input")); Tensor* filter_grad = ctx.Output(framework::GradVarName("Filter")); + PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && + input->format() != memory::format::format_undef, + "Wrong layout/format set for Input tensor"); + PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN && + filter->format() != memory::format::format_undef, + "Wrong layout/format set for Filter tensor"); + PADDLE_ENFORCE(output->layout() == DataLayout::kMKLDNN && + output->format() != memory::format::format_undef, + "Wrong layout/format set for Output tensor"); + PADDLE_ENFORCE(output_grad->layout() == DataLayout::kMKLDNN && + output_grad->format() != memory::format::format_undef, + "Wrong layout/format set for output_grad tensor"); + if (!input_grad && !filter_grad) return; // Get an unique name from "argument" name of "Output" variable @@ -167,108 +223,147 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { paddle::framework::vectorize2int(filter->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); - // TODO(pzelazko-intel): support more formats - auto src_md = platform::MKLDNNMemDesc( - src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - auto diff_src_md = platform::MKLDNNMemDesc( - src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - auto weights_md = - platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, - mkldnn::memory::format::oihw); - auto diff_weights_md = - platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, - mkldnn::memory::format::oihw); - auto diff_dst_md = platform::MKLDNNMemDesc( - dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); - - // create memory - auto diff_dst_memory = mkldnn::memory( - {diff_weights_md, mkldnn_engine}, - reinterpret_cast(const_cast(output_grad_data))); + // create mkldnn memory from input tensors (input/weights/output_grad) + auto user_src_memory = memory( + {{{src_tz}, memory::data_type::f32, input->format()}, mkldnn_engine}, + to_void_cast(input_data)); + auto user_weights_memory = + memory({{{weights_tz}, memory::data_type::f32, filter->format()}, + mkldnn_engine}, + to_void_cast(filter_data)); + auto user_diff_dst_memory = + memory({{{dst_tz}, memory::data_type::f32, output_grad->format()}, + mkldnn_engine}, + to_void_cast(output_grad_data)); + + /* create memory descriptor for conv backward without specified format + * ('any') which lets a primitive (conv backward in this case) choose + * the memory format preferred for best performance + */ + auto src_md = platform::MKLDNNMemDesc(src_tz, memory::data_type::f32, + memory::format::any); + auto diff_src_md = platform::MKLDNNMemDesc(src_tz, memory::data_type::f32, + memory::format::any); + auto weights_md = platform::MKLDNNMemDesc( + weights_tz, memory::data_type::f32, memory::format::any); + auto diff_weights_md = platform::MKLDNNMemDesc( + weights_tz, memory::data_type::f32, memory::format::any); + auto diff_dst_md = platform::MKLDNNMemDesc(dst_tz, memory::data_type::f32, + memory::format::any); + // Retrieve conv_pd from device context - auto conv_pd = - std::static_pointer_cast( - dev_ctx.GetBlob(key_conv_pd)); + auto conv_pd = std::static_pointer_cast( + dev_ctx.GetBlob(key_conv_pd)); PADDLE_ENFORCE(conv_pd != nullptr, "Fail to find conv_pd in device context"); // create backward conv primitive for weights if (filter_grad) { - // create primitive descriptor - mkldnn::convolution_backward_weights::primitive_desc conv_bwd_weights_pd = - ConvBwdWeightsPrimitiveDesc(src_md, diff_weights_md, diff_dst_md, - strides, paddings, *conv_pd, - mkldnn_engine); - - // create memory + // create backward convolution primitive descriptor + auto conv_bwd_weights_desc = conv_bwd_weights::desc( + mkldnn::convolution_direct, src_md, diff_weights_md, diff_dst_md, + strides, paddings, paddings, mkldnn::padding_kind::zero); + auto conv_bwd_weights_pd = conv_bwd_weights::primitive_desc( + conv_bwd_weights_desc, mkldnn_engine, *conv_pd); + + // create reorder primitive if the input format is not the preferred one + auto src_memory = user_src_memory; + primitive reorder_src; + bool is_src_reordered = false; + if (memory::primitive_desc(conv_bwd_weights_pd.src_primitive_desc()) != + user_src_memory.get_primitive_desc()) { + src_memory = memory(conv_bwd_weights_pd.src_primitive_desc()); + reorder_src = reorder(user_src_memory, src_memory); + is_src_reordered = true; + } + + auto diff_dst_memory_4filter = user_diff_dst_memory; + primitive reorder_diff_dst_4filter; + bool is_diff_dst_reordered_4filter = false; + if (memory::primitive_desc( + conv_bwd_weights_pd.diff_dst_primitive_desc()) != + user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory_4filter = + memory(conv_bwd_weights_pd.diff_dst_primitive_desc()); + reorder_diff_dst_4filter = + reorder(user_diff_dst_memory, diff_dst_memory_4filter); + is_diff_dst_reordered_4filter = true; + } + + // create mkldnn memory for output (i.e. diff weights) auto diff_weights_memory = - mkldnn::memory({diff_weights_md, mkldnn_engine}, - reinterpret_cast(filter_grad_data)); - auto src_memory = - mkldnn::memory({src_md, mkldnn_engine}, - reinterpret_cast(const_cast(input_data))); + memory(conv_bwd_weights_pd.diff_weights_primitive_desc(), + reinterpret_cast(filter_grad_data)); // create backward conv primitive for weights - auto conv_bwd_weights_prim = mkldnn::convolution_backward_weights( - conv_bwd_weights_pd, src_memory, diff_dst_memory, - diff_weights_memory); + auto conv_bwd_weights_prim = + conv_bwd_weights(conv_bwd_weights_pd, src_memory, + diff_dst_memory_4filter, diff_weights_memory); // push primitive and execute it - std::vector pipeline{conv_bwd_weights_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + std::vector pipeline; + if (is_src_reordered) pipeline.push_back(reorder_src); + if (is_diff_dst_reordered_4filter) + pipeline.push_back(reorder_diff_dst_4filter); + pipeline.push_back(conv_bwd_weights_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + filter_grad->set_layout(DataLayout::kMKLDNN); + filter_grad->set_format(GetMKLDNNFormat(diff_weights_memory)); } if (input_grad) { - // create primitive descriptor - mkldnn::convolution_backward_data::primitive_desc conv_bwd_data_pd = - ConvBwdDataPrimitiveDesc(diff_src_md, weights_md, diff_dst_md, - strides, paddings, *conv_pd, mkldnn_engine); - - // create memory - auto diff_src_memory = mkldnn::memory( - {diff_src_md, mkldnn_engine}, - reinterpret_cast(const_cast(input_grad_data))); - auto weights_memory = - mkldnn::memory({weights_md, mkldnn_engine}, - reinterpret_cast(const_cast(filter_data))); + // create backward convolution primitive descriptor + auto conv_bwd_data_desc = conv_bwd_data::desc( + mkldnn::convolution_direct, diff_src_md, weights_md, diff_dst_md, + strides, paddings, paddings, mkldnn::padding_kind::zero); + auto conv_bwd_data_pd = conv_bwd_data::primitive_desc( + conv_bwd_data_desc, mkldnn_engine, *conv_pd); + + // create reorder primitive if the input format is not the preferred one + auto weights_memory = user_weights_memory; + primitive reorder_weights; + bool is_weights_reordered = false; + if (memory::primitive_desc(conv_bwd_data_pd.weights_primitive_desc()) != + user_weights_memory.get_primitive_desc()) { + weights_memory = memory(conv_bwd_data_pd.weights_primitive_desc()); + reorder_weights = reorder(user_weights_memory, weights_memory); + is_weights_reordered = true; + } + + auto diff_dst_memory_4data = user_diff_dst_memory; + primitive reorder_diff_dst_4data; + bool is_diff_dst_reordered_4data = false; + if (memory::primitive_desc(conv_bwd_data_pd.diff_dst_primitive_desc()) != + user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory_4data = + memory(conv_bwd_data_pd.diff_dst_primitive_desc()); + reorder_diff_dst_4data = + reorder(user_diff_dst_memory, diff_dst_memory_4data); + is_diff_dst_reordered_4data = true; + } + + // create mkldnn memory for output (i.e. diff src) + auto diff_src_memory = memory(conv_bwd_data_pd.diff_src_primitive_desc(), + reinterpret_cast(input_grad_data)); // create backward conv primitive for data - auto conv_bwd_data_prim = mkldnn::convolution_backward_data( - conv_bwd_data_pd, diff_dst_memory, weights_memory, diff_src_memory); + auto conv_bwd_data_prim = + conv_bwd_data(conv_bwd_data_pd, diff_dst_memory_4data, weights_memory, + diff_src_memory); - // push primitive to stream and wait until it's executed - std::vector pipeline{conv_bwd_data_prim}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + // push primitive and execute it + std::vector pipeline; + if (is_weights_reordered) pipeline.push_back(reorder_weights); + if (is_diff_dst_reordered_4data) + pipeline.push_back(reorder_diff_dst_4data); + pipeline.push_back(conv_bwd_data_prim); + stream(stream::kind::eager).submit(pipeline).wait(); + + input_grad->set_layout(DataLayout::kMKLDNN); + input_grad->set_format(GetMKLDNNFormat(diff_src_memory)); } } // Compute() - - private: - mkldnn::convolution_backward_weights::primitive_desc - ConvBwdWeightsPrimitiveDesc( - const mkldnn::memory::desc& src, const mkldnn::memory::desc& diff_weights, - const mkldnn::memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const mkldnn::convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine) const { - auto conv_bwd_weights_desc = mkldnn::convolution_backward_weights::desc( - mkldnn::convolution_direct, src, diff_weights, diff_dst, strides, - paddings, paddings, mkldnn::padding_kind::zero); - return mkldnn::convolution_backward_weights::primitive_desc( - conv_bwd_weights_desc, engine, conv_pd); - } - - mkldnn::convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc( - const mkldnn::memory::desc& diff_src, const mkldnn::memory::desc& weights, - const mkldnn::memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const mkldnn::convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine) const { - auto conv_bwd_data_desc = mkldnn::convolution_backward_data::desc( - mkldnn::convolution_direct, diff_src, weights, diff_dst, strides, - paddings, paddings, mkldnn::padding_kind::zero); - return mkldnn::convolution_backward_data::primitive_desc(conv_bwd_data_desc, - engine, conv_pd); - } }; } // namespace operators diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 697d914842..37153d5843 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -75,6 +75,10 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { framework::LibraryType library{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library = framework::LibraryType::kCUDNN; @@ -84,6 +88,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library = framework::LibraryType::kMKLDNN; + layout = framework::DataLayout::kMKLDNN; } #endif @@ -99,9 +104,6 @@ framework::OpKernelType ConvOp::GetExpectedKernelType( "float16 can only be used when CUDNN is used"); } - std::string data_format = ctx.Attr("data_format"); - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout = framework::StringToDataLayout(data_format); return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, library); } @@ -122,7 +124,8 @@ void Conv2DOpMaker::Make() { "input image channels divided by the groups."); AddOutput("Output", "(Tensor) The output tensor of convolution operator. " - "The format of output tensor is also NCHW."); + "The format of output tensor is also NCHW.") + .Reuse("Input"); AddAttr>("strides", "(vector default:{1, 1}), the " "strides(h_stride, w_stride) of " @@ -217,7 +220,8 @@ void Conv3DOpMaker::Make() { "input image channels divided by the groups."); AddOutput("Output", "(Tensor) The output tensor of convolution operator." - "The format of output tensor is also NCDHW."); + "The format of output tensor is also NCDHW.") + .Reuse("Input"); AddAttr>("strides", "(vector, default:{1, 1, 1}), the " "strides(d_stride, h_stride, w_stride) of " @@ -309,6 +313,10 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvOpGrad::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -318,12 +326,10 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType( if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), layout_, library_); diff --git a/paddle/fluid/operators/crop_op.cc b/paddle/fluid/operators/crop_op.cc index 669b3bbe9d..5b5a220cf9 100644 --- a/paddle/fluid/operators/crop_op.cc +++ b/paddle/fluid/operators/crop_op.cc @@ -48,6 +48,13 @@ class CropOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", y_dim); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } }; class CropOpMaker : public framework::OpProtoAndCheckerMaker { @@ -60,13 +67,19 @@ class CropOpMaker : public framework::OpProtoAndCheckerMaker { "The input used as reference for cropping, " "which is of the same dimensions as X.") .AsDispensable(); + AddInput("Offsets", + "The input used to describe offsets in runtime, which is a " + "1-D vector whose size equals to the rank of input 'X'. The " + "elements data type must be int.") + .AsDispensable(); AddOutput("Out", "The output of crop op, " "which is of the same dimensions as X."); AddAttr>("offsets", "A list describing offsets to be cropped. " "The size of offsets list should be the same as " - "the dimension size of input X."); + "the dimension size of input X.") + .SetDefault(std::vector()); AddAttr>("shape", "A list describing the shape of output. " "The size of shape list should be the same as " @@ -77,6 +90,17 @@ Crop Operator. Crop input into output, as specified by offsets and shape. +There are two ways to set the offsets: +1. In runtime: Using the input 'Offsets', which is a Vairbale and can be + output of other operators. This way is suitable for + dynamic offsets. +2. In network configuration: Using the attribute 'offsets', which will be + set in Python configure script. This way is + suitable for fixed offsets. +You CANNOT use these two ways at the same time. An exception will be raised +if input 'Offset' is configured and meanwhile the attribute 'offsets' is +not empty. + There are two ways to set shape: 1. reference input: crop input X into the same shape as reference input. The dimension of reference input should @@ -146,6 +170,15 @@ class CropOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(x_grad_name, x_dims); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType( + ctx.Input(framework::GradVarName("Out")) + ->type()), + ctx.device_context()); + } }; } // namespace operators diff --git a/paddle/fluid/operators/crop_op.h b/paddle/fluid/operators/crop_op.h index f05c2e2328..91cfbbda73 100644 --- a/paddle/fluid/operators/crop_op.h +++ b/paddle/fluid/operators/crop_op.h @@ -27,6 +27,37 @@ template ; using framework::Tensor; +static std::vector GetOffsets(const framework::ExecutionContext& ctx) { + std::vector res; + int rank = ctx.Input("X")->dims().size(); + if (ctx.HasInput("Offsets")) { + PADDLE_ENFORCE(ctx.Attr>("offsets").empty(), + "Input 'Offsets' and attribute 'offsets' should not be used " + "at the same time."); + const auto* offsets_tensor = ctx.Input("Offsets"); + PADDLE_ENFORCE_EQ(offsets_tensor->dims().size(), 1); + PADDLE_ENFORCE_EQ( + rank, offsets_tensor->dims()[0], + "Offsets size should be equal to dimension size of input tensor."); + const int* offsets_data; + framework::Tensor cpu_tmp_tensor; + if (platform::is_cpu_place(offsets_tensor->place())) { + offsets_data = offsets_tensor->data(); + } else { + framework::TensorCopySync(*offsets_tensor, platform::CPUPlace(), + &cpu_tmp_tensor); + offsets_data = cpu_tmp_tensor.data(); + } + res = std::vector(offsets_data, offsets_data + rank); + } else { + res = ctx.Attr>("offsets"); + PADDLE_ENFORCE_EQ( + rank, res.size(), + "Offsets size should be equal to dimension size of input tensor."); + } + return res; +} + template class CropKernel : public framework::OpKernel { public: @@ -37,10 +68,7 @@ class CropKernel : public framework::OpKernel { T* out_data = out->mutable_data(context.GetPlace()); auto x_stride = framework::stride(x->dims()); auto out_stride = framework::stride(out->dims()); - auto offsets = context.Attr>("offsets"); - PADDLE_ENFORCE_EQ( - x->dims().size(), static_cast(offsets.size()), - "Offsets size should be equal to dimension size of input tensor."); + auto offsets = GetOffsets(context); int64_t offset = 0; for (size_t i = 0; i < offsets.size(); ++i) { offset += (x_stride[i] * offsets[i]); @@ -56,7 +84,7 @@ void CropGradFunction(const framework::ExecutionContext& context) { if (d_x != nullptr) { auto* d_out = context.Input(framework::GradVarName("Out")); d_x->mutable_data(context.GetPlace()); - auto offsets = context.Attr>("offsets"); + auto offsets = GetOffsets(context); Eigen::array, D> paddings; for (size_t i = 0; i < D; ++i) { paddings[i].first = offsets[i]; diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index a3bec3da45..d5e095f9ca 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -124,7 +124,8 @@ class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker { "Tensor with shape [N x D]."); AddOutput("Y", "(Tensor, default Tensor), a 2-D tensor with shape " - "[N x 1]. The cross entropy loss."); + "[N x 1]. The cross entropy loss.") + .Reuse("X"); AddAttr("soft_label", "(bool, default false), a flag indicating whether to " "interpretate the given labels as soft labels.") diff --git a/paddle/fluid/operators/detail/CMakeLists.txt b/paddle/fluid/operators/detail/CMakeLists.txt index b9a66474c9..abc5aad043 100644 --- a/paddle/fluid/operators/detail/CMakeLists.txt +++ b/paddle/fluid/operators/detail/CMakeLists.txt @@ -1,11 +1,38 @@ -if(WITH_DISTRIBUTE) +if(NOT WITH_DISTRIBUTE) + return() +endif() + + +if(WITH_GRPC) grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc - grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) + request_handler_impl.cc rpc_client.cc rpc_server.cc grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor + selected_rows memory) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - set_source_files_properties(serde_test.cc grpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - cc_test(serde_test SRCS serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr + set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(serde_test SRCS grpc_serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc SERIAL) - cc_test(grpc_server_test SRCS grpc_server_test.cc DEPS sendrecvop_grpc + cc_test(grpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_table_op SERIAL) + return() endif() + + +set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") +set_source_files_properties(brpc_server.cc brpc_client.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) +brpc_library(sendrecvop_brpc SRCS brpc_client.cc brpc_server.cc rpc_server.cc rpc_client.cc request_handler_impl.cc + PROTO send_recv.proto + DEPS lod_tensor selected_rows memory) + +find_library(OPENSSL_CRYPTO_LIBRARY_STATIC NAMES libcrypto.so) +ADD_LIBRARY(crypto SHARED IMPORTED GLOBAL) +SET_PROPERTY(TARGET crypto PROPERTY IMPORTED_LOCATION ${OPENSSL_CRYPTO_LIBRARY_STATIC}) + + +find_library(OPENSSL_SSL_LIBRARY_STATIC NAMES libssl.so) +ADD_LIBRARY(ssl SHARED IMPORTED GLOBAL) +SET_PROPERTY(TARGET ssl PROPERTY IMPORTED_LOCATION ${OPENSSL_SSL_LIBRARY_STATIC}) + +cc_test(brpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_brpc + brpc protobuf leveldb gflags glog + protobuf executor proto_desc lookup_table_op snappystream snappy ssl crypto SERIAL) diff --git a/paddle/fluid/operators/detail/brpc_client.cc b/paddle/fluid/operators/detail/brpc_client.cc new file mode 100644 index 0000000000..9a4e410f1d --- /dev/null +++ b/paddle/fluid/operators/detail/brpc_client.cc @@ -0,0 +1,180 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/detail/brpc_client.h" +#include "paddle/fluid/framework/threadpool.h" + +namespace paddle { +namespace operators { +namespace detail { + +DEFINE_int32(brpc_channel_num, 24, + "Number of channels to send requests connected to one server"); +DEFINE_int32(timeout_ms, 30000, "RPC timeout in milliseconds"); +DEFINE_int32(max_retry, 3, "Max retries(not including the first RPC)"); + +BRPCClient::~BRPCClient() { Wait(); } + +void HandleSendResponse(brpc::Controller* cntl, + sendrecv::VoidMessage* response) { + // std::unique_ptr makes sure cntl/response will be deleted before returning. + std::unique_ptr cntl_guard(cntl); + std::unique_ptr response_guard(response); + + if (cntl->Failed()) { + LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText(); + return; + } + LOG(INFO) << "Received response from " << cntl->remote_side() + << " latency=" << cntl->latency_us() << "us"; +} + +bool BRPCClient::AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string var_name_val = var_name; + const framework::Scope* p_scope = &scope; + const auto ch_ptr = GetChannel(ep_val); + + framework::AsyncIO( + [var_name_val, p_ctx, ep_val, p_scope, time_out, ch_ptr, this] { + auto ch_ctx = ch_ptr->Pop(); + brpc::Controller* cntl = new brpc::Controller(); + sendrecv::VoidMessage* response = new sendrecv::VoidMessage(); + cntl->set_timeout_ms(time_out); + + google::protobuf::Closure* done = + brpc::NewCallback(&HandleSendResponse, cntl, response); + + sendrecv::VariableMessage request; + ch_ctx->stub->SendVariable(cntl, &request, response, done); + }); + req_count_++; + + return true; +} + +void HandleGetResponse(brpc::Controller* cntl, + sendrecv::VariableMessage* response) { + // std::unique_ptr makes sure cntl/response will be deleted before returning. + std::unique_ptr cntl_guard(cntl); + std::unique_ptr response_guard(response); + + if (cntl->Failed()) { + LOG(WARNING) << "Fail to send EchoRequest, " << cntl->ErrorText(); + return; + } + LOG(INFO) << "Received response from " << cntl->remote_side() + << " latency=" << cntl->latency_us() << "us"; + + // framework::Variable* outvar = nullptr; + // DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, &outvar); +} + +bool BRPCClient::AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string var_name_val = var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::AsyncIO( + [var_name_val, ep_val, p_scope, p_ctx, time_out, ch, this] {}); + + req_count_++; + + return true; +} + +bool BRPCClient::AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string in_var_name_val = in_var_name; + const std::string out_var_name_val = out_var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, + time_out, ch, this] {}); + + req_count_++; + return true; +} + +void BRPCClient::AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out) { + req_count_++; +} + +void BRPCClient::AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out) { + req_count_++; +} + +void BRPCClient::Wait() { + std::unique_lock lk(sync_mutex_); + sync_cond_.wait(lk, [this] { return req_count_ == 0; }); +} + +ChannelQueuePtr BRPCClient::GetChannel(const std::string& ep) { + { + std::lock_guard guard(chan_mutex_); + auto it = channels_.find(ep); + if (it != channels_.end()) { + return it->second; + } + } + + ChannelQueuePtr q(new framework::BlockingQueue()); + + brpc::ChannelOptions options; + options.protocol = "baidu_std"; + options.connection_type = "pooled"; + options.connect_timeout_ms = 100; + options.timeout_ms = FLAGS_timeout_ms /*milliseconds*/; + options.max_retry = FLAGS_max_retry; + for (int i = 0; i < FLAGS_brpc_channel_num; ++i) { + std::shared_ptr c(new ChannelContext()); + if (c->channel.Init(ep.c_str(), &options) != 0) { + LOG(ERROR) << "Fail to initialize channel"; + return nullptr; + } + + c->stub.reset(new sendrecv::SendRecvService_Stub( + static_cast(&c->channel))); + q->Push(c); + } + + { + std::lock_guard guard(chan_mutex_); + channels_[ep] = q; + } + + return q; +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/brpc_client.h b/paddle/fluid/operators/detail/brpc_client.h new file mode 100644 index 0000000000..1e953ea431 --- /dev/null +++ b/paddle/fluid/operators/detail/brpc_client.h @@ -0,0 +1,100 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include // NOLINT +#include +#include +#include +#include +#include // NOLINT +#include +#include + +#include "brpc/channel.h" +#include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/detail/rpc_client.h" +#include "paddle/fluid/operators/detail/send_recv.pb.h" +#include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN + +namespace paddle { +namespace operators { +namespace detail { + +struct ChannelContext { + brpc::Channel channel; + std::shared_ptr stub; +}; + +typedef std::shared_ptr ChannelContextPtr; +typedef std::shared_ptr> + ChannelQueuePtr; + +class BRPCClient : public RPCClient { + public: + BRPCClient() {} + virtual ~BRPCClient(); + + bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = RPCClient::rpc_time_out) override; + + bool AsyncGetVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = RPCClient::rpc_time_out) override; + + bool AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out = RPCClient::rpc_time_out) override; + + void AsyncSendBatchBarrier( + const std::string& ep, + int64_t time_out = RPCClient::rpc_time_out) override; + + void AsyncSendFetchBarrier( + const std::string& ep, + int64_t time_out = RPCClient::rpc_time_out) override; + + void Wait() override; + + private: + void Proceed(); + ChannelQueuePtr GetChannel(const std::string& ep); + + private: + std::unordered_map channels_; + + // mutex for Wait client sync + std::mutex sync_mutex_; + std::condition_variable sync_cond_; + std::atomic req_count_{0}; + + // mutex for GetChannel thread safety + std::mutex chan_mutex_; + DISABLE_COPY_AND_ASSIGN(BRPCClient); +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/brpc_server.cc b/paddle/fluid/operators/detail/brpc_server.cc new file mode 100644 index 0000000000..2170abe679 --- /dev/null +++ b/paddle/fluid/operators/detail/brpc_server.cc @@ -0,0 +1,144 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/detail/brpc_server.h" +#include "paddle/fluid/operators/detail/request_handler.h" + +namespace sendrecv { + +typedef std::unordered_map + HandlerMap; + +class BRPCServiceImpl : public SendRecvService { + public: + explicit BRPCServiceImpl(const HandlerMap& rpc_call_map) + : request_send_h_(nullptr), + request_get_h_(nullptr), + request_prefetch_h_(nullptr) { + auto it = rpc_call_map.find(paddle::operators::detail::kRequestSend); + if (it != rpc_call_map.end()) { + request_send_h_ = it->second; + } + + it = rpc_call_map.find(paddle::operators::detail::kRequestSend); + if (it != rpc_call_map.end()) { + request_get_h_ = it->second; + } + + it = rpc_call_map.find(paddle::operators::detail::kRequestPrefetch); + if (it != rpc_call_map.end()) { + request_prefetch_h_ = it->second; + } + } + + virtual ~BRPCServiceImpl() {} + + void SendVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VoidMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE(request_send_h_ != nullptr, + "RequestSend handler should be registed first!"); + brpc::ClosureGuard done_guard(done); + + paddle::framework::Scope* local_scope = request_send_h_->scope(); + paddle::framework::Variable* outvar = nullptr; + paddle::framework::Variable* invar = nullptr; + + std::string varname = request->varname(); + + if (!request_send_h_->sync_mode()) { + local_scope = &request_send_h_->scope()->NewScope(); + invar = local_scope->Var(varname); + } else { + invar = local_scope->FindVar(varname); + } + + request_send_h_->Handle(varname, local_scope, invar, &outvar); + + if (!request_send_h_->sync_mode()) { + request_send_h_->scope()->DeleteScope(local_scope); + } + } + + void GetVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, VariableMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE(request_get_h_ != nullptr, + "RequestGet handler should be registed first!"); + } + + void PrefetchVariable(google::protobuf::RpcController* cntl_butil, + const VariableMessage* request, + VariableMessage* response, + google::protobuf::Closure* done) override { + PADDLE_ENFORCE(request_prefetch_h_ != nullptr, + "kRequestPrefetch handler should be registed first!"); + } + + private: + paddle::operators::detail::RequestHandler* request_send_h_; + paddle::operators::detail::RequestHandler* request_get_h_; + paddle::operators::detail::RequestHandler* request_prefetch_h_; +}; +} // namespace sendrecv + +namespace paddle { +namespace operators { +namespace detail { + +void AsyncBRPCServer::StartServer() { + // Instance of your service. + sendrecv::BRPCServiceImpl service_impl(rpc_call_map_); + + // Add the service into server. Notice the second parameter, because the + // service is put on stack, we don't want server to delete it, otherwise + // use brpc::SERVER_OWNS_SERVICE. + if (server_.AddService(&service_impl, brpc::SERVER_DOESNT_OWN_SERVICE) != 0) { + LOG(FATAL) << "Fail to add service"; + return; + } + + brpc::ServerOptions options; + options.idle_timeout_sec = idle_timeout_s_; + options.max_concurrency = max_concurrency_; + if (server_.Start(bind_address_.c_str(), &options) != 0) { + LOG(FATAL) << "Fail to start EchoServer" << bind_address_; + return; + } + + butil::EndPoint ep = server_.listen_address(); + selected_port_ = ep.port; + + { + std::lock_guard lock(this->mutex_ready_); + ready_ = 1; + } + condition_ready_.notify_all(); + + server_.Join(); +} + +void AsyncBRPCServer::ShutDownImpl() { server_.Stop(1000); } + +void AsyncBRPCServer::WaitServerReady() { + VLOG(3) << "AsyncGRPCServer is wait server ready"; + std::unique_lock lock(this->mutex_ready_); + condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); + VLOG(3) << "AsyncGRPCServer WaitSeverReady"; +} + +}; // namespace detail +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/detail/brpc_server.h b/paddle/fluid/operators/detail/brpc_server.h new file mode 100644 index 0000000000..0105c8074a --- /dev/null +++ b/paddle/fluid/operators/detail/brpc_server.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include // NOLINT +#include // NOLINT +#include + +#include "brpc/server.h" +#include "paddle/fluid/operators/detail/rpc_server.h" +#include "paddle/fluid/operators/detail/send_recv.pb.h" + +namespace paddle { +namespace operators { +namespace detail { + +class AsyncBRPCServer final : public RPCServer { + public: + explicit AsyncBRPCServer(const std::string& address, int client_num) + : RPCServer(address, client_num), ready_(0) {} + + virtual ~AsyncBRPCServer() {} + void StartServer() override; + void WaitServerReady() override; + + private: + void ShutDownImpl() override; + + brpc::Server server_; + + static constexpr int idle_timeout_s_ = -1; + static constexpr int max_concurrency_ = 0; + + std::mutex mutex_ready_; + std::condition_variable condition_ready_; + int ready_; +}; + +}; // namespace detail +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc index f7ce778687..02ffe3651e 100644 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ b/paddle/fluid/operators/detail/grpc_client.cc @@ -19,32 +19,43 @@ limitations under the License. */ #include #include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/detail/request_handler.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { namespace detail { -std::once_flag RPCClient::init_flag_; +void GRPCClient::InitImpl() { InitEventLoop(); } -std::unique_ptr RPCClient::rpc_client_(nullptr); +void GRPCClient::InitEventLoop() { + // start the client process thread + // TODO(wuyi): can make this in a threadpool + client_thread_.reset(new std::thread(std::bind(&GRPCClient::Proceed, this))); +} -RPCClient* RPCClient::GetInstance() { - std::call_once(init_flag_, &RPCClient::Init); - return rpc_client_.get(); +void GRPCClient::SendComplete() { + for (auto& it : channels_) { + this->AsyncSendComplete(it.first); + } } -void RPCClient::Init() { - if (rpc_client_.get() == nullptr) { - rpc_client_.reset(new RPCClient()); +GRPCClient::~GRPCClient() { + Wait(); + cq_.Shutdown(); + { + std::lock_guard guard(chan_mutex_); + for (auto& it : channels_) { + it.second.reset(); + } } + client_thread_->join(); } -bool RPCClient::AsyncSendVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, - int64_t time_out) { +bool GRPCClient::AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string var_name_val = var_name; @@ -94,11 +105,10 @@ void RequestToByteBuffer(const T& proto, ::grpc::ByteBuffer* result) { result->Swap(&tmp); } -bool RPCClient::AsyncGetVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, - int64_t time_out) { +bool GRPCClient::AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string var_name_val = var_name; @@ -136,12 +146,12 @@ bool RPCClient::AsyncGetVariable(const std::string& ep, return true; } -bool RPCClient::AsyncPrefetchVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& in_var_name, - const std::string& out_var_name, - int64_t time_out) { +bool GRPCClient::AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out) { const platform::DeviceContext* p_ctx = &ctx; const std::string ep_val = ep; const std::string in_var_name_val = in_var_name; @@ -179,7 +189,8 @@ bool RPCClient::AsyncPrefetchVariable(const std::string& ep, return true; } -void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { +void GRPCClient::AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out) { const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); @@ -192,7 +203,8 @@ void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { req_count_++; } -void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { +void GRPCClient::AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out) { const auto ch = GetChannel(ep); FetchBarrierProcessor* s = new FetchBarrierProcessor(ch); s->Prepare(time_out); @@ -204,68 +216,50 @@ void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { req_count_++; } -bool RPCClient::Wait() { - if (req_count_ <= 0) { - return true; - } - const size_t kReqCnt = req_count_; - bool a[kReqCnt]; - std::vector> waits(req_count_); - std::mutex mu; - - for (int i = 0; i < req_count_; i++) { - waits[i] = framework::AsyncIO([i, &a, &mu, this] { - bool ret = Proceed(); - std::lock_guard l(mu); - a[i] = ret; - }); - } - - for (int i = 0; i < req_count_; i++) { - waits[i].wait(); - } +void GRPCClient::AsyncSendComplete(const std::string& ep, int64_t time_out) { + const auto ch = GetChannel(ep); - int last_req_count = req_count_; - req_count_ = 0; + BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); + s->Prepare(time_out); - for (int i = 0; i < last_req_count; i++) { - if (!a[i]) { - return false; - } - } + sendrecv::VariableMessage req; + req.set_varname(COMPLETE_MESSAGE); + auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); + rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + req_count_++; +} - return true; +void GRPCClient::Wait() { + std::unique_lock lk(sync_mutex_); + sync_cond_.wait(lk, [this] { return req_count_ == 0; }); } -bool RPCClient::Proceed() { - void* tag = NULL; +void GRPCClient::Proceed() { + void* tag = nullptr; bool ok = false; - // request counts. - if (!cq_.Next(&tag, &ok)) { - LOG(ERROR) << "Get meets CompletionQueue error"; - return false; - } - - GPR_ASSERT(ok); - PADDLE_ENFORCE(tag); - - // TODO(gongwb): add more retries. - BaseProcessor* c = static_cast(tag); - if (!c->status_.ok()) { - LOG(ERROR) << "proc param error:" << c->var_h_.String() - << " grpc error:" << c->status_.error_message(); + while (cq_.Next(&tag, &ok)) { + BaseProcessor* c = static_cast(tag); + GPR_ASSERT(ok); + PADDLE_ENFORCE(c); + if (c->status_.ok()) { + c->Process(); + } else { + LOG(ERROR) << "var: " << c->var_h_.String() + << " grpc error:" << c->status_.error_message(); + } delete c; - return false; + { + std::lock_guard lk(sync_mutex_); + req_count_--; + } + sync_cond_.notify_all(); } - - c->Process(); - delete c; - return true; } -std::shared_ptr RPCClient::GetChannel(const std::string& ep) { + +std::shared_ptr GRPCClient::GetChannel(const std::string& ep) { // TODO(Yancey1989): make grpc client completely thread-safe - std::unique_lock lock(mutex_); + std::lock_guard guard(chan_mutex_); auto it = channels_.find(ep); if (it != channels_.end()) { return it->second; diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h index 449d5105af..44000c028b 100644 --- a/paddle/fluid/operators/detail/grpc_client.h +++ b/paddle/fluid/operators/detail/grpc_client.h @@ -16,15 +16,18 @@ limitations under the License. */ #include -#include // NOLINT +#include // NOLINT +#include // NOLINT #include #include #include #include #include // NOLINT #include +#include // NOLINT #include +#include "grpc++/channel.h" #include "grpc++/generic/generic_stub.h" #include "grpc++/grpc++.h" #include "grpc++/support/byte_buffer.h" @@ -35,6 +38,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/detail/rpc_client.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" #include "paddle/fluid/platform/macros.h" // for DISABLE_COPY_AND_ASSIGN @@ -161,53 +165,65 @@ class FetchBarrierProcessor : public BaseProcessor { std::unique_ptr stub_; }; -class RPCClient { +class GRPCClient : public RPCClient { public: - RPCClient() {} + GRPCClient() {} + virtual ~GRPCClient(); - static RPCClient* GetInstance(); + bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = RPCClient::rpc_time_out) override; - bool AsyncSendVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& var_name, - int64_t time_out = 600 * 1000); + bool AsyncGetVar(const std::string& ep, const platform::DeviceContext& ctx, + const framework::Scope& scope, const std::string& var_name, + int64_t time_out = RPCClient::rpc_time_out) override; - bool AsyncGetVariable(const std::string& ep, + bool AsyncPrefetchVar(const std::string& ep, const platform::DeviceContext& ctx, const framework::Scope& scope, - const std::string& var_name, - int64_t time_out = 600 * 1000); + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out = RPCClient::rpc_time_out) override; - bool AsyncPrefetchVariable(const std::string& ep, - const platform::DeviceContext& ctx, - const framework::Scope& scope, - const std::string& in_var_name, - const std::string& out_var_name, - int64_t time_out = 600 * 1000); + void AsyncSendBatchBarrier( + const std::string& ep, + int64_t time_out = RPCClient::rpc_time_out) override; - void AsyncSendBatchBarrier(const std::string& ep, - int64_t time_out = 600 * 1000); + void AsyncSendFetchBarrier( + const std::string& ep, + int64_t time_out = RPCClient::rpc_time_out) override; - void AsyncSendFetchBarrier(const std::string& ep, - int64_t time_out = 600 * 1000); + void Wait() override; - bool Wait(); + void SendComplete() override; + + protected: + void InitImpl() override; private: - bool Proceed(); + // InitEventLoop should only be called by Init() + void InitEventLoop(); + + void Proceed(); + + void AsyncSendComplete(const std::string& ep, + int64_t time_out = RPCClient::rpc_time_out); + std::shared_ptr GetChannel(const std::string& ep); - // Init is called by GetInstance. - static void Init(); private: grpc::CompletionQueue cq_; - std::map> channels_; + std::unordered_map> channels_; + std::unique_ptr client_thread_; + + // mutex for Wait client sync + std::mutex sync_mutex_; + std::condition_variable sync_cond_; std::atomic req_count_{0}; - std::mutex mutex_; - static std::unique_ptr rpc_client_; - static std::once_flag init_flag_; - DISABLE_COPY_AND_ASSIGN(RPCClient); + + // mutex for GetChannel thread safety + std::mutex chan_mutex_; + DISABLE_COPY_AND_ASSIGN(GRPCClient); }; } // namespace detail diff --git a/paddle/fluid/operators/detail/serde_test.cc b/paddle/fluid/operators/detail/grpc_serde_test.cc similarity index 100% rename from paddle/fluid/operators/detail/serde_test.cc rename to paddle/fluid/operators/detail/grpc_serde_test.cc diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index 361cc24b5b..5a87258901 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/*Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. @@ -12,19 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/detail/grpc_server.h" - #include #include -using ::grpc::ServerAsyncResponseWriter; +#include "paddle/fluid/operators/detail/grpc_server.h" -DEFINE_int32(rpc_server_handle_send_threads, 20, - "Number of threads used to handle send at rpc server."); -DEFINE_int32(rpc_server_handle_get_threads, 20, - "Number of threads used to handle get at rpc server."); -DEFINE_int32(rpc_server_handle_prefetch_threads, 1, - "Number of threads used to handle prefetch at rpc server."); +using ::grpc::ServerAsyncResponseWriter; namespace paddle { namespace operators { @@ -36,377 +29,323 @@ enum CallStatus { PROCESS = 0, FINISH }; class RequestBase { public: explicit RequestBase(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - const platform::DeviceContext* dev_ctx) + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) : service_(service), cq_(cq), - sync_mode_(sync_mode), status_(PROCESS), - dev_ctx_(dev_ctx) { + request_handler_(request_handler), + req_id_(req_id) { PADDLE_ENFORCE(cq_); } virtual ~RequestBase() {} - virtual void Process() { assert(false); } + virtual void Process() = 0; + + CallStatus Status() const { + std::lock_guard l(status_mu_); + return status_; + } - CallStatus Status() { return status_; } - void SetStatus(CallStatus status) { status_ = status; } - virtual std::string GetReqName() { - assert(false); - return ""; + template + void Finish(const T& reply, ServerAsyncResponseWriter* responder) { + std::lock_guard l(status_mu_); + status_ = FINISH; + responder->Finish(reply, ::grpc::Status::OK, + reinterpret_cast(static_cast(req_id_))); } + virtual std::string GetReqName() = 0; protected: + mutable std::mutex status_mu_; ::grpc::ServerContext ctx_; GrpcService::AsyncService* service_; ::grpc::ServerCompletionQueue* cq_; - const bool sync_mode_; CallStatus status_; - const platform::DeviceContext* dev_ctx_; + RequestHandler* request_handler_; + int req_id_; }; class RequestSend final : public RequestBase { public: explicit RequestSend(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - framework::Scope* scope, ReceivedQueue* queue, - const platform::DeviceContext* dev_ctx, int req_id) - : RequestBase(service, cq, sync_mode, dev_ctx), - queue_(queue), - responder_(&ctx_), - req_id_(req_id) { - if (sync_mode_) { - request_.reset(new VariableResponse(scope, dev_ctx_, false)); - } else { - request_.reset(new VariableResponse(scope, dev_ctx_, true)); - } + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { + request_.reset(new VariableResponse(request_handler->scope(), + request_handler->dev_ctx(), + !request_handler->sync_mode())); int method_id = static_cast(detail::GrpcMethod::kSendVariable); service_->RequestAsyncUnary( method_id, &ctx_, request_.get(), &responder_, cq_, cq_, reinterpret_cast(static_cast(req_id))); } - virtual ~RequestSend() {} + std::string GetReqName() override { return request_->Varname(); } - virtual std::string GetReqName() { return request_->Varname(); } + void Process() override { + std::string varname = GetReqName(); + VLOG(3) << "RequestSend var_name:" << varname; - virtual void Process() { - std::string var_name = GetReqName(); - VLOG(3) << "RequestSend " << var_name; - queue_->Push(std::make_pair(var_name, request_)); + auto scope = request_->GetMutableLocalScope(); + auto invar = request_->GetVar(); + framework::Variable* outvar = nullptr; - status_ = FINISH; - responder_.Finish(reply_, ::grpc::Status::OK, - reinterpret_cast(static_cast(req_id_))); + request_handler_->Handle(varname, scope, invar, &outvar); + Finish(reply_, &responder_); } protected: sendrecv::VoidMessage reply_; std::shared_ptr request_; - ReceivedQueue* queue_; ServerAsyncResponseWriter responder_; - int req_id_; }; class RequestGet final : public RequestBase { public: explicit RequestGet(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - framework::Scope* scope, - const platform::DeviceContext* dev_ctx, - framework::BlockingQueue* queue, - int req_id) - : RequestBase(service, cq, sync_mode, dev_ctx), - responder_(&ctx_), - scope_(scope), - queue_(queue), - req_id_(req_id) { + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { auto method_id = static_cast(detail::GrpcMethod::kGetVariable); service_->RequestAsyncUnary( method_id, &ctx_, &request_, &responder_, cq_, cq_, - reinterpret_cast(static_cast(req_id_))); + reinterpret_cast(static_cast(req_id))); } virtual ~RequestGet() {} - virtual std::string GetReqName() { return request_.varname(); } + std::string GetReqName() override { return request_.varname(); } - virtual void Process() { + void Process() override { // proc request. - std::string var_name = request_.varname(); - VLOG(3) << "RequestGet " << var_name; - auto* var = scope_->FindVar(var_name); + std::string varname = request_.varname(); + VLOG(3) << "RequestGet " << varname; - if (var_name != FETCH_BARRIER_MESSAGE) { - SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply_); - } + auto scope = request_handler_->scope(); + auto invar = scope->FindVar(varname); + framework::Variable* outvar = nullptr; - status_ = FINISH; - responder_.Finish(reply_, ::grpc::Status::OK, - reinterpret_cast(static_cast(req_id_))); + request_handler_->Handle(varname, scope, invar, &outvar); - if (var_name == FETCH_BARRIER_MESSAGE) { - sendrecv::VariableMessage msg; - MessageWithName msg_with_name = std::make_pair(var_name, msg); - queue_->Push(msg_with_name); + if (outvar) { + SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), + &reply_); } + Finish(reply_, &responder_); } protected: sendrecv::VariableMessage request_; ::grpc::ByteBuffer reply_; ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; - framework::Scope* scope_; - framework::BlockingQueue* queue_; - int req_id_; }; class RequestPrefetch final : public RequestBase { public: explicit RequestPrefetch(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - framework::Scope* scope, - const platform::DeviceContext* dev_ctx, - framework::Executor* executor, - framework::ProgramDesc* program, - framework::ExecutorPrepareContext* prefetch_ctx, - int req_id) - : RequestBase(service, cq, sync_mode, dev_ctx), + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_), - scope_(scope), - executor_(executor), - program_(program), - prefetch_ctx_(prefetch_ctx), - req_id_(req_id) { - // prefetch always create a new sub scope - request_.reset(new VariableResponse(scope, dev_ctx_, true)); + local_scope_(nullptr) { + request_.reset(new VariableResponse(request_handler->scope(), + request_handler->dev_ctx(), true)); int method_id = static_cast(detail::GrpcMethod::kPrefetchVariable); service_->RequestAsyncUnary( method_id, &ctx_, request_.get(), &responder_, cq_, cq_, - reinterpret_cast(static_cast(req_id_))); + reinterpret_cast(static_cast(req_id))); } virtual ~RequestPrefetch() {} - virtual std::string GetReqName() { return request_->Varname(); } + std::string GetReqName() override { return request_->Varname(); } - virtual void Process() { + void Process() override { // prefetch process... + std::string in_var_name = request_->Varname(); + std::string out_var_name = request_->OutVarname(); + VLOG(3) << "RequestPrefetch, in_var_name: " << in_var_name + << " out_var_name: " << out_var_name; - std::string var_name = request_->OutVarname(); - VLOG(3) << "RequestPrefetch " << var_name; - auto var_desc = program_->Block(0).FindVar(var_name); - framework::Scope* local_scope = request_->GetMutableLocalScope(); - auto* var = local_scope->FindVar(var_name); - InitializeVariable(var, var_desc->GetType()); - executor_->RunPreparedContext(prefetch_ctx_, local_scope); + auto scope = request_->GetMutableLocalScope(); + auto invar = scope->FindVar(in_var_name); + // out var must be created in local scope! + framework::Variable* outvar = scope->Var(out_var_name); - SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply_); + request_handler_->Handle(in_var_name, scope, invar, &outvar, out_var_name); - status_ = FINISH; - responder_.Finish(reply_, ::grpc::Status::OK, - reinterpret_cast(static_cast(req_id_))); + SerializeToByteBuffer(out_var_name, outvar, *request_handler_->dev_ctx(), + &reply_); + Finish(reply_, &responder_); } protected: std::shared_ptr request_; ::grpc::ByteBuffer reply_; ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; - framework::Scope* scope_; - framework::Executor* executor_; - framework::ProgramDesc* program_; - framework::ExecutorPrepareContext* prefetch_ctx_; - int req_id_; + framework::Scope* local_scope_; }; -void AsyncGRPCServer::WaitClientGet(int count) { - int fetch_barriers = 0; - while (fetch_barriers < count) { - auto msg = var_get_queue_.Pop(); - if (msg.first == FETCH_BARRIER_MESSAGE) { - fetch_barriers++; - } - } -} - void AsyncGRPCServer::WaitServerReady() { + VLOG(3) << "AsyncGRPCServer is wait server ready"; std::unique_lock lock(this->mutex_ready_); condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); + VLOG(3) << "AsyncGRPCServer WaitSeverReady"; } -void AsyncGRPCServer::RunSyncUpdate() { +void AsyncGRPCServer::StartServer() { ::grpc::ServerBuilder builder; - builder.AddListeningPort(address_, ::grpc::InsecureServerCredentials(), + builder.AddListeningPort(bind_address_, ::grpc::InsecureServerCredentials(), &selected_port_); + builder.SetMaxSendMessageSize(std::numeric_limits::max()); builder.SetMaxReceiveMessageSize(std::numeric_limits::max()); builder.RegisterService(&service_); - cq_send_ = builder.AddCompletionQueue(); - cq_get_ = builder.AddCompletionQueue(); - cq_prefetch_ = builder.AddCompletionQueue(); + for (auto t : rpc_call_map_) { + rpc_cq_[t.first].reset(builder.AddCompletionQueue().release()); + } server_ = builder.BuildAndStart(); - LOG(INFO) << "Server listening on " << address_ + LOG(INFO) << "Server listening on " << bind_address_ << " selected port: " << selected_port_; - std::function send_register = std::bind( - &AsyncGRPCServer::TryToRegisterNewSendOne, this, std::placeholders::_1); - std::function get_register = std::bind( - &AsyncGRPCServer::TryToRegisterNewGetOne, this, std::placeholders::_1); - std::function prefetch_register = - std::bind(&AsyncGRPCServer::TryToRegisterNewPrefetchOne, this, - std::placeholders::_1); + std::function f = + std::bind(&AsyncGRPCServer::TryToRegisterNewOne, this, + std::placeholders::_1, std::placeholders::_2); - for (int i = 0; i < kSendReqsBufSize; ++i) { - TryToRegisterNewSendOne(i); - } - for (int i = 0; i < kGetReqsBufSize; ++i) { - TryToRegisterNewGetOne(i); - } - for (int i = 0; i < kPrefetchReqsBufSize; ++i) { - TryToRegisterNewPrefetchOne(i); - } + for (auto& t : rpc_call_map_) { + auto& rpc_name = t.first; + auto& cq = rpc_cq_[rpc_name]; + auto threadnum = rpc_thread_num_[rpc_name]; + auto& reqs = rpc_reqs_[rpc_name]; - for (int i = 0; i < FLAGS_rpc_server_handle_send_threads; ++i) { - t_sends_.emplace_back( - new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, - cq_send_.get(), "cq_send", send_register))); - } - for (int i = 0; i < FLAGS_rpc_server_handle_get_threads; ++i) { - t_gets_.emplace_back( - new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, - cq_get_.get(), "cq_get", get_register))); - } - for (int i = 0; i < FLAGS_rpc_server_handle_prefetch_threads; ++i) { - t_prefetchs_.emplace_back(new std::thread( - std::bind(&AsyncGRPCServer::HandleRequest, this, cq_prefetch_.get(), - "cq_prefetch", prefetch_register))); + reqs.reserve(kRequestBufSize); + + for (int i = 0; i < kRequestBufSize; i++) { + TryToRegisterNewOne(rpc_name, i); + } + + for (int i = 0; i < threadnum; i++) { + rpc_threads_[rpc_name].emplace_back(new std::thread(std::bind( + &AsyncGRPCServer::HandleRequest, this, cq.get(), rpc_name, f))); + VLOG(3) << t.first << " creates threads!"; + } } + { std::lock_guard lock(this->mutex_ready_); ready_ = 1; } condition_ready_.notify_all(); + // wait server server_->Wait(); - for (int i = 0; i < FLAGS_rpc_server_handle_send_threads; ++i) { - t_sends_[i]->join(); - } - for (int i = 0; i < FLAGS_rpc_server_handle_get_threads; ++i) { - t_gets_[i]->join(); - } - for (int i = 0; i < FLAGS_rpc_server_handle_prefetch_threads; ++i) { - t_prefetchs_[i]->join(); + + for (auto& t : rpc_threads_) { + auto& threads = t.second; + for (size_t i = 0; i < threads.size(); ++i) { + threads[i]->join(); + VLOG(3) << t.first << " threads ends!"; + } } } void AsyncGRPCServer::ShutdownQueue() { - std::unique_lock lock(cq_mutex_); - cq_send_->Shutdown(); - cq_get_->Shutdown(); - cq_prefetch_->Shutdown(); + for (auto& t : rpc_cq_) { + t.second->Shutdown(); + VLOG(3) << t.first << " shutdown!"; + } } -// This URL explains why shutdown is complicate: -void AsyncGRPCServer::ShutDown() { +void AsyncGRPCServer::ShutDownImpl() { + std::unique_lock lock(cq_mutex_); is_shut_down_ = true; ShutdownQueue(); + + VLOG(3) << "server_ shutdown!"; server_->Shutdown(); } -void AsyncGRPCServer::TryToRegisterNewSendOne(int i) { +void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, + int req_id) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { VLOG(3) << "shutdown, do not TryToRegisterNewSendOne"; return; } - RequestSend* send = new RequestSend(&service_, cq_send_.get(), sync_mode_, - scope_, &var_recv_queue_, dev_ctx_, i); - send_reqs_[i] = static_cast(send); - VLOG(4) << "Create RequestSend status:" << send->Status(); -} -void AsyncGRPCServer::TryToRegisterNewGetOne(int req_id) { - std::unique_lock lock(cq_mutex_); - if (is_shut_down_) { - VLOG(3) << "shutdown, do not TryToRegisterNewGetOne"; - return; + VLOG(4) << "register send rpc_name:" << rpc_name + << ", handler:" << rpc_call_map_[kRequestSend]; + + auto& reqs = rpc_reqs_[rpc_name]; + auto& handler = rpc_call_map_[rpc_name]; + auto& cq = rpc_cq_[rpc_name]; + + RequestBase* b = nullptr; + if (rpc_name == kRequestSend) { + b = new RequestSend(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestGet) { + b = new RequestGet(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestPrefetch) { + b = new RequestPrefetch(&service_, cq.get(), handler, req_id); + } else { + PADDLE_ENFORCE(false, "not supported rpc"); } - RequestGet* get = new RequestGet(&service_, cq_get_.get(), sync_mode_, scope_, - dev_ctx_, &var_get_queue_, req_id); - get_reqs_[req_id] = static_cast(get); - VLOG(4) << "Create RequestGet status:" << get->Status(); -} -void AsyncGRPCServer::TryToRegisterNewPrefetchOne(int req_id) { - std::unique_lock lock(cq_mutex_); - if (is_shut_down_) { - VLOG(3) << "shutdown, do not TryToRegisterNewPrefetchOne"; - return; - } - RequestPrefetch* prefetch = new RequestPrefetch( - &service_, cq_prefetch_.get(), sync_mode_, scope_, dev_ctx_, executor_, - program_, prefetch_ctx_.get(), req_id); - prefetch_reqs_[req_id] = static_cast(prefetch); + reqs[req_id] = b; - VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status(); + VLOG(4) << "Create RequestSend status:" << b->Status(); } -// FIXME(typhoonzero): change cq_name to enum. void AsyncGRPCServer::HandleRequest( - ::grpc::ServerCompletionQueue* cq, const std::string& cq_name, - std::function TryToRegisterNewOne) { + ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, + std::function TryToRegisterNewOne) { void* tag = NULL; bool ok = false; while (true) { - VLOG(3) << "HandleRequest for " << cq_name << " wait Next"; + VLOG(3) << "HandleRequest " << rpc_name << " wait next"; if (!cq->Next(&tag, &ok)) { - LOG(INFO) << cq_name << " CompletionQueue shutdown!"; + LOG(INFO) << "CompletionQueue " << rpc_name << " shutdown!"; break; } - VLOG(3) << "HandleRequest for " << cq_name << " get Next"; - int req_id = static_cast(reinterpret_cast(tag)); - if (sync_mode_) { - // FIXME(typhoonzero): de-couple the barriers with recv_op - if (!is_shut_down_ && cq_name == "cq_get") WaitCond(1); - if (!is_shut_down_ && cq_name == "cq_send") WaitCond(0); - VLOG(3) << "HandleRequest for " << cq_name << " after WaitCond"; - } + int req_id = static_cast(reinterpret_cast(tag)); + VLOG(3) << "HandleRequest " << rpc_name << ", req_id:" << req_id + << " get next"; + auto& reqs = rpc_reqs_[rpc_name]; RequestBase* base = nullptr; { - std::lock_guard l(cq_mutex_); - if (cq_name == "cq_get") { - base = get_reqs_[req_id]; - } else if (cq_name == "cq_send") { - base = send_reqs_[req_id]; - } else if (cq_name == "cq_prefetch") { - base = prefetch_reqs_[req_id]; - } + PADDLE_ENFORCE(req_id >= 0 && req_id < kRequestBufSize); + std::unique_lock lock(cq_mutex_); + base = reqs[req_id]; } + // reference: // https://github.com/tensorflow/tensorflow/issues/5596 // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I if (!ok) { - LOG(WARNING) << cq_name << " recv no regular event:argument name[" + LOG(WARNING) << "completion queue:" << rpc_name + << " recv no regular event:argument name[" << base->GetReqName() << "]"; - TryToRegisterNewOne(req_id); + TryToRegisterNewOne(rpc_name, req_id); delete base; continue; } + VLOG(3) << "queue id:" << rpc_name << ", req_id:" << req_id + << ", status:" << base->Status(); + switch (base->Status()) { case PROCESS: { base->Process(); - VLOG(4) << cq_name << " PROCESS status:" << base->Status(); break; } case FINISH: { - TryToRegisterNewOne(req_id); - VLOG(4) << cq_name << " FINISH status:" << base->Status(); + TryToRegisterNewOne(rpc_name, req_id); delete base; break; } @@ -415,20 +354,6 @@ void AsyncGRPCServer::HandleRequest( } } -void AsyncGRPCServer::WaitCond(int cond) { - std::unique_lock lock(this->barrier_mutex_); - barrier_condition_.wait(lock, - [=] { return this->barrier_cond_step_ == cond; }); -} - -void AsyncGRPCServer::SetCond(int cond) { - { - std::lock_guard lock(this->barrier_mutex_); - barrier_cond_step_ = cond; - } - barrier_condition_.notify_all(); -} - } // namespace detail } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index bdff9801a9..f1db7590f6 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include +#include #include #include // NOLINT #include @@ -28,6 +30,8 @@ limitations under the License. */ #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/operators/detail/grpc_service.h" +#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/detail/rpc_server.h" #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" #include "paddle/fluid/operators/detail/send_recv.pb.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" @@ -37,106 +41,47 @@ namespace paddle { namespace operators { namespace detail { -typedef std::pair> - ReceivedMessage; -typedef framework::BlockingQueue ReceivedQueue; - -typedef std::pair MessageWithName; class RequestBase; -class AsyncGRPCServer final { +class AsyncGRPCServer final : public RPCServer { public: - explicit AsyncGRPCServer(const std::string &address, bool sync_mode) - : address_(address), sync_mode_(sync_mode), ready_(0) {} - - ~AsyncGRPCServer() {} - void WaitServerReady(); - void RunSyncUpdate(); - - // functions to sync server barrier status. - void WaitCond(int cond); - void SetCond(int cond); - void WaitClientGet(int count); - - void SetScope(framework::Scope *scope) { scope_ = scope; } - - void SetDevCtx(const platform::DeviceContext *dev_ctx) { dev_ctx_ = dev_ctx; } - - void SetProgram(framework::ProgramDesc *program) { program_ = program; } - - void SetExecutor(framework::Executor *executor) { executor_ = executor; } - - void SetPrefetchPreparedCtx( - std::unique_ptr prepared) { - prefetch_ctx_.reset(prepared.release()); - } - - int GetSelectedPort() const { return selected_port_; } - - const ReceivedMessage Get() { return this->var_recv_queue_.Pop(); } + explicit AsyncGRPCServer(const std::string& address, int client_num) + : RPCServer(address, client_num), ready_(0) {} - void Push(const std::string &msg_name) { - this->var_recv_queue_.Push(std::make_pair(msg_name, nullptr)); - } + virtual ~AsyncGRPCServer() {} + void WaitServerReady() override; + void StartServer() override; - void ShutDown(); + private: + // HandleRequest needs to be thread-safe. + void HandleRequest( + ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, + std::function TryToRegisterNewOne); - protected: - void HandleRequest(::grpc::ServerCompletionQueue *cq, - const std::string &cq_name, - std::function TryToRegisterNewOne); - void TryToRegisterNewSendOne(int req_id); - void TryToRegisterNewGetOne(int req_id); - void TryToRegisterNewPrefetchOne(int req_id); + void TryToRegisterNewOne(const std::string& rpc_name, int req_id); void ShutdownQueue(); + void ShutDownImpl() override; private: - static const int kSendReqsBufSize = 100; - static const int kGetReqsBufSize = 100; - static const int kPrefetchReqsBufSize = 10; + static const int kRequestBufSize = 100; std::mutex cq_mutex_; volatile bool is_shut_down_ = false; - std::unique_ptr<::grpc::ServerCompletionQueue> cq_send_; - std::unique_ptr<::grpc::ServerCompletionQueue> cq_get_; - std::unique_ptr<::grpc::ServerCompletionQueue> cq_prefetch_; - - RequestBase *send_reqs_[kSendReqsBufSize]; - RequestBase *get_reqs_[kGetReqsBufSize]; - RequestBase *prefetch_reqs_[kPrefetchReqsBufSize]; GrpcService::AsyncService service_; std::unique_ptr<::grpc::Server> server_; - std::string address_; - const bool sync_mode_; - framework::Scope *scope_; - const platform::DeviceContext *dev_ctx_; - - // received variable from RPC, operators fetch variable from this queue. - framework::BlockingQueue var_get_queue_; - // client send variable to this queue. - ReceivedQueue var_recv_queue_; - // condition of the sub program - std::mutex barrier_mutex_; - mutable int barrier_cond_step_; std::condition_variable barrier_condition_; - std::vector> t_sends_; - std::vector> t_gets_; - std::vector> t_prefetchs_; - - std::unique_ptr t_prefetch_; - - std::unique_ptr prefetch_ctx_; - framework::ProgramDesc *program_; - framework::Executor *executor_; - int selected_port_; - std::mutex mutex_ready_; std::condition_variable condition_ready_; + int ready_; + + std::map> rpc_cq_; + std::map>> rpc_threads_; + std::map> rpc_reqs_; }; }; // namespace detail diff --git a/paddle/fluid/operators/detail/macros.h b/paddle/fluid/operators/detail/macros.h new file mode 100644 index 0000000000..da1de72dad --- /dev/null +++ b/paddle/fluid/operators/detail/macros.h @@ -0,0 +1,27 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#ifdef PADDLE_WITH_GRPC +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/grpc_server.h" +#define RPCSERVER_T detail::AsyncGRPCServer +#define RPCCLIENT_T detail::GRPCClient +#else +#include "paddle/fluid/operators/detail/brpc_client.h" +#include "paddle/fluid/operators/detail/brpc_server.h" +#define RPCSERVER_T detail::AsyncBRPCServer +#define RPCCLIENT_T detail::BRPCClient +#endif diff --git a/paddle/fluid/operators/detail/request_handler.h b/paddle/fluid/operators/detail/request_handler.h new file mode 100644 index 0000000000..a2d08747d5 --- /dev/null +++ b/paddle/fluid/operators/detail/request_handler.h @@ -0,0 +1,129 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" + +namespace paddle { +namespace operators { +namespace detail { + +constexpr char kRequestSend[] = "RequestSend"; +constexpr char kRequestGet[] = "RequestGet"; +constexpr char kRequestPrefetch[] = "RequestPrefetch"; + +#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" +#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV" +#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV" +#define COMPLETE_MESSAGE "COMPLETE@RECV" + +class RPCServer; + +class RequestHandler { + public: + explicit RequestHandler(bool sync_mode) + : sync_mode_(sync_mode), + dev_ctx_(nullptr), + executor_(nullptr), + scope_(nullptr), + program_(nullptr), + rpc_server_(nullptr) {} + + virtual ~RequestHandler() {} + + // Set attributes. + void SetScope(framework::Scope* scope) { scope_ = scope; } + void SetDevCtx(const platform::DeviceContext* dev_ctx) { dev_ctx_ = dev_ctx; } + void SetProgram(framework::ProgramDesc* program) { program_ = program; } + void SetExecutor(framework::Executor* executor) { executor_ = executor; } + + // Used for dist lookup table prefetch + void SetPrefetchPreparedCtx( + std::unordered_map< + std::string, std::shared_ptr>* g) { + prefetch_var_name_to_prepared_ctx_ = g; + } + + // Used for async. + void SetGradToPreparedCtx( + std::unordered_map< + std::string, std::shared_ptr>* g) { + grad_to_prepared_ctx_ = g; + } + + void SetRPCServer(RPCServer* rpc_server) { rpc_server_ = rpc_server; } + + // Get attributes. + bool sync_mode() { return sync_mode_; } + framework::Scope* scope() { return scope_; } + const platform::DeviceContext* dev_ctx() { return dev_ctx_; } + framework::ProgramDesc* program() { return program_; } + framework::Executor* executor() { return executor_; } + + // This function processes user's rpc request. + // The implemention is in request_handler_impl. + // example: + // std::string varname = request_.varname(); + // + // auto scope = request_handler_->scope(); + // auto invar = scope->FindVar(varname); + // framework::Variable* outvar = nullptr; + // + // request_handler_->Handle(varname, scope, invar, &outvar); + // if (outvar) { + // SerializeToByteBuffer(varname, outvar, + // *request_handler_->dev_ctx(), &reply_); + // } + virtual bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") = 0; + + protected: + const bool sync_mode_; + + const platform::DeviceContext* dev_ctx_; + framework::Executor* executor_; + framework::Scope* scope_; + framework::ProgramDesc* program_; + + // used for distribute lookup table prefetch + std::unordered_map>* + prefetch_var_name_to_prepared_ctx_; + + // Used for async. + std::unordered_map>* + grad_to_prepared_ctx_; + + RPCServer* rpc_server_; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.cc b/paddle/fluid/operators/detail/request_handler_impl.cc new file mode 100644 index 0000000000..7425bee798 --- /dev/null +++ b/paddle/fluid/operators/detail/request_handler_impl.cc @@ -0,0 +1,124 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "paddle/fluid/operators/detail/rpc_server.h" + +namespace paddle { +namespace operators { +namespace detail { + +bool RequestSendHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar, + const std::string& out_var_name) { + VLOG(4) << "RequestSendHandler:" << varname; + + // Async + if (!sync_mode_) { + try { + executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(), + scope); + } catch (std::exception& e) { + LOG(ERROR) << "async: run sub program error " << e.what(); + return false; + } + return true; + } + + // Sync + if (varname == BATCH_BARRIER_MESSAGE) { + VLOG(3) << "sync: recv batch barrier message"; + rpc_server_->IncreaseBatchBarrier(kRequestSend); + } else if (varname == COMPLETE_MESSAGE) { + VLOG(3) << "sync: recv complete message"; + rpc_server_->DecreaseClientNum(); + } else { + VLOG(3) << "sync: received var_name: " << varname; + if (sync_mode_) { + rpc_server_->WaitCond(kRequestSend); + } + + if (invar == nullptr) { + LOG(ERROR) << "sync: Can not find server side var: " << varname; + PADDLE_THROW("sync: Can not find server side var"); + return false; + } + if (invar->IsType()) { + std::unique_lock lock(mutex_sparse_vars_); + sparse_vars_.push_back(invar); + } + } + return true; +} + +void RequestSendHandler::ResetSparseVarRecorder() { + std::unique_lock lock(mutex_sparse_vars_); + for (auto* var : sparse_vars_) { + var->GetMutable()->mutable_rows()->clear(); + } + sparse_vars_.clear(); +} + +bool RequestGetHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar, + const std::string& out_var_name) { + VLOG(4) << "RequestGetHandler:" << varname; + + if (varname != FETCH_BARRIER_MESSAGE) { + if (sync_mode_) { + rpc_server_->WaitCond(kRequestGet); + } + *outvar = scope_->FindVar(varname); + return true; + } + + // FETCH_BARRIER_MESSAGE + if (sync_mode_) { + VLOG(3) << "sync: recv fetch barrier message"; + rpc_server_->IncreaseBatchBarrier(kRequestGet); + } + + return true; +} + +bool RequestPrefetchHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar, + const std::string& out_var_name) { + VLOG(4) << "RequestPrefetchHandler " << varname; + + auto var_desc = program_->Block(0).FindVar(out_var_name); + InitializeVariable(*outvar, var_desc->GetType()); + executor_->RunPreparedContext( + (*prefetch_var_name_to_prepared_ctx_)[varname].get(), scope); + + return true; +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.h b/paddle/fluid/operators/detail/request_handler_impl.h new file mode 100644 index 0000000000..3f77c09a95 --- /dev/null +++ b/paddle/fluid/operators/detail/request_handler_impl.h @@ -0,0 +1,71 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include +#include +#include +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/detail/request_handler.h" + +namespace paddle { +namespace operators { +namespace detail { + +class RequestSendHandler final : public RequestHandler { + public: + explicit RequestSendHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestSendHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") override; + void ResetSparseVarRecorder(); + + private: + std::mutex mutex_sparse_vars_; + std::vector sparse_vars_; +}; + +class RequestGetHandler final : public RequestHandler { + public: + explicit RequestGetHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestGetHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") override; +}; + +class RequestPrefetchHandler final : public RequestHandler { + public: + explicit RequestPrefetchHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestPrefetchHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar, + const std::string& out_var_name = "") override; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_client.cc b/paddle/fluid/operators/detail/rpc_client.cc new file mode 100644 index 0000000000..9a791403e3 --- /dev/null +++ b/paddle/fluid/operators/detail/rpc_client.cc @@ -0,0 +1,26 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/detail/rpc_client.h" + +namespace paddle { +namespace operators { +namespace detail { + +std::once_flag RPCClient::init_flag_; +std::unique_ptr RPCClient::rpc_client_(nullptr); + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_client.h b/paddle/fluid/operators/detail/rpc_client.h new file mode 100644 index 0000000000..47c6ffb4fd --- /dev/null +++ b/paddle/fluid/operators/detail/rpc_client.h @@ -0,0 +1,89 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" + +namespace paddle { +namespace operators { +namespace detail { + +class RPCClient { + public: + RPCClient() {} + virtual ~RPCClient() {} + virtual bool AsyncSendVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out = rpc_time_out) = 0; + + virtual bool AsyncGetVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& var_name, + int64_t time_out = rpc_time_out) = 0; + + virtual bool AsyncPrefetchVar(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out = rpc_time_out) = 0; + + virtual void AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out = rpc_time_out) = 0; + + virtual void AsyncSendFetchBarrier(const std::string& ep, + int64_t time_out = rpc_time_out) = 0; + + // SendComplete tells all the server that current trainer have no more data + // to train, so that the pserver can reduce it's barrier count, and continue + // to train with other trainers. + virtual void SendComplete() = 0; + + virtual void Wait() = 0; + + static constexpr int64_t rpc_time_out = 120 * 1000; + + template + static RPCClient* GetInstance() { + std::call_once(init_flag_, &RPCClient::Init); + return rpc_client_.get(); + } + + // Init is called by GetInstance. + template + static void Init() { + if (rpc_client_.get() == nullptr) { + rpc_client_.reset(new T()); + rpc_client_->InitImpl(); + } + } + + protected: + virtual void InitImpl() {} + + private: + static std::once_flag init_flag_; + static std::unique_ptr rpc_client_; +}; +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.cc b/paddle/fluid/operators/detail/rpc_server.cc new file mode 100644 index 0000000000..cd0fe96e23 --- /dev/null +++ b/paddle/fluid/operators/detail/rpc_server.cc @@ -0,0 +1,117 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include +#include + +#include "paddle/fluid/operators/detail/rpc_server.h" + +namespace paddle { +namespace operators { +namespace detail { + +void RPCServer::ShutDown() { + LOG(INFO) << "RPCServer ShutDown "; + ShutDownImpl(); + + exit_flag_ = true; + barrier_cond_.notify_all(); + rpc_cond_.notify_all(); +} + +void RPCServer::SavePort() const { + auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); + std::ofstream port_file; + port_file.open(file_path); + port_file << selected_port_; + port_file.close(); + VLOG(4) << "selected port written to " << file_path; +} + +void RPCServer::WaitBarrier(const std::string& rpc_name) { + std::unique_lock lock(this->mutex_); + barrier_cond_.wait(lock, [this, &rpc_name] { + return (barrier_counter_[rpc_name] >= client_num_ || exit_flag_.load()); + }); + + VLOG(3) << "batch_barrier_:" << barrier_counter_[rpc_name]; +} + +void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) { + VLOG(3) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; + int b = 0; + std::unique_lock lock(mutex_); + b = ++barrier_counter_[rpc_name]; + if (b >= client_num_) { + lock.unlock(); + barrier_cond_.notify_all(); + lock.lock(); + } +} + +void RPCServer::DecreaseClientNum() { + { + std::unique_lock lock(mutex_); + client_num_--; + } + barrier_cond_.notify_all(); +} + +void RPCServer::ResetBarrierCounter() { + VLOG(3) << "RPCServer ResetBarrierCounter "; + std::unique_lock lock(mutex_); + for (auto& t : barrier_counter_) { + t.second = 0; + } +} + +void RPCServer::RegisterRPC(const std::string& rpc_name, + RequestHandler* handler, int thread_num) { + rpc_call_map_[rpc_name] = handler; + rpc_thread_num_[rpc_name] = thread_num; + + static int cond = -1; + rpc_cond_map_[rpc_name] = ++cond; + VLOG(4) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler + << ", cond:" << rpc_cond_map_[rpc_name]; +} + +void RPCServer::SetCond(const std::string& rpc_name) { + VLOG(3) << "RPCServer SetCond " << rpc_name; + { + std::unique_lock lock(mutex_); + cur_cond_ = rpc_cond_map_[rpc_name]; + } + + rpc_cond_.notify_all(); +} + +void RPCServer::WaitCond(const std::string& rpc_name) { + VLOG(3) << "RPCServer WaitCond " << rpc_name; + int cond = 0; + { + std::unique_lock lock(mutex_); + cond = rpc_cond_map_[rpc_name]; + } + + std::unique_lock lock(mutex_); + rpc_cond_.wait( + lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); }); +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.h b/paddle/fluid/operators/detail/rpc_server.h new file mode 100644 index 0000000000..2e3342428c --- /dev/null +++ b/paddle/fluid/operators/detail/rpc_server.h @@ -0,0 +1,91 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include // NOLINT +#include +#include +#include "paddle/fluid/operators/detail/request_handler.h" + +namespace paddle { +namespace operators { +namespace detail { + +class RPCServer { + public: + explicit RPCServer(const std::string& address, int client_num) + : cur_cond_(0), + bind_address_(address), + exit_flag_(false), + selected_port_(0), + client_num_(client_num) {} + + virtual ~RPCServer() {} + virtual void StartServer() = 0; + virtual void WaitServerReady() = 0; + + void ShutDown(); + + bool IsExit() { return exit_flag_.load(); } + + int GetSelectedPort() const { return selected_port_; } + void SavePort() const; + + // RegisterRPC, register the rpc method name to a handler + // class, and auto generate a condition id for this call + // to be used for the barrier. + void RegisterRPC(const std::string& rpc_name, RequestHandler* handler, + int thread_num = 5); + + // Wait util all the clients have reached the barrier for one + // rpc method. This function should be called in the + // RequestHandler if you want to run the server/client in a + // synchronous mode. + void WaitBarrier(const std::string& rpc_name); + + void SetCond(const std::string& rpc_name); + void WaitCond(const std::string& rpc_name); + void IncreaseBatchBarrier(const std::string rpc_name); + void DecreaseClientNum(); + void ResetBarrierCounter(); + + protected: + virtual void ShutDownImpl() = 0; + + private: + std::mutex mutex_; + std::unordered_map barrier_counter_; + std::condition_variable barrier_cond_; + + std::unordered_map rpc_cond_map_; + std::atomic cur_cond_; + std::condition_variable rpc_cond_; + + protected: + std::string bind_address_; + std::atomic exit_flag_; + int selected_port_; + int client_num_; + + std::unordered_map rpc_call_map_; + std::unordered_map rpc_thread_num_; + friend class RequestHandler; +}; + +}; // namespace detail +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_server_test.cc b/paddle/fluid/operators/detail/rpc_server_test.cc similarity index 59% rename from paddle/fluid/operators/detail/grpc_server_test.cc rename to paddle/fluid/operators/detail/rpc_server_test.cc index 350a7ee123..463a7b80cf 100644 --- a/paddle/fluid/operators/detail/grpc_server_test.cc +++ b/paddle/fluid/operators/detail/rpc_server_test.cc @@ -17,20 +17,23 @@ limitations under the License. */ #include // NOLINT #include "gtest/gtest.h" -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/detail/grpc_server.h" - #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" +#include "paddle/fluid/operators/detail/rpc_client.h" +#include "paddle/fluid/operators/detail/rpc_server.h" + namespace framework = paddle::framework; namespace platform = paddle::platform; namespace detail = paddle::operators::detail; USE_OP(lookup_table); -std::unique_ptr rpc_service_; +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) { auto root_block = program->MutableBlock(0); @@ -88,53 +91,72 @@ void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place, } } -void StartServer(const std::string& endpoint) { - rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, true)); +void StartServer() { framework::ProgramDesc program; framework::Scope scope; platform::CPUPlace place; framework::Executor exe(place); platform::CPUDeviceContext ctx(place); auto* block = AppendPrefetchBlcok(&program); - auto prepared = exe.Prepare(program, block->ID()); + std::string in_var_name("ids"); + std::vector prefetch_block_ids{block->ID()}; + auto prepared = exe.Prepare(program, prefetch_block_ids); InitTensorsOnServer(&scope, &place, 10); - rpc_service_->SetProgram(&program); - rpc_service_->SetPrefetchPreparedCtx(std::move(prepared)); - rpc_service_->SetDevCtx(&ctx); - rpc_service_->SetScope(&scope); - rpc_service_->SetExecutor(&exe); + std::unordered_map> + prefetch_var_name_to_prepared; + prefetch_var_name_to_prepared[in_var_name] = prepared[0]; + g_req_handler->SetProgram(&program); + g_req_handler->SetPrefetchPreparedCtx(&prefetch_var_name_to_prepared); + g_req_handler->SetDevCtx(&ctx); + g_req_handler->SetScope(&scope); + g_req_handler->SetExecutor(&exe); + + g_rpc_service->RegisterRPC(detail::kRequestPrefetch, g_req_handler.get()); + g_req_handler->SetRPCServer(g_rpc_service.get()); - rpc_service_->RunSyncUpdate(); + std::thread server_thread( + std::bind(&detail::RPCServer::StartServer, g_rpc_service.get())); + + server_thread.join(); } -TEST(PREFETCH, DISABLED_CPU) { - // start up a server instance backend - std::thread server_thread(StartServer, "127.0.0.1:8889"); - sleep(2); +TEST(PREFETCH, CPU) { + g_req_handler.reset(new detail::RequestPrefetchHandler(true)); + g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); + detail::RPCClient* client = detail::RPCClient::GetInstance(); + + std::thread server_thread(StartServer); + g_rpc_service->WaitServerReady(); + + int port = g_rpc_service->GetSelectedPort(); + std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port); + framework::Scope scope; platform::CPUPlace place; platform::CPUDeviceContext ctx(place); - // create var on local scope - int64_t rows_numel = 5; - InitTensorsOnClient(&scope, &place, rows_numel); - std::string in_var_name("ids"); - std::string out_var_name("out"); - - auto client = detail::RPCClient::GetInstance(); - client->AsyncPrefetchVariable("127.0.0.1:8889", ctx, scope, in_var_name, - out_var_name); - client->Wait(); - - auto var = scope.Var(out_var_name); - auto value = var->GetMutable()->value(); - auto ptr = value.mutable_data(place); + { + // create var on local scope + int64_t rows_numel = 5; + InitTensorsOnClient(&scope, &place, rows_numel); + std::string in_var_name("ids"); + std::string out_var_name("out"); + + client->AsyncPrefetchVar(ep, ctx, scope, in_var_name, out_var_name); + client->Wait(); + auto var = scope.Var(out_var_name); + auto value = var->GetMutable()->value(); + auto ptr = value.mutable_data(place); + + for (int64_t i = 0; i < rows_numel; ++i) { + EXPECT_EQ(ptr[0 + i * value.dims()[1]], static_cast(i * 2)); + } + } - rpc_service_->ShutDown(); + g_rpc_service->ShutDown(); server_thread.join(); - rpc_service_.reset(nullptr); - - for (int64_t i = 0; i < rows_numel; ++i) { - EXPECT_EQ(ptr[0 + i * value.dims()[1]], static_cast(i * 2)); - } + LOG(INFO) << "begin reset"; + g_rpc_service.reset(nullptr); + g_req_handler.reset(nullptr); } diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto index a244afc46f..54cb93e04d 100644 --- a/paddle/fluid/operators/detail/send_recv.proto +++ b/paddle/fluid/operators/detail/send_recv.proto @@ -14,6 +14,8 @@ limitations under the License. */ syntax = "proto3"; package sendrecv; +// option cc_generic_services = true; + service SendRecvService { // For parameter server round-robin like hashing, do not split tensors. // Send and recv only one tensor diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.h b/paddle/fluid/operators/detail/sendrecvop_utils.h index c72e1bd076..bd16bf1dab 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.h +++ b/paddle/fluid/operators/detail/sendrecvop_utils.h @@ -32,16 +32,6 @@ namespace paddle { namespace operators { namespace detail { -#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" -#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV" -#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV" - -static int64_t GetTimestamp() { - struct timeval tp; - gettimeofday(&tp, NULL); - return tp.tv_sec * 1000 + tp.tv_usec / 1000; -} - typedef void (*DestroyCallback)(void*); void SerializeToByteBuffer(const std::string& name, framework::Variable* var, diff --git a/paddle/fluid/operators/detail/variable_response.h b/paddle/fluid/operators/detail/variable_response.h index bf624da2a6..69cfd784f8 100644 --- a/paddle/fluid/operators/detail/variable_response.h +++ b/paddle/fluid/operators/detail/variable_response.h @@ -67,8 +67,8 @@ class VariableResponse { framework::Scope* GetMutableLocalScope() const { return local_scope_; } - inline std::string Varname() { return meta_.varname(); } - inline std::string OutVarname() { return meta_.out_varname(); } + inline std::string Varname() const { return meta_.varname(); } + inline std::string OutVarname() const { return meta_.out_varname(); } // should call parse first. framework::Variable* GetVar() { diff --git a/paddle/fluid/operators/detection/box_coder_op.cc b/paddle/fluid/operators/detection/box_coder_op.cc index 74848005d0..8c4b4321b7 100644 --- a/paddle/fluid/operators/detection/box_coder_op.cc +++ b/paddle/fluid/operators/detection/box_coder_op.cc @@ -22,21 +22,21 @@ class BoxCoderOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("PriorBox"), "Input(PriorBox) of BoxCoderOp should not be null."); - PADDLE_ENFORCE(ctx->HasInput("PriorBoxVar"), - "Input(PriorBoxVar) of BoxCoderOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("TargetBox"), "Input(TargetBox) of BoxCoderOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("OutputBox"), "Output(OutputBox) of BoxCoderOp should not be null."); auto prior_box_dims = ctx->GetInputDim("PriorBox"); - auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar"); auto target_box_dims = ctx->GetInputDim("TargetBox"); PADDLE_ENFORCE_EQ(prior_box_dims.size(), 2, "The rank of Input of PriorBoxVar must be 2"); PADDLE_ENFORCE_EQ(prior_box_dims[1], 4, "The shape of PriorBox is [N, 4]"); - PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims); + if (ctx->HasInput("PriorBoxVar")) { + auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar"); + PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims); + } auto code_type = GetBoxCodeType(ctx->Attrs().Get("code_type")); if (code_type == BoxCodeType::kEncodeCenterSize) { @@ -71,9 +71,11 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { "of the coordinate system. [xmax, ymax] is the right bottom " "coordinate of the anchor box."); AddInput("PriorBoxVar", - "(Tensor, default Tensor) " + "(Tensor, default Tensor, optional) " "PriorBoxVar is a 2-D Tensor with shape [M, 4] holds M group " - "of variance."); + "of variance. PriorBoxVar will set all elements to 1 by " + "default.") + .AsDispensable(); AddInput( "TargetBox", "(LoDTensor or Tensor) This input can be a 2-D LoDTensor with shape " @@ -91,6 +93,10 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker { "the code type used with the target box") .SetDefault("encode_center_size") .InEnum({"encode_center_size", "decode_center_size"}); + AddAttr("box_normalized", + "(bool, default true) " + "whether treat the priorbox as a noramlized box") + .SetDefault(true); AddOutput("OutputBox", "(LoDTensor or Tensor) " "When code_type is 'encode_center_size', the output tensor of " @@ -127,5 +133,6 @@ width and height. namespace ops = paddle::operators; REGISTER_OPERATOR(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker, paddle::framework::EmptyGradOpMaker); -REGISTER_OP_CPU_KERNEL(box_coder, ops::BoxCoderKernel, - ops::BoxCoderKernel); +REGISTER_OP_CPU_KERNEL( + box_coder, ops::BoxCoderKernel, + ops::BoxCoderKernel); diff --git a/paddle/fluid/operators/detection/box_coder_op.cu b/paddle/fluid/operators/detection/box_coder_op.cu index 8cef8e0343..a7af111f63 100644 --- a/paddle/fluid/operators/detection/box_coder_op.cu +++ b/paddle/fluid/operators/detection/box_coder_op.cu @@ -20,15 +20,16 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data, const T* prior_box_var_data, const T* target_box_data, const int row, const int col, const int len, - T* output) { + const bool normalized, T* output) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < row * col) { const int row_idx = idx / col; const int col_idx = idx % col; - T prior_box_width = - prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len]; - T prior_box_height = - prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1]; + T prior_box_width = prior_box_data[col_idx * len + 2] - + prior_box_data[col_idx * len] + (normalized == false); + T prior_box_height = prior_box_data[col_idx * len + 3] - + prior_box_data[col_idx * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2; T prior_box_center_y = (prior_box_data[col_idx * len + 3] + @@ -41,20 +42,24 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data, T target_box_center_y = (target_box_data[row_idx * len + 3] + target_box_data[row_idx * len + 1]) / 2; - T target_box_width = - target_box_data[row_idx * len + 2] - target_box_data[row_idx * len]; - T target_box_height = - target_box_data[row_idx * len + 3] - target_box_data[row_idx * len + 1]; + T target_box_width = target_box_data[row_idx * len + 2] - + target_box_data[row_idx * len] + (normalized == false); + T target_box_height = target_box_data[row_idx * len + 3] - + target_box_data[row_idx * len + 1] + + (normalized == false); - output[idx * len] = (target_box_center_x - prior_box_center_x) / - prior_box_width / prior_box_var_data[col_idx * len]; - output[idx * len + 1] = (target_box_center_y - prior_box_center_y) / - prior_box_height / - prior_box_var_data[col_idx * len + 1]; - output[idx * len + 2] = log(fabs(target_box_width / prior_box_width)) / - prior_box_var_data[col_idx * len + 2]; - output[idx * len + 3] = log(fabs(target_box_height / prior_box_height)) / - prior_box_var_data[col_idx * len + 3]; + output[idx * len] = + (target_box_center_x - prior_box_center_x) / prior_box_width; + output[idx * len + 1] = + (target_box_center_y - prior_box_center_y) / prior_box_height; + output[idx * len + 2] = log(fabs(target_box_width / prior_box_width)); + output[idx * len + 3] = log(fabs(target_box_height / prior_box_height)); + if (prior_box_var_data) { + output[idx * len] /= prior_box_var_data[col_idx * len]; + output[idx * len + 1] /= prior_box_var_data[col_idx * len + 1]; + output[idx * len + 2] /= prior_box_var_data[col_idx * len + 2]; + output[idx * len + 3] /= prior_box_var_data[col_idx * len + 3]; + } } } @@ -63,42 +68,56 @@ __global__ void DecodeCenterSizeKernel(const T* prior_box_data, const T* prior_box_var_data, const T* target_box_data, const int row, const int col, const int len, - T* output) { + const bool normalized, T* output) { const int idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < row * col) { const int col_idx = idx % col; - T prior_box_width = - prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len]; - T prior_box_height = - prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1]; + T prior_box_width = prior_box_data[col_idx * len + 2] - + prior_box_data[col_idx * len] + (normalized == false); + T prior_box_height = prior_box_data[col_idx * len + 3] - + prior_box_data[col_idx * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2; T prior_box_center_y = (prior_box_data[col_idx * len + 3] + prior_box_data[col_idx * len + 1]) / 2; - - T target_box_width = exp(prior_box_var_data[col_idx * len + 2] * + T target_box_width, target_box_height; + T target_box_center_x, target_box_center_y; + if (prior_box_var_data) { + target_box_width = exp(prior_box_var_data[col_idx * len + 2] * target_box_data[idx * len + 2]) * prior_box_width; - T target_box_height = exp(prior_box_var_data[col_idx * len + 3] * + target_box_height = exp(prior_box_var_data[col_idx * len + 3] * target_box_data[idx * len + 3]) * prior_box_height; - T target_box_center_x = prior_box_var_data[col_idx * len] * + target_box_center_x = prior_box_var_data[col_idx * len] * target_box_data[idx * len] * prior_box_width + prior_box_center_x; - T target_box_center_y = prior_box_var_data[col_idx * len + 1] * + target_box_center_y = prior_box_var_data[col_idx * len + 1] * target_box_data[idx * len + 1] * prior_box_height + prior_box_center_y; + } else { + target_box_width = exp(target_box_data[idx * len + 2]) * prior_box_width; + target_box_height = + exp(target_box_data[idx * len + 3]) * prior_box_height; + target_box_center_x = + target_box_data[idx * len] * prior_box_width + prior_box_center_x; + target_box_center_y = target_box_data[idx * len + 1] * prior_box_height + + prior_box_center_y; + } output[idx * len] = target_box_center_x - target_box_width / 2; output[idx * len + 1] = target_box_center_y - target_box_height / 2; - output[idx * len + 2] = target_box_center_x + target_box_width / 2; - output[idx * len + 3] = target_box_center_y + target_box_height / 2; + output[idx * len + 2] = + target_box_center_x + target_box_width / 2 - (normalized == false); + output[idx * len + 3] = + target_box_center_y + target_box_height / 2 - (normalized == false); } } -template +template class BoxCoderCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { @@ -109,6 +128,11 @@ class BoxCoderCUDAKernel : public framework::OpKernel { auto* target_box = context.Input("TargetBox"); auto* output_box = context.Output("OutputBox"); + const T* prior_box_data = prior_box->data(); + const T* target_box_data = target_box->data(); + const T* prior_box_var_data = nullptr; + if (prior_box_var) prior_box_var_data = prior_box_var->data(); + if (target_box->lod().size()) { PADDLE_ENFORCE_EQ(target_box->lod().size(), 1, "Only support 1 level of LoD."); @@ -120,22 +144,19 @@ class BoxCoderCUDAKernel : public framework::OpKernel { int grid = (row * col + block - 1) / block; auto& device_ctx = context.cuda_device_context(); - const T* prior_box_data = prior_box->data(); - const T* prior_box_var_data = prior_box_var->data(); - const T* target_box_data = target_box->data(); - output_box->mutable_data({row, col, len}, context.GetPlace()); T* output = output_box->data(); auto code_type = GetBoxCodeType(context.Attr("code_type")); + bool normalized = context.Attr("box_normalized"); if (code_type == BoxCodeType::kEncodeCenterSize) { EncodeCenterSizeKernel<<>>( prior_box_data, prior_box_var_data, target_box_data, row, col, len, - output); + normalized, output); } else if (code_type == BoxCodeType::kDecodeCenterSize) { DecodeCenterSizeKernel<<>>( prior_box_data, prior_box_var_data, target_box_data, row, col, len, - output); + normalized, output); } } }; @@ -144,5 +165,7 @@ class BoxCoderCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(box_coder, ops::BoxCoderCUDAKernel, - ops::BoxCoderCUDAKernel); +REGISTER_OP_CUDA_KERNEL( + box_coder, + ops::BoxCoderCUDAKernel, + ops::BoxCoderCUDAKernel); diff --git a/paddle/fluid/operators/detection/box_coder_op.h b/paddle/fluid/operators/detection/box_coder_op.h index 77fc6c2b62..5ed8520acd 100644 --- a/paddle/fluid/operators/detection/box_coder_op.h +++ b/paddle/fluid/operators/detection/box_coder_op.h @@ -28,26 +28,28 @@ inline BoxCodeType GetBoxCodeType(const std::string& type) { PADDLE_THROW("Not support type %s.", type); } -template +template class BoxCoderKernel : public framework::OpKernel { public: - void EncodeCenterSize(const framework::Tensor& target_box, - const framework::Tensor& prior_box, - const framework::Tensor& prior_box_var, - T* output) const { - int64_t row = target_box.dims()[0]; - int64_t col = prior_box.dims()[0]; - int64_t len = prior_box.dims()[1]; - auto* target_box_data = target_box.data(); - auto* prior_box_data = prior_box.data(); - auto* prior_box_var_data = prior_box_var.data(); + void EncodeCenterSize(const framework::Tensor* target_box, + const framework::Tensor* prior_box, + const framework::Tensor* prior_box_var, + const bool normalized, T* output) const { + int64_t row = target_box->dims()[0]; + int64_t col = prior_box->dims()[0]; + int64_t len = prior_box->dims()[1]; + auto* target_box_data = target_box->data(); + auto* prior_box_data = prior_box->data(); + const T* prior_box_var_data = nullptr; + if (prior_box_var) prior_box_var_data = prior_box_var->data(); for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { - T prior_box_width = - prior_box_data[j * len + 2] - prior_box_data[j * len]; - T prior_box_height = - prior_box_data[j * len + 3] - prior_box_data[j * len + 1]; + T prior_box_width = prior_box_data[j * len + 2] - + prior_box_data[j * len] + (normalized == false); + T prior_box_height = prior_box_data[j * len + 3] - + prior_box_data[j * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2; T prior_box_center_y = @@ -57,67 +59,89 @@ class BoxCoderKernel : public framework::OpKernel { (target_box_data[i * len + 2] + target_box_data[i * len]) / 2; T target_box_center_y = (target_box_data[i * len + 3] + target_box_data[i * len + 1]) / 2; - T target_box_width = - target_box_data[i * len + 2] - target_box_data[i * len]; - T target_box_height = - target_box_data[i * len + 3] - target_box_data[i * len + 1]; + T target_box_width = target_box_data[i * len + 2] - + target_box_data[i * len] + (normalized == false); + T target_box_height = target_box_data[i * len + 3] - + target_box_data[i * len + 1] + + (normalized == false); size_t offset = i * col * len + j * len; - output[offset] = (target_box_center_x - prior_box_center_x) / - prior_box_width / prior_box_var_data[j * len]; - output[offset + 1] = (target_box_center_y - prior_box_center_y) / - prior_box_height / prior_box_var_data[j * len + 1]; + output[offset] = + (target_box_center_x - prior_box_center_x) / prior_box_width; + output[offset + 1] = + (target_box_center_y - prior_box_center_y) / prior_box_height; output[offset + 2] = - std::log(std::fabs(target_box_width / prior_box_width)) / - prior_box_var_data[j * len + 2]; + std::log(std::fabs(target_box_width / prior_box_width)); output[offset + 3] = - std::log(std::fabs(target_box_height / prior_box_height)) / - prior_box_var_data[j * len + 3]; + std::log(std::fabs(target_box_height / prior_box_height)); + if (prior_box_var) { + output[offset] /= prior_box_var_data[j * len]; + output[offset + 1] /= prior_box_var_data[j * len + 1]; + output[offset + 2] /= prior_box_var_data[j * len + 2]; + output[offset + 3] /= prior_box_var_data[j * len + 3]; + } } } } - void DecodeCenterSize(const framework::Tensor& target_box, - const framework::Tensor& prior_box, - const framework::Tensor& prior_box_var, - T* output) const { - int64_t row = target_box.dims()[0]; - int64_t col = prior_box.dims()[0]; - int64_t len = prior_box.dims()[1]; - - auto* target_box_data = target_box.data(); - auto* prior_box_data = prior_box.data(); - auto* prior_box_var_data = prior_box_var.data(); + void DecodeCenterSize(const framework::Tensor* target_box, + const framework::Tensor* prior_box, + const framework::Tensor* prior_box_var, + const bool normalized, T* output) const { + int64_t row = target_box->dims()[0]; + int64_t col = prior_box->dims()[0]; + int64_t len = prior_box->dims()[1]; + + auto* target_box_data = target_box->data(); + auto* prior_box_data = prior_box->data(); + const T* prior_box_var_data = nullptr; + if (prior_box_var) prior_box_var_data = prior_box_var->data(); for (int64_t i = 0; i < row; ++i) { for (int64_t j = 0; j < col; ++j) { size_t offset = i * col * len + j * len; - T prior_box_width = - prior_box_data[j * len + 2] - prior_box_data[j * len]; - T prior_box_height = - prior_box_data[j * len + 3] - prior_box_data[j * len + 1]; + T prior_box_width = prior_box_data[j * len + 2] - + prior_box_data[j * len] + (normalized == false); + T prior_box_height = prior_box_data[j * len + 3] - + prior_box_data[j * len + 1] + + (normalized == false); T prior_box_center_x = (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2; T prior_box_center_y = (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2; - T target_box_center_x = prior_box_var_data[j * len] * + T target_box_center_x = 0, target_box_center_y = 0; + T target_box_width = 0, target_box_height = 0; + if (prior_box_var) { + target_box_center_x = prior_box_var_data[j * len] * target_box_data[offset] * prior_box_width + prior_box_center_x; - T target_box_center_y = prior_box_var_data[j * len + 1] * + target_box_center_y = prior_box_var_data[j * len + 1] * target_box_data[offset + 1] * prior_box_height + prior_box_center_y; - T target_box_width = std::exp(prior_box_var_data[j * len + 2] * + target_box_width = std::exp(prior_box_var_data[j * len + 2] * target_box_data[offset + 2]) * prior_box_width; - T target_box_height = std::exp(prior_box_var_data[j * len + 3] * + target_box_height = std::exp(prior_box_var_data[j * len + 3] * target_box_data[offset + 3]) * prior_box_height; + } else { + target_box_center_x = + target_box_data[offset] * prior_box_width + prior_box_center_x; + target_box_center_y = target_box_data[offset + 1] * prior_box_height + + prior_box_center_y; + target_box_width = + std::exp(target_box_data[offset + 2]) * prior_box_width; + target_box_height = + std::exp(target_box_data[offset + 3]) * prior_box_height; + } output[offset] = target_box_center_x - target_box_width / 2; output[offset + 1] = target_box_center_y - target_box_height / 2; - output[offset + 2] = target_box_center_x + target_box_width / 2; - output[offset + 3] = target_box_center_y + target_box_height / 2; + output[offset + 2] = + target_box_center_x + target_box_width / 2 - (normalized == false); + output[offset + 3] = + target_box_center_y + target_box_height / 2 - (normalized == false); } } } @@ -139,11 +163,14 @@ class BoxCoderKernel : public framework::OpKernel { output_box->mutable_data({row, col, len}, context.GetPlace()); auto code_type = GetBoxCodeType(context.Attr("code_type")); + bool normalized = context.Attr("box_normalized"); T* output = output_box->data(); if (code_type == BoxCodeType::kEncodeCenterSize) { - EncodeCenterSize(*target_box, *prior_box, *prior_box_var, output); + EncodeCenterSize(target_box, prior_box, prior_box_var, normalized, + output); } else if (code_type == BoxCodeType::kDecodeCenterSize) { - DecodeCenterSize(*target_box, *prior_box, *prior_box_var, output); + DecodeCenterSize(target_box, prior_box, prior_box_var, normalized, + output); } } }; diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h index f4cec8ad97..12364fff96 100644 --- a/paddle/fluid/operators/elementwise_op.h +++ b/paddle/fluid/operators/elementwise_op.h @@ -59,47 +59,48 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { void Make() final { AddInput("X", "(Tensor), The first input tensor of elementwise op."); AddInput("Y", "(Tensor), The second input tensor of elementwise op."); - AddOutput("Out", "The output of elementwise op."); + AddOutput("Out", "The output of elementwise op.").Reuse("X"); AddAttr("axis", "(int, default -1). The start dimension index " "for broadcasting Y onto X.") .SetDefault(-1) .EqualGreaterThan(-1); AddComment(string::Sprintf(R"DOC( -Limited Elementwise %s Operator. +Limited Elementwise %s Operator The equation is: $$%s$$ -$X$ is a tensor of any dimension and the dimensions of tensor $Y$ must be -smaller than or equal to the dimensions of $X$. +- $X$: a tensor of any dimension. +- $Y$: a tensor whose dimensions must be less than or equal to the dimensions of $X$. There are two cases for this operator: -1. The shape of $Y$ is same with $X$; -2. The shape of $Y$ is a congiguous subsequencet of $X$. The trailing dimensions - of size 1 for $Y$ will be ignored for the consideration of subsequence. +1. The shape of $Y$ is the same with $X$. +2. The shape of $Y$ is a continuous subsequence of $X$. For case 2: -$Y$ will be broadcasted to match the shape of $X$ and axis should be -set to index of the start dimension to broadcast $Y$ onto $X$. +1. Broadcast $Y$ to match the shape of $X$, where $axis$ is the start dimension index + for broadcasting $Y$ onto $X$. +2. If $axis$ is -1 (default), $axis = rank(X) - rank(Y)$. +3. The trailing dimensions of size 1 for $Y$ will be ignored for the consideration of + subsequence, such as shape(Y) = (2, 1) => (2). -If axis is -1, it is treated as axis=rank(X)-rank(Y). +For example: -For example .. code-block:: python shape(X) = (2, 3, 4, 5), shape(Y) = (,) shape(X) = (2, 3, 4, 5), shape(Y) = (5,) - shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5) + shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5), with axis=-1(default) or axis=2 shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1 shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0 shape(X) = (2, 3, 4, 5), shape(Y) = (2, 1), with axis=0 -Either of the inputs $X$ and $Y$ or none can carry the LoD (Level of Details) -information. However, the output only shares the LoD information with input $X$. +The inputs $X$ and $Y$ can carry the different LoD information. +But the output only shares the LoD information with the input $X$. )DOC", GetName(), GetEquation())); diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index 8843a1c44b..a9ae1396db 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -43,7 +43,7 @@ void FCOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType FCOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { framework::LibraryType library{framework::LibraryType::kMKLDNN}; - framework::DataLayout layout{framework::DataLayout::kAnyLayout}; + framework::DataLayout layout{framework::DataLayout::kMKLDNN}; return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), @@ -65,7 +65,7 @@ void FCOpGrad::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType FCOpGrad::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { framework::LibraryType library{framework::LibraryType::kMKLDNN}; - framework::DataLayout layout{framework::DataLayout::kAnyLayout}; + framework::DataLayout layout{framework::DataLayout::kMKLDNN}; return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc index 79ec02f520..98b051afb5 100644 --- a/paddle/fluid/operators/fetch_barrier_op.cc +++ b/paddle/fluid/operators/fetch_barrier_op.cc @@ -19,8 +19,7 @@ limitations under the License. */ #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" - -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { @@ -43,15 +42,16 @@ class FetchBarrierOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - auto rpc_client = detail::RPCClient::GetInstance(); + detail::RPCClient* rpc_client = + detail::RPCClient::GetInstance(); - PADDLE_ENFORCE(rpc_client->Wait()); + rpc_client->Wait(); for (auto& ep : eps) { VLOG(3) << "fetch barrier, ep: " << ep; rpc_client->AsyncSendFetchBarrier(ep); } - PADDLE_ENFORCE(rpc_client->Wait()); + rpc_client->Wait(); } }; diff --git a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc index 1ae78675a0..453a1b32a0 100644 --- a/paddle/fluid/operators/fill_constant_batch_size_like_op.cc +++ b/paddle/fluid/operators/fill_constant_batch_size_like_op.cc @@ -32,16 +32,16 @@ class FillConstantBatchSizeLikeOp : public BatchSizeLikeOp { class FillConstantBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker { protected: void Apply() override { - AddAttr("dtype", - "(int, default 5 (FP32)) " - "Output data type") + AddAttr( + "dtype", + "It could be numpy.dtype. Output data type. Default is float32") .SetDefault(framework::proto::VarType::FP32); - AddAttr("value", "(float, default 0) The value to be filled") + AddAttr("value", "default 0. The value to be filled") .SetDefault(0.0f); AddComment(R"DOC( -FillConstantBatchSizeLike Operator. - -Fill up a variable with specified constant value. +This function creates a tensor of specified *shape*, *dtype* and batch size, +and initializes this with a constant supplied in *value*. The batch size is +obtained from the `input` tensor. )DOC"); } diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc index e21b572589..aa3e05b83b 100644 --- a/paddle/fluid/operators/gather_op.cc +++ b/paddle/fluid/operators/gather_op.cc @@ -33,7 +33,6 @@ class GatherOp : public framework::OperatorWithKernel { auto index_dims = ctx->GetInputDim("Index"); PADDLE_ENFORCE(index_dims.size() == 1); int batch_size = ctx->GetInputDim("Index")[0]; - PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0"); framework::DDim output_dims(ctx->GetInputDim("X")); output_dims[0] = batch_size; ctx->SetOutputDim("Out", output_dims); diff --git a/paddle/fluid/operators/gather_test.cc b/paddle/fluid/operators/gather_test.cc index 9c0561b016..f6b156eb30 100644 --- a/paddle/fluid/operators/gather_test.cc +++ b/paddle/fluid/operators/gather_test.cc @@ -43,7 +43,8 @@ TEST(Gather, GatherData) { auto* cpu_place = new paddle::platform::CPUPlace(); paddle::platform::CPUDeviceContext ctx(*cpu_place); paddle::operators::CPUGather(ctx, *src, *index, output); - + delete cpu_place; + cpu_place = NULL; for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4); for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4); diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc index a5678f6346..f824eee4e7 100644 --- a/paddle/fluid/operators/gen_nccl_id_op.cc +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -21,8 +21,8 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" #include "paddle/fluid/platform/nccl_helper.h" namespace paddle { @@ -60,12 +60,17 @@ class GenNCCLIdOp : public framework::OperatorBase { std::vector endpoint_list = Attr>("endpoint_list"); - detail::RPCClient client; + detail::RPCClient* client = detail::RPCClient::GetInstance(); + for (auto& ep : endpoint_list) { VLOG(3) << "sending nccl id to " << ep; - client.AsyncSendVariable(ep, dev_ctx, *scope, NCCL_ID_VARNAME); + client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME); + } + client->Wait(); + for (auto& ep : endpoint_list) { + client->AsyncSendBatchBarrier(ep); } - client.Wait(); + client->Wait(); VLOG(3) << "sending completed..."; } @@ -75,21 +80,28 @@ class GenNCCLIdOp : public framework::OperatorBase { // NOTE: Can not use unique_ptr here because the default // deleter will call GRPC Server's base class's dtor and // that will cause a wired crash. - detail::AsyncGRPCServer rpc_service(endpoint, true); + detail::RequestSendHandler rpc_h(true); + std::unique_ptr rpc_service( + new RPCSERVER_T(endpoint, 1)); + + rpc_service->RegisterRPC(detail::kRequestSend, &rpc_h); + rpc_h.SetRPCServer(rpc_service.get()); + framework::ProgramDesc empty_program; framework::Executor executor(dev_ctx.GetPlace()); - rpc_service.SetScope(scope); - rpc_service.SetDevCtx(&dev_ctx); - rpc_service.SetProgram(&empty_program); - rpc_service.SetExecutor(&executor); + rpc_h.SetScope(scope); + rpc_h.SetDevCtx(&dev_ctx); + rpc_h.SetProgram(&empty_program); + rpc_h.SetExecutor(&executor); std::thread server_thread( - std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, &rpc_service)); - rpc_service.SetCond(0); + std::bind(&detail::RPCServer::StartServer, rpc_service.get())); + + rpc_service->SetCond(detail::kRequestSend); VLOG(3) << "start getting nccl id from trainer 0..."; - auto recv = rpc_service.Get(); + rpc_service->WaitBarrier(detail::kRequestSend); VLOG(3) << "got nccl id and stop server..."; - rpc_service.ShutDown(); + rpc_service->ShutDown(); VLOG(3) << "rpc server stopped"; server_thread.join(); } diff --git a/paddle/fluid/operators/get_places_op.cc b/paddle/fluid/operators/get_places_op.cc index eafc364a15..db6ff78256 100644 --- a/paddle/fluid/operators/get_places_op.cc +++ b/paddle/fluid/operators/get_places_op.cc @@ -85,7 +85,7 @@ class GetPlacesOpProtoMaker : public framework::OpProtoAndCheckerMaker { .InEnum({"CUDA", "CPU", "AUTO"}) .SetDefault("AUTO"); AddComment(R"DOC( -Returns a list of places based on flags. The list will be used for parallel +Returns a list of places based on arguments. The list will be used for parallel execution. )DOC"); } diff --git a/paddle/fluid/operators/linear_chain_crf_op.cc b/paddle/fluid/operators/linear_chain_crf_op.cc index e38525cd7f..a711da3627 100644 --- a/paddle/fluid/operators/linear_chain_crf_op.cc +++ b/paddle/fluid/operators/linear_chain_crf_op.cc @@ -67,8 +67,6 @@ class LinearChainCRFOpMaker : public framework::OpProtoAndCheckerMaker { "mini-batch. Note: S is equal to the sequence number in a mini-batch. " "The output is no longer a LoDTensor."); AddComment(R"DOC( -LinearChainCRF Operator. - Conditional Random Field defines an undirected probabilistic graph with nodes denoting random variables and edges denoting dependencies between these variables. CRF learns the conditional probability $P(Y|X)$, where diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index df5f229acd..4d12278799 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -19,14 +19,17 @@ limitations under the License. */ #include // NOLINT #include +#include "paddle/fluid/operators/detail/macros.h" + +#include "paddle/fluid/operators/detail/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -void RunServer(std::shared_ptr service) { - service->RunSyncUpdate(); +void RunServer(std::shared_ptr service) { + service->StartServer(); VLOG(4) << "RunServer thread end"; } static void split(const std::string &str, char sep, @@ -67,8 +70,6 @@ static void ParallelExecuteBlocks( for (size_t i = 0; i < fs.size(); ++i) fs[i].wait(); } -std::atomic_int ListenAndServOp::selected_port_{0}; - ListenAndServOp::ListenAndServOp(const std::string &type, const framework::VariableNameMap &inputs, const framework::VariableNameMap &outputs, @@ -78,7 +79,6 @@ ListenAndServOp::ListenAndServOp(const std::string &type, ListenAndServOp::~ListenAndServOp() { Stop(); } void ListenAndServOp::Stop() { - rpc_service_->Push(LISTEN_TERMINATE_MESSAGE); rpc_service_->ShutDown(); server_thread_->join(); auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); @@ -87,159 +87,88 @@ void ListenAndServOp::Stop() { void ListenAndServOp::SavePort() const { // NOTE: default write file to /tmp/paddle.selected_port - selected_port_ = rpc_service_->GetSelectedPort(); - auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); - std::ofstream port_file; - port_file.open(file_path); - port_file << selected_port_.load(); - port_file.close(); - VLOG(4) << "selected port written to " << file_path; + rpc_service_->SavePort(); } -void ListenAndServOp::WaitServerReady() { - while (selected_port_.load() == 0) { - } +static int64_t GetTimestamp() { + struct timeval tp; + gettimeofday(&tp, NULL); + return tp.tv_sec * 1000 + tp.tv_usec / 1000; } -void ListenAndServOp::RunSyncLoop(framework::Executor *executor, - framework::ProgramDesc *program, - framework::Scope *recv_scope, - framework::BlockDesc *prefetch_block) const { - auto fan_in = Attr("Fanin"); - +void ListenAndServOp::RunSyncLoop( + framework::Executor *executor, framework::ProgramDesc *program, + framework::Scope *recv_scope, + const std::vector &prefetch_block_id_list) const { size_t num_blocks = program->Size(); PADDLE_ENFORCE_GE(num_blocks, 2, "server program should have at least 2 blocks"); - std::vector block_list; - for (size_t blkid = 1; blkid < num_blocks; ++blkid) { - block_list.push_back(blkid); + std::vector optimize_block_id_list; + for (int blkid = 1; blkid < num_blocks; ++blkid) { + if (std::find(prefetch_block_id_list.begin(), prefetch_block_id_list.end(), + blkid) == prefetch_block_id_list.end()) { + optimize_block_id_list.push_back(blkid); + } } - auto optimize_prepared = executor->Prepare(*program, block_list); + auto optimize_prepared = executor->Prepare(*program, optimize_block_id_list); // Insert placeholder for block0 which holds current op itself. optimize_prepared.insert( optimize_prepared.begin(), std::shared_ptr(nullptr)); - bool exit_flag = false; - // Record received sparse variables, so that - // we could reset those after execute optimize program - std::vector sparse_vars; - while (!exit_flag && !SignalHandler::IsProgramExit()) { + rpc_service_->ResetBarrierCounter(); + while (true) { // Get from multiple trainers, we don't care about the order in which // the gradients arrives, just add suffix 0~n and merge the gradient. - rpc_service_->SetCond(0); - size_t recv_var_cnt = 0; - int batch_barrier = 0; - while (batch_barrier != fan_in) { - const detail::ReceivedMessage v = rpc_service_->Get(); - auto recv_var_name = v.first; - if (recv_var_name == LISTEN_TERMINATE_MESSAGE) { - LOG(INFO) << "received terminate message and exit"; - exit_flag = true; - break; - } else if (recv_var_name == BATCH_BARRIER_MESSAGE) { - VLOG(3) << "recv batch barrier message"; - batch_barrier++; - continue; - } else { - VLOG(3) << "received grad: " << recv_var_name; - recv_var_cnt++; - auto var = v.second->GetVar(); - if (var == nullptr) { - LOG(ERROR) << "Can not find server side var: " << recv_var_name; - PADDLE_THROW("Can not find server side var"); - } - if (var->IsType()) { - sparse_vars.push_back(var); - } - } - } - if (exit_flag) { - rpc_service_->SetCond(1); - rpc_service_->ShutDown(); + rpc_service_->SetCond(detail::kRequestSend); + rpc_service_->WaitBarrier(detail::kRequestSend); + + if (rpc_service_->IsExit()) { + LOG(WARNING) << "get exit!rpc_processor break!"; + rpc_service_->SetCond(detail::kRequestGet); break; } // NOTE: if is_gpu_place, CUDA kernels are launched by multiple threads // and this will still work. - // The optimize blocks which have the same parent ID would run parallel // TODO(Yancey1989): need to use ParallelExecutor for future int32_t last_parent_blkid = program->Block(1).Parent(); std::vector parallel_blkids; parallel_blkids.push_back(1); - double ts = detail::GetTimestamp(); - for (size_t blkid = 2; blkid < num_blocks; ++blkid) { - if (blkid != static_cast(prefetch_block->ID())) { - if (program->Block(blkid).Parent() != last_parent_blkid) { - ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, - program, recv_scope); - parallel_blkids.clear(); - last_parent_blkid = program->Block(blkid).Parent(); - } - parallel_blkids.push_back(blkid); + double ts = GetTimestamp(); + for (size_t i = 1; i < optimize_block_id_list.size(); ++i) { + // skip the first optimize block because it is already in the + // parallel_blkids. + int blkid = optimize_block_id_list[i]; + if (program->Block(blkid).Parent() != last_parent_blkid) { + ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, + program, recv_scope); + parallel_blkids.clear(); + last_parent_blkid = program->Block(blkid).Parent(); } + parallel_blkids.push_back(blkid); } ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, program, recv_scope); - VLOG(2) << "run all blocks spent " << detail::GetTimestamp() - ts << "(ms)"; - - // Reset the received sparse variables, the sum operator would not - // sum the input sparse variables which rows is empty at the next - // mini-batch. - // TODO(Yancey1989): move the reset action into an operator, we couldn't - // have any hide logic in the operator. - for (framework::Variable *var : sparse_vars) { - var->GetMutable()->mutable_rows()->clear(); - } - - rpc_service_->SetCond(1); - // FIXME(typhoonzero): use another condition to sync wait clients get. - rpc_service_->WaitClientGet(fan_in); - sparse_vars.clear(); + VLOG(2) << "run all blocks spent " << GetTimestamp() - ts << "(ms)"; + + rpc_service_->SetCond(detail::kRequestGet); + rpc_service_->WaitBarrier(detail::kRequestGet); + rpc_service_->ResetBarrierCounter(); + // reset received sparse vars to avoid reuse it in the next mini-batch + dynamic_cast(request_send_handler_.get()) + ->ResetSparseVarRecorder(); } // while(true) } -static void AsyncUpdateThread( - const std::string &var_name, const bool &exit_flag, - const std::shared_ptr &queue, - framework::Executor *executor, - framework::ExecutorPrepareContext *prepared) { - VLOG(3) << "update thread for " << var_name << " started"; - while (!exit_flag && !SignalHandler::IsProgramExit()) { - const detail::ReceivedMessage v = queue->Pop(); - if (SignalHandler::IsProgramExit()) { - VLOG(3) << "update thread for " << var_name << " exit"; - break; - } - auto recv_var_name = v.first; - VLOG(4) << "async update " << recv_var_name; - auto var = v.second->GetVar(); - if (var == nullptr) { - LOG(ERROR) << "Can not find server side var: " << recv_var_name; - PADDLE_THROW("Can not find server side var"); - } - auto fs = framework::Async([var_name, &executor, &v, prepared] { - try { - executor->RunPreparedContext(prepared, - v.second->GetMutableLocalScope()); - } catch (const std::exception &e) { - LOG(ERROR) << "run sub program error " << e.what(); - } - }); - fs.wait(); - } -} - void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, framework::ProgramDesc *program) const { VLOG(3) << "RunAsyncLoop in"; // grad name to block id std::unordered_map grad_to_block_id; std::unordered_map id_to_grad; - std::unordered_map> - grad_to_queue; auto grad_to_block_id_str = Attr>("grad_to_block_id"); @@ -249,13 +178,9 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, VLOG(3) << "after split, grad = " << pieces[0] << ", id=" << pieces[1]; PADDLE_ENFORCE_EQ(pieces.size(), 2); PADDLE_ENFORCE_EQ(grad_to_block_id.count(pieces[0]), 0); + int block_id = std::stoi(pieces[1]); grad_to_block_id[pieces[0]] = block_id; - std::shared_ptr queue = - std::make_shared(); - grad_to_queue[pieces[0]] = queue; - // record blocking queue in SignalHandler - SignalHandler::RegisterBlockingQueue(queue); id_to_grad[block_id] = pieces[0]; } size_t num_blocks = program->Size(); @@ -274,39 +199,37 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, grad_to_prepared_ctx[id_to_grad[block_list[i]]] = optimize_prepared[i]; } - bool exit_flag = false; + request_send_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); + request_get_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); + request_prefetch_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); - VLOG(3) << "start async optimize threads"; - std::vector> fs; - for (auto iter = grad_to_queue.begin(); iter != grad_to_queue.end(); iter++) { - std::string grad_name = iter->first; - VLOG(3) << "create async update thread for " << grad_name; - fs.push_back(framework::AsyncIO([grad_name, &exit_flag, &executor, - &grad_to_queue, &grad_to_prepared_ctx]() { - AsyncUpdateThread(grad_name, exit_flag, grad_to_queue[grad_name], - executor, grad_to_prepared_ctx[grad_name].get()); - })); - } VLOG(3) << "RunAsyncLoop into while"; - while (!exit_flag && !SignalHandler::IsProgramExit()) { - const detail::ReceivedMessage v = rpc_service_->Get(); - auto recv_var_name = v.first; - if (recv_var_name == LISTEN_TERMINATE_MESSAGE) { - LOG(INFO) << "received terminate message and exit"; - exit_flag = true; + while (true) { + if (rpc_service_->IsExit()) { + LOG(INFO) << "get exit!rpc_processor break!"; break; - } else { - VLOG(3) << "received grad: " << recv_var_name; - grad_to_queue[recv_var_name]->Push(v); } - if (exit_flag) { - rpc_service_->ShutDown(); - break; - } + sleep(1); } // while(true) } +static void FillRequestCtx( + detail::RequestHandler *h, framework::Scope *scope, + platform::DeviceContext *dev_ctx, framework::Executor *executor, + framework::ProgramDesc *program, + std::unordered_map> + *prefetch_ctx, + detail::RPCServer *rpc_server) { + h->SetScope(scope); + h->SetDevCtx(dev_ctx); + h->SetExecutor(executor); + h->SetProgram(program); + h->SetPrefetchPreparedCtx(prefetch_ctx); + h->SetRPCServer(rpc_server); +} + void ListenAndServOp::RunImpl(const framework::Scope &scope, const platform::Place &dev_place) const { // Mark this as PS that it should decide profiling by listening from trainer. @@ -316,27 +239,67 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, framework::Scope &recv_scope = scope.NewScope(); bool sync_mode = Attr("sync_mode"); + auto fan_in = Attr("Fanin"); PADDLE_ENFORCE(!rpc_service_); std::string endpoint = Attr("endpoint"); - rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, sync_mode)); + LOG(INFO) << "sync_mode:" << sync_mode << ", fan_in:" << fan_in + << ", end_point:" << endpoint; + + rpc_service_.reset(new RPCSERVER_T(endpoint, fan_in)); + + request_send_handler_.reset(new detail::RequestSendHandler(sync_mode)); + request_get_handler_.reset(new detail::RequestGetHandler(sync_mode)); + request_prefetch_handler_.reset( + new detail::RequestPrefetchHandler(sync_mode)); + + rpc_service_->RegisterRPC(detail::kRequestSend, request_send_handler_.get()); + rpc_service_->RegisterRPC(detail::kRequestGet, request_get_handler_.get()); + rpc_service_->RegisterRPC(detail::kRequestPrefetch, + request_prefetch_handler_.get()); auto *optimize_block = Attr(kOptimizeBlock); - auto *prefetch_block = Attr(kPrefetchBlock); auto *program = optimize_block->Program(); framework::Executor executor(dev_place); - // prepare rpc_service - rpc_service_->SetScope(&recv_scope); - rpc_service_->SetDevCtx(&dev_ctx); - rpc_service_->SetProgram(program); - rpc_service_->SetExecutor(&executor); - // prepare for prefetch - VLOG(3) << "prefetch block id is " << prefetch_block->ID(); - auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID()); - rpc_service_->SetPrefetchPreparedCtx(std::move(prefetch_prepared)); + std::vector prefetch_block_id_list; + std::unordered_map block_id_to_prefetch_var_name; + + auto prefetch_var_name_to_block_id_str = + Attr>(kPrefetchVarNameToBlockId); + for (const auto &prefetch_var_name_and_id : + prefetch_var_name_to_block_id_str) { + std::vector pieces; + split(prefetch_var_name_and_id, ':', &pieces); + VLOG(3) << "after split, prefetch_var = " << pieces[0] + << ", id=" << pieces[1]; + PADDLE_ENFORCE_EQ(pieces.size(), 2); + + int block_id = std::stoi(pieces[1]); + prefetch_block_id_list.push_back(block_id); + block_id_to_prefetch_var_name[block_id] = pieces[0]; + } + + auto prefetch_prepared = executor.Prepare(*program, prefetch_block_id_list); + + std::unordered_map> + prefetch_var_name_to_prepared_ctx; + for (size_t i = 0; i < prefetch_block_id_list.size(); ++i) { + auto block_id = prefetch_block_id_list[i]; + auto prefetch_var_name = block_id_to_prefetch_var_name[block_id]; + prefetch_var_name_to_prepared_ctx[prefetch_var_name] = prefetch_prepared[i]; + } + + auto f = std::bind(FillRequestCtx, std::placeholders::_1, &recv_scope, + &dev_ctx, &executor, program, + &prefetch_var_name_to_prepared_ctx, rpc_service_.get()); + + f(request_send_handler_.get()); + f(request_get_handler_.get()); + f(request_prefetch_handler_.get()); // start the server listening after all member initialized. server_thread_.reset(new std::thread(RunServer, rpc_service_)); @@ -348,11 +311,9 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, signal(SIGTERM, SignalHandler::StopAndExit); // Write to a file of server selected port for python use. - std::string file_path = string::Sprintf("/tmp/paddle.%d.selected_port", - static_cast(::getpid())); SavePort(); if (sync_mode) { - RunSyncLoop(&executor, program, &recv_scope, prefetch_block); + RunSyncLoop(&executor, program, &recv_scope, prefetch_block_id_list); } else { RunAsyncLoop(&executor, program); } @@ -378,34 +339,17 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { AddAttr("sync_mode", "if works at sync_mode or not").SetDefault(true); AddAttr(kOptimizeBlock, "BlockID to run on server side."); - AddAttr(kPrefetchBlock, - "prefetch block to run on server side."); + AddAttr>(kPrefetchVarNameToBlockId, + "prefetch blocks to run on server side.") + .SetDefault({}); AddAttr("Fanin", "How many clients send to this server.") .SetDefault(1); } }; -bool SignalHandler::program_exit_flag_ = false; - -SignalHandler::BlockingQueueSet SignalHandler::blocking_queue_set_{}; - void SignalHandler::StopAndExit(int signal_num) { VLOG(3) << "Catch interrupt signal: " << signal_num << ", program will exit"; - - program_exit_flag_ = true; - - // awake all blocking queues - for (BlockingQueueSet::iterator iter = blocking_queue_set_.begin(); - iter != blocking_queue_set_.end(); iter++) { - iter->get()->Push( - std::make_pair(std::string(LISTEN_TERMINATE_MESSAGE), nullptr)); - } - - exit(EXIT_SUCCESS); -} - -void SignalHandler::RegisterBlockingQueue(BlockingQueue &queue) { - blocking_queue_set_.insert(queue); + exit(0); } } // namespace operators diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 6f868369dc..46c3a19e20 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -18,20 +18,22 @@ limitations under the License. */ #include #include #include +#include #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/detail/rpc_server.h" namespace paddle { namespace operators { constexpr char kOptimizeBlock[] = "OptimizeBlock"; -constexpr char kPrefetchBlock[] = "PrefetchBlock"; +constexpr char kPrefetchVarNameToBlockId[] = "prefetch_var_name_to_block_id"; -void RunServer(std::shared_ptr service); +void RunServer(std::shared_ptr service); class ListenAndServOp : public framework::OperatorBase { public: @@ -45,48 +47,34 @@ class ListenAndServOp : public framework::OperatorBase { void RunSyncLoop(framework::Executor* executor, framework::ProgramDesc* program, framework::Scope* recv_scope, - framework::BlockDesc* prefetch_block) const; + const std::vector& prefetch_block_id_list) const; void RunAsyncLoop(framework::Executor* executor, framework::ProgramDesc* program) const; void SavePort() const; - void WaitServerReady(); - - int GetSelectedPort() { return selected_port_; } + int GetSelectedPort() { return rpc_service_->GetSelectedPort(); } void Stop() override; void RunImpl(const framework::Scope& scope, const platform::Place& dev_place) const override; - static void ResetPort() { selected_port_ = 0; } - protected: - mutable std::shared_ptr rpc_service_; + mutable std::shared_ptr rpc_service_; + mutable std::shared_ptr request_send_handler_; + mutable std::shared_ptr request_get_handler_; + mutable std::shared_ptr request_prefetch_handler_; + mutable std::shared_ptr server_thread_; - // FIXME(wuyi): it's static so that the operator can be cloned. - static std::atomic_int selected_port_; }; class SignalHandler { - public: - typedef std::shared_ptr BlockingQueue; - typedef std::unordered_set BlockingQueueSet; - public: static void StopAndExit(int signal_num); - static void RegisterBlockingQueue(BlockingQueue&); - - static inline bool IsProgramExit() { return program_exit_flag_; } - private: - static bool program_exit_flag_; - - static BlockingQueueSet blocking_queue_set_; - DISABLE_COPY_AND_ASSIGN(SignalHandler); }; diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc index 93f45cff8a..8f4b504927 100644 --- a/paddle/fluid/operators/load_op.cc +++ b/paddle/fluid/operators/load_op.cc @@ -74,25 +74,18 @@ class LoadOp : public framework::OperatorBase { class LoadOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddOutput("Out", "(Tensor) The tensor need to be loaded"); + AddOutput("Out", "The tensor need to be loaded"); AddAttr( "load_as_fp16", - "(boolean, default false)" "If true, the tensor will be first loaded and then " "converted to float16 data type. Otherwise, the tensor will be " - "directly loaded without data type conversion.") + "directly loaded without data type conversion. Default is false.") .SetDefault(false); AddAttr("file_path", - "(string) " - "Variable will be loaded from \"file_path\".") + R"(Variable will be loaded from "file_path")") .AddCustomChecker( [](const std::string &path) { return !path.empty(); }); - AddComment(R"DOC( -Load Operator. - -Load operator will load a tensor variable from disk file. - -)DOC"); + AddComment("Load operator will load a tensor variable from disk file."); } }; } // namespace operators diff --git a/paddle/fluid/operators/lrn_op.cc b/paddle/fluid/operators/lrn_op.cc index 52b9cd7fb7..52b459a6a2 100644 --- a/paddle/fluid/operators/lrn_op.cc +++ b/paddle/fluid/operators/lrn_op.cc @@ -124,16 +124,17 @@ namespace { framework::OpKernelType GetExpectedLRNKernel( const framework::ExecutionContext& ctx) { framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); #ifdef PADDLE_WITH_MKLDNN if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), layout_, library_); diff --git a/paddle/fluid/operators/math/blas.h b/paddle/fluid/operators/math/blas.h index 1a37cb39d5..6207d14ecd 100644 --- a/paddle/fluid/operators/math/blas.h +++ b/paddle/fluid/operators/math/blas.h @@ -20,13 +20,16 @@ #ifdef PADDLE_WITH_MKLML #include #include +#include #include #endif #ifdef PADDLE_USE_OPENBLAS #include +#ifdef LAPACK_FOUND #include #endif +#endif #ifndef LAPACK_FOUND extern "C" { @@ -46,6 +49,18 @@ namespace paddle { namespace operators { namespace math { +static void SetNumThreads(int num_threads) { +#ifdef PADDLE_USE_OPENBLAS + int real_num_threads = num_threads > 1 ? num_threads : 1; + openblas_set_num_threads(real_num_threads); +#elif defined(PADDLE_WITH_MKLML) + int real_num_threads = num_threads > 1 ? num_threads : 1; + mkl_set_num_threads(real_num_threads); +#else + PADDLE_ENFORCE(false, "To be implemented."); +#endif +} + /** * Matrix Descriptor of a memory buffer. * diff --git a/paddle/fluid/operators/math/math_function.h b/paddle/fluid/operators/math/math_function.h index d4b0e17ed4..8b296b6a07 100644 --- a/paddle/fluid/operators/math/math_function.h +++ b/paddle/fluid/operators/math/math_function.h @@ -21,8 +21,10 @@ limitations under the License. */ #ifdef PADDLE_USE_OPENBLAS #include +#ifdef LAPACK_FOUND #include #endif +#endif #ifndef LAPACK_FOUND extern "C" { diff --git a/paddle/fluid/operators/math/math_function_test.cc b/paddle/fluid/operators/math/math_function_test.cc index 3719a264e9..b545671b43 100644 --- a/paddle/fluid/operators/math/math_function_test.cc +++ b/paddle/fluid/operators/math/math_function_test.cc @@ -77,6 +77,8 @@ TEST(math_function, gemm_trans_clbas) { paddle::platform::CPUDeviceContext context(*cpu_place); GetBlas(context).GEMM(false, true, m, n, k, 1, input1_ptr, 3, input2_ptr + 3, 3, 1, input3_ptr + 1, 4); + delete cpu_place; + cpu_place = NULL; EXPECT_EQ(input3_ptr[0], 0); EXPECT_EQ(input3_ptr[1], 24); diff --git a/paddle/fluid/operators/max_sequence_len_op.cc b/paddle/fluid/operators/max_sequence_len_op.cc index 8e508b68ee..b1e69f375d 100644 --- a/paddle/fluid/operators/max_sequence_len_op.cc +++ b/paddle/fluid/operators/max_sequence_len_op.cc @@ -42,10 +42,15 @@ class MaxSeqenceLenOp : public framework::OperatorBase { class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("RankTable", "The lod_rank_table."); - AddOutput("Out", "The max sequence length."); - AddComment( - R"DOC(Calculate the max sequence length through lod_rank_table.)DOC"); + AddInput("RankTable", "Input variable which is a LoDRankTable object"); + AddOutput("Out", "The max sequence length"); + AddComment(R"DOC( + Given a LoDRankTable object, this layer returns the max length of + a batch of sequences. In fact, a LoDRankTable object contains a list of + tuples() and the list is already sorted by + sequence length in descending order, so the operator just returns the + sequence length of the first tuple element +)DOC"); } }; diff --git a/paddle/fluid/operators/mean_iou_op.cc b/paddle/fluid/operators/mean_iou_op.cc new file mode 100644 index 0000000000..a60f245f53 --- /dev/null +++ b/paddle/fluid/operators/mean_iou_op.cc @@ -0,0 +1,110 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/mean_iou_op.h" + +namespace paddle { +namespace operators { + +class MeanIoUOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Predictions"), + "Input (Predictions) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Labels"), + "Input (labels) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("OutMeanIou"), + "Output (OutMeanIou) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("OutWrong"), + "Output (OutWrong) of MeanIoU op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("OutCorrect"), + "Output (OutWrong) of MeanIoU op should not be null."); + + int64_t num_classes = + static_cast(ctx->Attrs().Get("num_classes")); + + ctx->SetOutputDim("OutMeanIou", {1}); + ctx->SetOutputDim("OutWrong", {num_classes}); + ctx->SetOutputDim("OutCorrect", {num_classes}); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Predictions")->type()), + ctx.GetPlace()); + } +}; + +class MeanIoUOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Predictions", + "(Tensor), A Tensor of prediction results for semantic labels" + " with type int32 or int64. The rank should be greater than 1."); + AddInput( + "Labels", + "(Tensor), A Tensor of ground truth labels with type int32 or int64." + "Its shape should be the same as Input(Predictions)."); + AddInput("InWrongs", + "(vector), A list of Tensor with shape " + "[num_classes]. They are used to collect wrong number among " + "batches. Empty list is also valid here.") + .AsDuplicable() + .AsDispensable(); + AddInput( + "InCorrects", + "(vector), A list of Tensor with shape " + "[num_classes]. They are used to collect correct number among batches. " + "Empty list is also valid here.") + .AsDuplicable() + .AsDispensable(); + AddInput("InMeanIou", + "(vector), A list of Tensor that Output(mean_iou) should " + "be added to. Empty list is also valid here.") + .AsDuplicable() + .AsDispensable(); + AddOutput("OutMeanIou", + "(vector), A Tensor representing the" + " mean intersection-over-union with shape [1]."); + AddOutput("OutWrong", "(Tensor), A Tensor with shape [num_classes]. "); + AddOutput("OutCorrect", "(Tensor), A Tensor with shape [num_classes]. "); + AddAttr("num_classes", "(int), The possible number of labels."); + + AddComment(R"DOC( +mean-IOU Operator. +Mean Intersection-Over-Union is a common evaluation metric for +semantic image segmentation, which first computes the IOU for each +semantic class and then computes the average over classes. +IOU is defined as follows: + IOU = true_positive / (true_positive + false_positive + false_negative). +It is based on pixel level area while "IOU Similarity Operator" +is based on area of rectangle. + +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(mean_iou, ops::MeanIoUOp, ops::MeanIoUOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(mean_iou, ops::MeanIoUKernel, + ops::MeanIoUKernel, + ops::MeanIoUKernel); diff --git a/paddle/fluid/operators/mean_iou_op.cu b/paddle/fluid/operators/mean_iou_op.cu new file mode 100644 index 0000000000..83bb4dde46 --- /dev/null +++ b/paddle/fluid/operators/mean_iou_op.cu @@ -0,0 +1,164 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/mean_iou_op.h" +#include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/gpu_info.h" + +namespace paddle { +namespace operators { + +using platform::PADDLE_CUDA_NUM_THREADS; + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__global__ void CountCUDAKernel(const int num_classes, const int count, + const T* predictions, const T* labels, + int* wrong, int* correct) { + extern __shared__ int blcok_cache[]; + int* wrong_c = blcok_cache; + int* correct_c = blcok_cache + num_classes; + // init cache + for (int i = threadIdx.x; i < num_classes * 2; i += blockDim.x) { + blcok_cache[i] = 0; + } + __syncthreads(); + + T pred; + T label; + CUDA_1D_KERNEL_LOOP(i, count) { + pred = predictions[i]; + label = labels[i]; + if (pred == label) { + atomicAdd(correct_c + pred, 1); + } else { + atomicAdd(wrong_c + pred, 1); + atomicAdd(wrong_c + label, 1); + } + } + + __syncthreads(); + + for (int i = threadIdx.x; i < num_classes; i += blockDim.x) { + atomicAdd(wrong + i, wrong_c[i]); + atomicAdd(correct + i, correct_c[i]); + } +} + +__global__ void ComputeIoUCUDAKernel(const int num_classes, int* wrong, + int* correct, float* ious, float* iou) { + __shared__ int valid_count_c; + if (threadIdx.x == 0) { + valid_count_c = 0; + } + __syncthreads(); + CUDA_1D_KERNEL_LOOP(i, num_classes) { + int wrong_n = wrong[i]; + int correct_n = correct[i]; + int denominator = wrong_n + correct_n; + if (denominator > 0) { + atomicAdd(&valid_count_c, 1); + ious[i] = static_cast(correct_n) / denominator; + } else { + ious[i] = 0; + } + } + __syncthreads(); + if (threadIdx.x == 0) { + float iou_sum = 0; + for (int i = 0; i < num_classes; ++i) { + iou_sum += ious[i]; + } + iou[0] += iou_sum / valid_count_c; + } +} + +template +class MeanIoUCUDAOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& place = *ctx.template device_context() + .eigen_device(); + // get input and output tensor + auto* predictions = ctx.Input("Predictions"); + auto* labels = ctx.Input("Labels"); + auto* out_mean_iou = ctx.Output("OutMeanIou"); + auto* out_wrong = ctx.Output("OutWrong"); + auto* out_correct = ctx.Output("OutCorrect"); + int num_classes = static_cast(ctx.Attr("num_classes")); + + // Get data ptr + const T* predictions_data = predictions->data(); + const T* labels_data = labels->data(); + int* out_wrong_data = out_wrong->mutable_data(ctx.GetPlace()); + int* out_correct_data = out_correct->mutable_data(ctx.GetPlace()); + float* out_mean_iou_data = + out_mean_iou->mutable_data(ctx.GetPlace()); + + // Get Eigen tensor + auto out_mean_iou_t = EigenTensor::From(*out_mean_iou); + auto out_wrong_t = EigenTensor::From(*out_wrong); + auto out_correct_t = EigenTensor::From(*out_correct); + + // Temporary tensor + Tensor ious; + float* ious_data = ious.mutable_data( + {static_cast(num_classes)}, ctx.GetPlace()); + auto ious_t = EigenTensor::From(ious); + + // Init out_wrong, out_correct and out_mean_iou + out_wrong_t.device(place) = out_wrong_t.constant(0); + out_correct_t.device(place) = out_correct_t.constant(0); + out_mean_iou_t.device(place) = out_mean_iou_t.constant(0.0f); + + // collect pre wrong, correct and mean_iou + auto in_mean_ious = ctx.MultiInput("InMeanIou"); + for (int i = 0; i < in_mean_ious.size(); ++i) { + out_mean_iou_t.device(place) += + EigenTensor::From(*in_mean_ious[i]); + } + auto in_wrongs = ctx.MultiInput("InWrongs"); + for (int i = 0; i < in_wrongs.size(); ++i) { + out_wrong_t.device(place) += EigenTensor::From(*in_wrongs[i]); + } + auto in_corrects = ctx.MultiInput("InCorrects"); + for (int i = 0; i < in_corrects.size(); ++i) { + out_correct_t.device(place) += EigenTensor::From(*in_corrects[i]); + } + // compute + auto stream = ctx.cuda_device_context().stream(); + int block = PADDLE_CUDA_NUM_THREADS; + int grid = (predictions->numel() + block - 1) / block; + int cache_size = (num_classes * 2 + 1) * sizeof(int); + CountCUDAKernel<<>>( + num_classes, predictions->numel(), predictions_data, labels_data, + out_wrong_data, out_correct_data); + ctx.device_context().Wait(); + ComputeIoUCUDAKernel<<<1, block, 0, stream>>>(num_classes, out_wrong_data, + out_correct_data, ious_data, + out_mean_iou_data); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(mean_iou, ops::MeanIoUCUDAOpKernel, + ops::MeanIoUCUDAOpKernel, + ops::MeanIoUCUDAOpKernel); diff --git a/paddle/fluid/operators/mean_iou_op.h b/paddle/fluid/operators/mean_iou_op.h new file mode 100644 index 0000000000..9fa00e60e0 --- /dev/null +++ b/paddle/fluid/operators/mean_iou_op.h @@ -0,0 +1,117 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +using Tensor = framework::Tensor; + +template +using EigenTensor = framework::EigenTensor; + +template +class MeanIoUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& place = *ctx.template device_context() + .eigen_device(); + // get input and output tensor + auto* predictions = ctx.Input("Predictions"); + auto* labels = ctx.Input("Labels"); + auto* out_mean_iou = ctx.Output("OutMeanIou"); + auto* out_wrong = ctx.Output("OutWrong"); + auto* out_correct = ctx.Output("OutCorrect"); + int num_classes = static_cast(ctx.Attr("num_classes")); + + // get data ptr + const T* predictions_data = predictions->data(); + const T* labels_data = labels->data(); + float* out_mean_iou_data = + out_mean_iou->mutable_data(ctx.GetPlace()); + int* out_wrong_data = out_wrong->mutable_data(ctx.GetPlace()); + int* out_correct_data = out_correct->mutable_data(ctx.GetPlace()); + + // get eigen tensor + auto out_mean_iou_t = EigenTensor::From(*out_mean_iou); + auto out_wrong_t = EigenTensor::From(*out_wrong); + auto out_correct_t = EigenTensor::From(*out_correct); + + // Tmp tensor + Tensor denominator; + Tensor valid_count; + Tensor iou_sum; + + // get data ptr of tmp tensor + int* denominator_data = denominator.mutable_data( + {static_cast(num_classes)}, ctx.GetPlace()); + int* valid_count_data = valid_count.mutable_data({1}, ctx.GetPlace()); + float* iou_sum_data = iou_sum.mutable_data({1}, ctx.GetPlace()); + + // get eigen tensor of tmp tensor + auto denominator_t = EigenTensor::From(denominator); + auto valid_count_t = EigenTensor::From(valid_count); + auto iou_sum_t = EigenTensor::From(iou_sum); + + // init out_wrong, out_correct and out_mean_iou + out_wrong_t = out_wrong_t.constant(0); + out_correct_t = out_correct_t.constant(0); + out_mean_iou_t = out_mean_iou_t.constant(0); + + // collect pre wrong, correct and mean_iou + auto in_mean_ious = ctx.MultiInput("InMeanIou"); + for (size_t i = 0; i < in_mean_ious.size(); ++i) { + out_mean_iou_t.device(place) += + EigenTensor::From(*in_mean_ious[i]); + } + auto in_wrongs = ctx.MultiInput("InWrongs"); + for (size_t i = 0; i < in_wrongs.size(); ++i) { + out_wrong_t.device(place) += EigenTensor::From(*in_wrongs[i]); + } + auto in_corrects = ctx.MultiInput("InCorrects"); + for (size_t i = 0; i < in_corrects.size(); ++i) { + out_correct_t.device(place) += EigenTensor::From(*in_corrects[i]); + } + + // compute + for (int64_t i = 0; i < predictions->numel(); ++i) { + if (predictions_data[i] == labels_data[i]) { + out_correct_data[predictions_data[i]] += 1; + } else { + out_wrong_data[labels_data[i]] += 1; + out_wrong_data[predictions_data[i]] += 1; + } + } + + denominator_t = out_wrong_t + out_correct_t; + valid_count_t = + (denominator_t > denominator_t.constant(0.0f)).cast().sum(); + + for (int i = 0; i < num_classes; ++i) { + if (denominator_data[i] == 0) { + denominator_data[i] = 1; + } + } + + iou_sum_t = + (out_correct_t.cast() / denominator_t.cast()).sum(); + out_mean_iou_data[0] += (iou_sum_data[0] / valid_count_data[0]); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/mean_op.cc b/paddle/fluid/operators/mean_op.cc index 74477eb439..4881cff4a3 100644 --- a/paddle/fluid/operators/mean_op.cc +++ b/paddle/fluid/operators/mean_op.cc @@ -34,7 +34,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "The input of mean op"); - AddOutput("Out", "The output of mean op"); + AddOutput("Out", "The output of mean op").Reuse("X"); AddComment(R"DOC( Mean Operator. diff --git a/paddle/fluid/operators/merge_ids_op.cc b/paddle/fluid/operators/merge_ids_op.cc new file mode 100644 index 0000000000..c6ec4ab047 --- /dev/null +++ b/paddle/fluid/operators/merge_ids_op.cc @@ -0,0 +1,128 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/merge_ids_op.h" + +namespace paddle { +namespace operators { + +class MergeIdsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); + AddInput( + "X", + "(LoDTensors) multi input tensor with shape{batch_num, N}, N is the " + "size of embedding table") + .AsDuplicable(); + AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors."); + + AddComment(R"DOC( +Merge multi LoDTensor's into one according to Ids's shard num. + + +split_ids_op -> prefetch_op -> merge_ids_op + + +merge_ids_op should be used after split_ids_op and prefetch_op, split_ids_op + will split input Ids into multiple tensors according to Id's shard number. +prefetch_op will send them to parameter server to prefetch embedding value +back. During split, the order of ids is disordered. In merge_ids_op we use +the original Ids to restore the order of the fetched embedding value and + also pass the lod information to the merged output. + + +Example: + + Ids = [1,2,3,4,5,6] # 3 shared + +split_ids_op -> + + Id0 = [3, 6] # id % 3 == 0 + Id1 = [1, 4] # id % 3 == 1 + Id2 = [2, 5] # id % 3 == 2 + +prefetch_op -> + + X0 = [[0.3 0.3] # 3 + [0.6 0.6]] # 6 + X1 = [[0.1 0.1] # 1 + [0.4 0.4]] # 4 + X2 = [[0.2 0.2] # 2 + [0.5 0.5]] # 5 + +merge_ids_op -> + + Out = [[0.1 0.1] # 1 + [0.2 0.2] # 2 + [0.3 0.3] # 3 + [0.4 0.4] # 4 + [0.5 0.5] # 5 + [0.6 0.6]] # 6 +)DOC"); + } +}; + +class MergeIdsOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Ids"), "MergeIdsOp must has input Ids."); + PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has input X."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "MergeIdsOp must has output Out."); + + auto ids_var_type = ctx->GetInputsVarType("Ids").front(); + auto ids_dims = ctx->GetInputDim("Ids"); + if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { + PADDLE_ENFORCE_EQ(ids_dims.size(), 2); + PADDLE_ENFORCE_EQ(ids_dims[1], 1); + } + auto x_var_type = ctx->GetInputsVarType("X"); + for (auto &var_type : x_var_type) { + PADDLE_ENFORCE_EQ(var_type, framework::proto::VarType::LOD_TENSOR, + "input X only support lod tensors"); + } + ctx->ShareLoD("Ids", "Out"); + } + + private: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType( + ctx.MultiInput("X").front()->type()), + ctx.GetPlace()); + } +}; + +class MergeIdsOpInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + auto *input_var = block->Var(op_desc.Input("Ids")[0]); + for (auto &out_var : op_desc.Output("Out")) { + block->Var(out_var)->SetType(input_var->GetType()); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(merge_ids, ops::MergeIdsOp, ops::MergeIdsOpMaker, + ops::MergeIdsOpInferVarType); +REGISTER_OP_CPU_KERNEL( + merge_ids, ops::MergeIdsOpKernel); diff --git a/paddle/fluid/operators/merge_ids_op.h b/paddle/fluid/operators/merge_ids_op.h new file mode 100644 index 0000000000..83712a8519 --- /dev/null +++ b/paddle/fluid/operators/merge_ids_op.h @@ -0,0 +1,92 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" + +namespace paddle { +namespace operators { + +template +class MergeIdsOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto place = ctx.GetPlace(); + if (!platform::is_cpu_place(place)) { + PADDLE_THROW("MergeIds do not support GPU kernel"); + } + VLOG(3) << "run in MergeIdsOpKernel"; + + const auto *ids_var = ctx.InputVar("Ids"); + PADDLE_ENFORCE(ids_var->IsType(), + "only support to merge Ids of LoDTensor"); + + const auto &ids_tensor = ids_var->Get(); + const auto &ids_dims = ids_tensor.dims(); + const int64_t *ids = ids_tensor.data(); + + auto x_tensors = ctx.MultiInput("X"); + + auto *out = ctx.Output("Out"); + + int batch_size = 0; + int embedding_size = 0; + for (auto &input : x_tensors) { + if (framework::product(input->dims()) != 0) { + if (embedding_size == 0) { + embedding_size = input->dims()[1]; + } + PADDLE_ENFORCE_EQ(embedding_size, input->dims()[1], + "embedding size of all input should be the same"); + batch_size += input->dims()[0]; + } + } + PADDLE_ENFORCE_EQ( + batch_size, ids_dims[0], + "the batch size of ids and merged embedding value should be the same"); + + const size_t shard_num = x_tensors.size(); + + if (shard_num == 1) { + VLOG(3) << "only one shard, we can copy the data directly"; + TensorCopy(*x_tensors[0], place, out); + } else { + std::vector in_indexs(shard_num, 0); + auto *out_data = out->mutable_data( + framework::make_ddim({batch_size, embedding_size}), place); + // copy data from ins[shard_num] to out. + for (int i = 0; i < ids_dims[0]; ++i) { + int64_t id = ids[i]; + size_t shard_id = static_cast(id) % shard_num; + int index = in_indexs[shard_id]; + memcpy(out_data + embedding_size * i, + x_tensors[shard_id]->data() + index * embedding_size, + sizeof(T) * embedding_size); + in_indexs[shard_id] += 1; + } + + for (size_t i = 0; i < shard_num; ++i) { + PADDLE_ENFORCE_EQ(in_indexs[i], x_tensors[i]->dims()[0], + "after merge, all data in x_tensor should be used"); + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/norm_op.cc b/paddle/fluid/operators/norm_op.cc index cdbc975c02..aa19c62c83 100644 --- a/paddle/fluid/operators/norm_op.cc +++ b/paddle/fluid/operators/norm_op.cc @@ -16,40 +16,34 @@ limitations under the License. */ namespace paddle { namespace operators { -template class NormOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput( - "X", - "(Tensor) The input tensor of norm operator. " - "The format of input tensor is NCHW. Where N is batch size, C is the " - "number of channels, H and W is the height and width of feature."); - AddInput("Scale", - "(Tensor) The input tensor of norm operator. " - "The format of input tensor is C * 1."); - AddAttr("epsilon", - "(float, default 1e-10) Constant " - "for numerical stability.") + AddInput("X", "(Tensor) A tensor of rank >= axis."); + AddAttr("axis", + "The axis on which to apply normalization. If axis < 0, " + "the dimension to normalization is rank(X) + axis. -1 is " + "the last dimension."); + AddAttr("epsilon", + "(float, default 1e-10) The epsilon value is used " + "to avoid division by zero.") .SetDefault(1.0e-10f); - AddOutput("Out", - "(Tensor) The output tensor of norm operator." - "N * M." - "M = C * H * W"); + AddOutput("Norm", + "(Tensor) A tensor saved the `sqrt(sum(x) + epsion)` will " + "be used in backward kernel.") + .AsIntermediate(); + AddOutput("Out", "(Tensor) A tensor of the same shape as X."); AddComment(R"DOC( - "Input shape: $(N, C, H, W)$ - Scale shape: $(C, 1)$ - Output shape: $(N, C, H, W)$ - Where - forward - $$ - [\frac {x_{1}}{\sqrt{\sum{x_{i}^{2}}}} \frac {x_{2}}{\sqrt{\sum{x_{i}^{2}}}} \frac {x_{3}}{\sqrt{\sum{x_{i}^{2}}}} \cdot \cdot \cdot \frac {x_{n}}{\sqrt{\sum{x_{i}^{2}}}}] - $$ - backward - $$ - \frac{\frac{\mathrm{d}L }{\mathrm{d}y_{1}} - \frac {x_{1}\sum {\frac{\mathrm{d} L}{\mathrm{d} y_{j}}}x_{j}}{\sum x_{j}^{2}} }{\sqrt{\sum{x_{j}^{2}}}} - $$ - )DOC"); + +Given a tensor, apply 2-normalization along the provided axis. + +$$ +y = \frac{x}{ \sqrt{\sum {x^2} + epsion }} +$$ + +where, $\sum {x^2}$ is calculated along the `axis` dimension. + +)DOC"); } }; @@ -58,15 +52,15 @@ class NormOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of NormOp" - "should not be null."); - PADDLE_ENFORCE(ctx->HasInput("Scale"), - "Input(Scale) of NormOp" - "should not be null."); + "Input(X) of NormOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of NormOp should not be null."); - auto in_x_dims = ctx->GetInputDim("X"); - ctx->SetOutputDim("Out", in_x_dims); + auto xdim = ctx->GetInputDim("X"); + ctx->SetOutputDim("Out", xdim); + int axis = ctx->Attrs().Get("axis"); + if (axis < 0) axis = xdim.size() + axis; + xdim[axis] = 1; + ctx->SetOutputDim("Norm", xdim); } }; @@ -84,12 +78,12 @@ class NormOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(norm, ops::NormOp, ops::NormOpMaker, +using CPU = paddle::platform::CPUDeviceContext; + +REGISTER_OPERATOR(norm, ops::NormOp, ops::NormOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(norm_grad, ops::NormOpGrad); -REGISTER_OP_CPU_KERNEL( - norm, ops::NormKernel, - ops::NormKernel); -REGISTER_OP_CPU_KERNEL( - norm_grad, ops::NormGradKernel, - ops::NormGradKernel); +REGISTER_OP_CPU_KERNEL(norm, ops::NormKernel, + ops::NormKernel); +REGISTER_OP_CPU_KERNEL(norm_grad, ops::NormGradKernel, + ops::NormGradKernel); diff --git a/paddle/fluid/operators/norm_op.cu b/paddle/fluid/operators/norm_op.cu index d1d9be5074..1d0021d33f 100644 --- a/paddle/fluid/operators/norm_op.cu +++ b/paddle/fluid/operators/norm_op.cu @@ -16,9 +16,9 @@ limitations under the License. */ #include "paddle/fluid/operators/norm_op.h" namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL( - norm, ops::NormKernel, - ops::NormKernel); -REGISTER_OP_CUDA_KERNEL( - norm_grad, ops::NormGradKernel, - ops::NormGradKernel); +using CUDA = paddle::platform::CUDADeviceContext; + +REGISTER_OP_CUDA_KERNEL(norm, ops::NormKernel, + ops::NormKernel); +REGISTER_OP_CUDA_KERNEL(norm_grad, ops::NormGradKernel, + ops::NormGradKernel); diff --git a/paddle/fluid/operators/norm_op.h b/paddle/fluid/operators/norm_op.h index 0ad29e8a03..3167bdc8ac 100644 --- a/paddle/fluid/operators/norm_op.h +++ b/paddle/fluid/operators/norm_op.h @@ -19,156 +19,110 @@ limitations under the License. */ namespace paddle { namespace operators { -template +inline void GetDims(const framework::DDim& dim, int axis, int* pre, int* n, + int* post) { + *pre = 1; + *post = 1; + *n = dim[axis]; + for (int i = 0; i < axis; ++i) { + (*pre) *= dim[i]; + } + for (int i = axis + 1; i < dim.size(); ++i) { + (*post) *= dim[i]; + } +} + +template class NormKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - const framework::Tensor* in_x = context.Input("X"); - const framework::Tensor* scale = context.Input("Scale"); - auto* out = context.Output("Out"); - auto epsilon = static_cast(context.Attr("epsilon")); - out->mutable_data(context.GetPlace()); - int batch_size = in_x->dims()[0]; - int channels = in_x->dims()[1]; - int height = in_x->dims()[2]; - int width = in_x->dims()[3]; - int fea_len = height * width; - auto* place = - context.template device_context().eigen_device(); - auto x = - framework::EigenMatrix::From( - *in_x, framework::make_ddim({batch_size, fea_len * channels})); - // get square - framework::Tensor x_square; - x_square.mutable_data(in_x->dims(), context.GetPlace()); - auto x_square_eigen = - framework::EigenMatrix::From( - x_square, framework::make_ddim({batch_size, fea_len * channels})); - x_square_eigen.device(*place) = x.square(); - auto scale_eigen = - framework::EigenVector::Flatten( - *scale); - for (int n = 0; n < batch_size; ++n) { - framework::Tensor in_x_batch = in_x->Slice(n, n + 1); - auto in_x_batch_eigen = - framework::EigenMatrix::From( - in_x_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor x_square_batch = x_square.Slice(n, n + 1); - auto x_square_batch_eigen = - framework::EigenMatrix::From( - x_square_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor out_batch = out->Slice(n, n + 1); - auto out_batch_eigen = - framework::EigenMatrix::From( - out_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor tmp_tensor; - tmp_tensor.mutable_data(framework::make_ddim({1, fea_len}), - context.GetPlace()); - auto tmp = framework::EigenVector::Flatten(tmp_tensor); - // get colsum and sqrt , inverse - auto dim = Eigen::array({{0}}); - tmp.device(*place) = x_square_batch_eigen.sum(dim); - tmp.device(*place) = (tmp + epsilon).sqrt().inverse(); - Eigen::array broadcast_dim_col; - broadcast_dim_col[1] = 1; - broadcast_dim_col[0] = channels; - out_batch_eigen.device(*place) = - in_x_batch_eigen * (tmp.broadcast(broadcast_dim_col)); - Eigen::array broadcast_dim_row; - broadcast_dim_row[1] = fea_len; - broadcast_dim_row[0] = 1; - out_batch_eigen.device(*place) = - out_batch_eigen * (scale_eigen.broadcast(broadcast_dim_row)); - } + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in_x = ctx.Input("X"); + auto* out_y = ctx.Output("Out"); + auto* out_norm = ctx.Output("Norm"); + out_y->mutable_data(ctx.GetPlace()); + out_norm->mutable_data(ctx.GetPlace()); + + auto xdim = in_x->dims(); + auto ndim = out_norm->dims(); + T eps = static_cast(ctx.Attr("epsilon")); + int axis = ctx.Attr("axis"); + if (axis < 0) axis = xdim.size() + axis; + int pre, n, post; + GetDims(xdim, axis, &pre, &n, &post); + + auto* place = ctx.template device_context().eigen_device(); + + Eigen::DSizes shape(pre, n, post); + Eigen::DSizes norm_shape(pre, post); + + auto x_e = framework::EigenVector::Flatten(*in_x); + auto y_e = framework::EigenVector::Flatten(*out_y); + auto norm_e = framework::EigenVector::Flatten(*out_norm); + auto x = x_e.reshape(shape); + auto y = y_e.reshape(shape); + auto norm = norm_e.reshape(norm_shape); + + Eigen::DSizes rdim(1); + // y = x / sqrt((sum(x * x) + epsilon)) + // norm = sqrt(sum(x * x) + epsilon) + auto sum = x.pow(2).sum(rdim) + eps; + norm.device(*place) = sum.sqrt(); + // y = x / norm + Eigen::DSizes rshape(pre, 1, post); + Eigen::DSizes bcast(1, n, 1); + y.device(*place) = x / norm.reshape(rshape).broadcast(bcast); } }; template class NormGradKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - const framework::Tensor* in_x = context.Input("X"); - const framework::Tensor* scale = context.Input("Scale"); - const framework::Tensor* out_grad = - context.Input(framework::GradVarName("Out")); - auto epsilon = static_cast(context.Attr("epsilon")); - framework::Tensor* in_x_grad = - context.Output(framework::GradVarName("X")); - in_x_grad->mutable_data(context.GetPlace()); - int batch_size = in_x->dims()[0]; - int channels = in_x->dims()[1]; - int height = in_x->dims()[2]; - int width = in_x->dims()[3]; - int fea_len = height * width; - auto* place = - context.template device_context().eigen_device(); - - auto scale_eigen = - framework::EigenVector::Flatten( - *scale); - auto x = - framework::EigenMatrix::From( - *in_x, framework::make_ddim({batch_size, fea_len * channels})); - // get square - framework::Tensor x_square; - x_square.mutable_data(in_x->dims(), context.GetPlace()); - auto x_square_eigen = - framework::EigenMatrix::From( - x_square, framework::make_ddim({batch_size, fea_len * channels})); - x_square_eigen.device(*place) = x.square(); - - for (int n = 0; n < batch_size; ++n) { - framework::Tensor in_x_batch = in_x->Slice(n, n + 1); - auto in_x_batch_eigen = - framework::EigenMatrix::From( - in_x_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor in_g_batch = in_x_grad->Slice(n, n + 1); - auto in_g_batch_eigen = - framework::EigenMatrix::From( - in_g_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor x_square_batch = x_square.Slice(n, n + 1); - auto x_square_batch_eigen = - framework::EigenMatrix::From( - x_square_batch, framework::make_ddim({channels, fea_len})); - framework::Tensor outg_batch = out_grad->Slice(n, n + 1); - auto outg_batch_eigen = - framework::EigenMatrix::From( - outg_batch, framework::make_ddim({channels, fea_len})); - - framework::Tensor tmp_tensor; - tmp_tensor.mutable_data(framework::make_ddim({1, fea_len}), - context.GetPlace()); - auto tmp_eigen = - framework::EigenVector::Flatten(tmp_tensor); - auto dim = Eigen::array({{0}}); - tmp_eigen.device(*place) = (in_x_batch_eigen * outg_batch_eigen).sum(dim); - framework::Tensor norm_tmp_tensor; - norm_tmp_tensor.mutable_data(framework::make_ddim({1, fea_len}), - context.GetPlace()); - auto norm_tmp_eigen = - framework::EigenVector::Flatten(norm_tmp_tensor); - norm_tmp_eigen.device(*place) = - (x_square_batch_eigen.sum(dim) + epsilon).sqrt(); - Eigen::array broadcast_dim_col; - broadcast_dim_col[1] = 1; - broadcast_dim_col[0] = channels; - in_g_batch_eigen.device(*place) = - in_x_batch_eigen * tmp_eigen.broadcast(broadcast_dim_col); - in_g_batch_eigen.device(*place) = - in_g_batch_eigen / - (norm_tmp_eigen * norm_tmp_eigen).broadcast(broadcast_dim_col); - in_g_batch_eigen.device(*place) = outg_batch_eigen - in_g_batch_eigen; - // outg_batch_eigen + (in_g_batch_eigen * -1); - in_g_batch_eigen.device(*place) = - in_g_batch_eigen / norm_tmp_eigen.broadcast(broadcast_dim_col); - Eigen::array broadcast_dim_row; - broadcast_dim_row[1] = fea_len; - broadcast_dim_row[0] = 1; - in_g_batch_eigen.device(*place) = - in_g_batch_eigen * (scale_eigen.broadcast(broadcast_dim_row)); - } + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in_x = ctx.Input("X"); + auto* in_norm = ctx.Input("Norm"); + auto* in_dy = ctx.Input(framework::GradVarName("Out")); + auto* out_dx = ctx.Output(framework::GradVarName("X")); + out_dx->mutable_data(ctx.GetPlace()); + + auto xdim = in_x->dims(); + int axis = ctx.Attr("axis"); + if (axis < 0) axis = xdim.size() + axis; + int pre, n, post; + GetDims(xdim, axis, &pre, &n, &post); + + auto* place = ctx.template device_context().eigen_device(); + + auto x_e = framework::EigenVector::Flatten(*in_x); + auto dy_e = framework::EigenVector::Flatten(*in_dy); + auto norm_e = framework::EigenVector::Flatten(*in_norm); + auto dx_e = framework::EigenVector::Flatten(*out_dx); + + Eigen::DSizes shape(pre, n, post); + Eigen::DSizes norm_shape(pre, post); + auto x = x_e.reshape(shape); + auto dy = dy_e.reshape(shape); + auto norm = norm_e.reshape(norm_shape); + auto dx = dx_e.reshape(shape); + + framework::Tensor rsum; + rsum.mutable_data({pre, post}, ctx.GetPlace()); + auto sum = framework::EigenTensor::From(rsum); + + Eigen::DSizes rdim(1); + Eigen::DSizes bcast(1, n, 1); + Eigen::DSizes rshape(pre, 1, post); + + // dx = ( dy/sqrt(sum(x*x)) ) * [1 - x*sum(x) / (sum(x*x) + e)] + // = [dy - dy * x * sum(x) / (sum(x*x) + e)] / sqrt(sum(x*x)) + // = [dy - x * sum(x*dy) / (sum(x*x) + e)] / sqrt(sum(x*x)) + // 1. sum = sum(x*dy) + sum.device(*place) = (x * dy).sum(rdim); + // 2. dx = x * sum + dx.device(*place) = sum.reshape(rshape).broadcast(bcast) * x; + // 3. dx / (sum(x*x) + e) + // where, norm.pow(2) = sum(x*x) + e, which is calculated in forward. + dx.device(*place) = dx / norm.pow(2).broadcast(bcast); + // 4. [dy - dx] / sqrt(sum(x*x)) + dx.device(*place) = (dy - dx) / norm.broadcast(bcast); } }; } // namespace operators diff --git a/paddle/fluid/operators/pool_cudnn_op.cu.cc b/paddle/fluid/operators/pool_cudnn_op.cu.cc index d60a99994e..be55bc43b1 100644 --- a/paddle/fluid/operators/pool_cudnn_op.cu.cc +++ b/paddle/fluid/operators/pool_cudnn_op.cu.cc @@ -135,7 +135,11 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { PoolingMode pooling_mode; if (pooling_type == "max") { - pooling_mode = PoolingMode::kMaximum; + if (FLAGS_cudnn_deterministic) { + pooling_mode = PoolingMode::kMaximumDeterministic; + } else { + pooling_mode = PoolingMode::kMaximum; + } } else { pooling_mode = PoolingMode::kAverage; } diff --git a/paddle/fluid/operators/pool_mkldnn_op.cc b/paddle/fluid/operators/pool_mkldnn_op.cc index 60e936298d..5341187d1c 100644 --- a/paddle/fluid/operators/pool_mkldnn_op.cc +++ b/paddle/fluid/operators/pool_mkldnn_op.cc @@ -18,16 +18,24 @@ limitations under the License. */ namespace paddle { namespace operators { -using mkldnn::memory; // Note: paddle has also "memory" namespace -using mkldnn::pooling_forward; +using framework::DataLayout; +using mkldnn::memory; using mkldnn::pooling_backward; +using mkldnn::pooling_forward; +using mkldnn::primitive; +using mkldnn::reorder; +using mkldnn::stream; +using platform::to_void_cast; // Generate keys for storing/retriving primitives for this operator // TODO(jczaja): Make hashing function more optimial -static std::string gethash(memory::dims& input_dims, std::string& pooling_type, - std::vector& ksize, std::vector& strides, - std::vector& paddings, std::string suffix) { - auto dims2str = [](memory::dims& operand_dims) { +static std::string gethash(const memory::dims& input_dims, + const std::string& pooling_type, + const std::vector& ksize, + const std::vector& strides, + const std::vector& paddings, + const std::string& suffix) { + auto dims2str = [](const memory::dims& operand_dims) { std::string dstr = ""; for (size_t i = 0; i < operand_dims.size(); ++i) { dstr += std::to_string(operand_dims[i]) + "-"; @@ -52,8 +60,9 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { const Tensor* input = ctx.Input("X"); Tensor* output = ctx.Output("Out"); - // Get an unique name from "argument" name of "Out" variable - // This name will be used as key when saving info into device context + PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && + input->format() != memory::format::format_undef, + "Wrong layout/format set for Input tensor"); std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); @@ -79,6 +88,9 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { std::vector src_tz = paddle::framework::vectorize2int(input->dims()); std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); + auto input_format = input->format(); + memory::format output_format{memory::format::format_undef}; + const std::string key = gethash(src_tz, pooling_type, ksize, strides, paddings, ctx.op().Output("Out")); const std::string key_pool_p = key + "@pool_p"; @@ -91,16 +103,17 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { auto pool_p = std::static_pointer_cast(dev_ctx.GetBlob(key_pool_p)); if (pool_p == nullptr) { - // TODO(pzelazko-intel): support more formats + auto src_md = platform::MKLDNNMemDesc( + src_tz, platform::MKLDNNGetDataType(), input_format); - auto src_md = - platform::MKLDNNMemDesc(src_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); - auto dst_md = - platform::MKLDNNMemDesc(dst_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); + /* create memory descriptor for pooling without specified format + * ('any') which lets a primitive (pooling in this case) choose + * the memory format preferred for best performance + */ + auto dst_md = platform::MKLDNNMemDesc(dst_tz, mkldnn::memory::f32, + mkldnn::memory::format::any); - std::shared_ptr pool_pd = + std::shared_ptr pool_pd = CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize, pooling_type, mkldnn_engine); @@ -113,20 +126,22 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { // save pool_workspace_memory to be referred in backward path dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory); - auto pool_src_memory_p = std::make_shared( - memory::primitive_desc{src_md, mkldnn_engine}, - static_cast(const_cast(input_data))); - dev_ctx.SetBlob(key_pool_src_mem_p, pool_src_memory_p); + auto src_memory = std::make_shared(pool_pd->src_primitive_desc(), + to_void_cast(input_data)); + auto dst_memory = + std::make_shared(pool_pd->dst_primitive_desc(), output_data); - auto pool_dst_memory_p = std::make_shared( - memory::primitive_desc{dst_md, mkldnn_engine}, - static_cast(output_data)); - dev_ctx.SetBlob(key_pool_dst_mem_p, pool_dst_memory_p); + dev_ctx.SetBlob(key_pool_src_mem_p, src_memory); + dev_ctx.SetBlob(key_pool_dst_mem_p, dst_memory); + + pool_p = std::make_shared(*pool_pd, *(src_memory.get()), + *(dst_memory.get()), + *workspace_memory); - pool_p = std::make_shared( - *pool_pd, *(pool_src_memory_p.get()), *(pool_dst_memory_p.get()), - *workspace_memory); dev_ctx.SetBlob(key_pool_p, pool_p); + + output_format = + (memory::format)dst_memory->get_primitive_desc().desc().data.format; } else { // Primitives already exist auto pool_src_memory_p = @@ -137,14 +152,20 @@ class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { std::static_pointer_cast(dev_ctx.GetBlob(key_pool_dst_mem_p)); PADDLE_ENFORCE(pool_dst_memory_p != nullptr, "Fail to find pooling dst mem_p in device context"); - pool_src_memory_p->set_data_handle( - reinterpret_cast(const_cast(input_data))); + pool_src_memory_p->set_data_handle(to_void_cast(input_data)); pool_dst_memory_p->set_data_handle(output_data); + + output_format = (memory::format)pool_dst_memory_p->get_primitive_desc() + .desc() + .data.format; } // push primitive to stream and wait until it's executed std::vector pipeline{*(pool_p.get())}; - mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + stream(stream::kind::eager).submit(pipeline).wait(); + + output->set_layout(DataLayout::kMKLDNN); + output->set_format(output_format); } private: @@ -191,6 +212,13 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const Tensor* out_grad = ctx.Input(framework::GradVarName("Out")); Tensor* in_x_grad = ctx.Output(framework::GradVarName("X")); + PADDLE_ENFORCE(in_x->layout() == DataLayout::kMKLDNN && + in_x->format() != memory::format::format_undef, + "Wrong layout/format set for Input X tensor"); + PADDLE_ENFORCE(out_grad->layout() == DataLayout::kMKLDNN && + out_grad->format() != memory::format::format_undef, + "Wrong layout/format set for Input output_grad tensor"); + std::string pooling_type = ctx.Attr("pooling_type"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); @@ -209,6 +237,7 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const T* out_grad_data = out_grad->data(); T* in_x_grad_data = in_x_grad->mutable_data(ctx.GetPlace()); + memory::format in_x_grad_format{memory::format::format_undef}; std::vector diff_src_tz = paddle::framework::vectorize2int(in_x_grad->dims()); @@ -222,39 +251,48 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { const std::string key_pool_bwd_p = key + "@pool_bwd_p"; const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p"; const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p"; + const std::string key_pool_src_mem_p = key + "@pool_src_mem_p"; + const std::string key_pool_dst_mem_p = key + "@pool_dst_mem_p"; const std::string key_pool_pd = key + "@pool_pd"; const std::string key_pool_workspace_memory = key + "@pool_workspace_memory"; + auto user_diff_dst_memory = + memory({{{diff_dst_tz}, memory::data_type::f32, out_grad->format()}, + mkldnn_engine}, + to_void_cast(out_grad_data)); + + std::shared_ptr diff_src_memory; + std::shared_ptr diff_dst_memory; + auto dst_memory = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_dst_mem_p)); + PADDLE_ENFORCE(dst_memory != nullptr, + "Fail to find dst_memory in device context"); + + primitive reorder_diff_dst; + bool is_diff_dst_reordered = false; auto pool_bwd_p = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_bwd_p)); if (pool_bwd_p == nullptr) { - auto diff_src_md = - platform::MKLDNNMemDesc(diff_src_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); - auto diff_dst_md = - platform::MKLDNNMemDesc(diff_dst_tz, platform::MKLDNNGetDataType(), - mkldnn::memory::format::nchw); + // Retrieve src_memory/dst_memory saved in forward pass + auto src_memory = + std::static_pointer_cast(dev_ctx.GetBlob(key_pool_src_mem_p)); + PADDLE_ENFORCE(src_memory != nullptr, + "Fail to find src_memory in device context"); // Retrieve pool_pd/pool_workspace_memory from device context auto pool_pd = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_pd)); PADDLE_ENFORCE(pool_pd != nullptr, "Fail to find pool_pd in device context"); - - auto workspace_memory = std::static_pointer_cast( + auto workspace_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_workspace_memory)); PADDLE_ENFORCE(workspace_memory != nullptr, "Fail to find workspace_memory in device context"); - auto pool_diff_src_memory_p = std::make_shared(memory( - {diff_src_md, mkldnn_engine}, static_cast(in_x_grad_data))); - dev_ctx.SetBlob(key_pool_diff_src_mem_p, pool_diff_src_memory_p); - - auto pool_diff_dst_memory_p = std::make_shared( - memory({diff_dst_md, mkldnn_engine}, - static_cast(const_cast(out_grad_data)))); - dev_ctx.SetBlob(key_pool_diff_dst_mem_p, pool_diff_dst_memory_p); + // create memory descriptors for pooling + auto diff_src_md = src_memory.get()->get_primitive_desc().desc(); + auto diff_dst_md = dst_memory.get()->get_primitive_desc().desc(); auto pool_bwd_desc = mkldnn::pooling_backward::desc( pooling_type == "max" ? mkldnn::algorithm::pooling_max @@ -264,35 +302,74 @@ class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc( pool_bwd_desc, mkldnn_engine, *pool_pd); + // reorder between user_diff_dst and pool diff_dst if needed + diff_dst_memory = std::make_shared(user_diff_dst_memory); + if (memory::primitive_desc(dst_memory->get_primitive_desc()) != + user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory = + std::make_shared(dst_memory.get()->get_primitive_desc()); + reorder_diff_dst = reorder(user_diff_dst_memory, *diff_dst_memory); + is_diff_dst_reordered = true; + } + + diff_src_memory = std::make_shared( + pool_bwd_pd.diff_src_primitive_desc(), in_x_grad_data); + + dev_ctx.SetBlob(key_pool_diff_src_mem_p, diff_src_memory); + dev_ctx.SetBlob(key_pool_diff_dst_mem_p, diff_dst_memory); + pool_bwd_p = std::make_shared( - pool_bwd_pd, *(pool_diff_dst_memory_p.get()), *workspace_memory, - *(pool_diff_src_memory_p)); + pool_bwd_pd, *(diff_dst_memory.get()), *workspace_memory, + *(diff_src_memory)); dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p); + } else { // Primitives already exist - auto pool_diff_src_memory_p = std::static_pointer_cast( + diff_src_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_diff_src_mem_p)); - PADDLE_ENFORCE(pool_diff_src_memory_p != nullptr, + PADDLE_ENFORCE(diff_src_memory != nullptr, "Fail to find pooling src mem_p in device context"); - auto pool_diff_dst_memory_p = std::static_pointer_cast( + diff_dst_memory = std::static_pointer_cast( dev_ctx.GetBlob(key_pool_diff_dst_mem_p)); - PADDLE_ENFORCE(pool_diff_dst_memory_p != nullptr, + PADDLE_ENFORCE(diff_dst_memory != nullptr, "Fail to find pooling dst mem_p in device context"); - pool_diff_src_memory_p->set_data_handle( - reinterpret_cast(in_x_grad_data)); - pool_diff_dst_memory_p->set_data_handle(const_cast(out_grad_data)); + + diff_src_memory->set_data_handle(reinterpret_cast(in_x_grad_data)); + diff_dst_memory->set_data_handle(const_cast(out_grad_data)); + + // reorder between user_diff_dst and pool diff_dst if needed + if (memory::primitive_desc(dst_memory->get_primitive_desc()) != + user_diff_dst_memory.get_primitive_desc()) { + diff_dst_memory = + std::make_shared(dst_memory.get()->get_primitive_desc()); + reorder_diff_dst = reorder(user_diff_dst_memory, *diff_dst_memory); + is_diff_dst_reordered = true; + } } + in_x_grad_format = (memory::format)diff_src_memory->get_primitive_desc() + .desc() + .data.format; + // push primitive to stream and wait until it's executed - std::vector pipeline{*(pool_bwd_p.get())}; + std::vector pipeline; + if (is_diff_dst_reordered) { + pipeline.push_back(reorder_diff_dst); + } + pipeline.push_back(*(pool_bwd_p.get())); mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + + in_x_grad->set_layout(DataLayout::kMKLDNN); + in_x_grad->set_format(in_x_grad_format); } // Compute() }; } // namespace operators } // namespace paddle +namespace ops = paddle::operators; + REGISTER_OP_KERNEL(pool2d, MKLDNN, ::paddle::platform::CPUPlace, - paddle::operators::PoolMKLDNNOpKernel); + ops::PoolMKLDNNOpKernel); REGISTER_OP_KERNEL(pool2d_grad, MKLDNN, ::paddle::platform::CPUPlace, - paddle::operators::PoolMKLDNNGradOpKernel); + ops::PoolMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index f4fb2b132f..6707cdded4 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -83,6 +83,9 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const { framework::OpKernelType PoolOp::GetExpectedKernelType( const framework::ExecutionContext &ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -92,11 +95,10 @@ framework::OpKernelType PoolOp::GetExpectedKernelType( if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif - std::string data_format = ctx.Attr("data_format"); - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType( framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), layout_, library_); @@ -112,6 +114,9 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const { framework::OpKernelType PoolOpGrad::GetExpectedKernelType( const framework::ExecutionContext &ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -121,6 +126,7 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType( if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif @@ -129,8 +135,6 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType( PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN, "float16 can only be used when CUDNN is used"); } - std::string data_format = ctx.Attr("data_format"); - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_, library_); } @@ -147,7 +151,8 @@ void Pool2dOpMaker::Make() { "The format of output tensor is also NCHW, " "where N is batch size, C is the number of channels, " "H is the height of the feature, " - "and W is the width of the feature."); + "and W is the width of the feature.") + .Reuse("X"); AddAttr("pooling_type", "(string), pooling type, can be \"max\" for max-pooling " @@ -240,7 +245,8 @@ void Pool3dOpMaker::Make() { "The format of output tensor is also NCDHW, " "where N is batch size, C is " "the number of channels, and D, H and W is the depth, height and " - "width of the feature, respectively."); + "width of the feature, respectively.") + .Reuse("X"); AddAttr("pooling_type", "(string) Pooling type, can be \"max\" for max-pooling " diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc index e0a9b24ac8..f71ba84b31 100644 --- a/paddle/fluid/operators/prefetch_op.cc +++ b/paddle/fluid/operators/prefetch_op.cc @@ -18,7 +18,7 @@ limitations under the License. */ #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/operators/send_recv_util.h" namespace paddle { @@ -41,19 +41,19 @@ class PrefetchOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); - auto rpc_client = detail::RPCClient::GetInstance(); + detail::RPCClient* rpc_client = + detail::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { VLOG(3) << "sending " << ins[i] << " to " << epmap[i] << " to get " << outs[i] << " back"; - rpc_client->AsyncPrefetchVariable(epmap[i], ctx, scope, ins[i], - outs[i]); + rpc_client->AsyncPrefetchVar(epmap[i], ctx, scope, ins[i], outs[i]); } else { VLOG(3) << "don't send no-initialied variable: " << ins[i]; } } - PADDLE_ENFORCE(rpc_client->Wait()); + rpc_client->Wait(); } }; diff --git a/paddle/fluid/operators/random_crop_op.cc b/paddle/fluid/operators/random_crop_op.cc index b14b559e31..528a6e4a1b 100644 --- a/paddle/fluid/operators/random_crop_op.cc +++ b/paddle/fluid/operators/random_crop_op.cc @@ -20,7 +20,6 @@ class RandomCropOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( @@ -36,11 +35,11 @@ class RandomCropOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Seed", "The random seed."); AddOutput("Out", "The cropped instance batch."); AddOutput("SeedOut", "The random seed after random cropping.") - .AsDispensable(); + .AsIntermediate(); AddAttr>("shape", "The shape of a cropped instance."); AddComment(R"DOC( - This operator takes a batch of instance, and do random cropping on each instance. - It means that cropping positions differs on each instance, which is determined + This operator takes a batch of instance, and do random cropping on each instance. + It means that cropping positions differs on each instance, which is determined by an uniform random generator. All cropped instances have the same shape, which is determined by the operator's attribute 'shape'. )DOC"); diff --git a/paddle/fluid/operators/reader/create_batch_reader_op.cc b/paddle/fluid/operators/reader/create_batch_reader_op.cc index 4cc7cbc6e8..ecbae3894d 100644 --- a/paddle/fluid/operators/reader/create_batch_reader_op.cc +++ b/paddle/fluid/operators/reader/create_batch_reader_op.cc @@ -20,7 +20,7 @@ namespace reader { class BatchReader : public framework::DecoratedReader { public: - BatchReader(ReaderBase* reader, int batch_size) + BatchReader(const std::shared_ptr& reader, int batch_size) : DecoratedReader(reader), batch_size_(batch_size) { buffer_.reserve(batch_size_); } diff --git a/paddle/fluid/operators/reader/create_custom_reader_op.cc b/paddle/fluid/operators/reader/create_custom_reader_op.cc index 331224a598..0a02fcdeaa 100644 --- a/paddle/fluid/operators/reader/create_custom_reader_op.cc +++ b/paddle/fluid/operators/reader/create_custom_reader_op.cc @@ -22,7 +22,8 @@ namespace reader { class CustomReader : public framework::DecoratedReader { public: - CustomReader(ReaderBase* reader, const framework::BlockDesc& sub_block, + CustomReader(const std::shared_ptr& reader, + const framework::BlockDesc& sub_block, const std::vector& source_var_names, const std::vector& sink_var_names) : DecoratedReader(reader), diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc index bc830a2b72..5f35b9b3ea 100644 --- a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc +++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc @@ -34,7 +34,8 @@ static constexpr size_t kChannelSize = 1; // kCacheSize - 2 class DoubleBufferReader : public framework::DecoratedReader { public: explicit DoubleBufferReader( - ReaderBase* reader, platform::Place target_place = platform::CPUPlace()) + const std::shared_ptr& reader, + platform::Place target_place = platform::CPUPlace()) : DecoratedReader(reader), place_(target_place) { cpu_tensor_cache_.resize(kCacheSize); gpu_tensor_cache_.resize(kCacheSize); diff --git a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc index 249b0b7c6d..19b54110b9 100644 --- a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc +++ b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc @@ -21,7 +21,7 @@ namespace reader { class MultiPassReader : public framework::DecoratedReader { public: - MultiPassReader(ReaderBase* reader, int pass_num) + MultiPassReader(const std::shared_ptr& reader, int pass_num) : DecoratedReader(reader), pass_num_(pass_num), pass_count_(0) {} void ReadNext(std::vector* out) override { diff --git a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc index fd233be945..57e8e21214 100644 --- a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc +++ b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc @@ -23,7 +23,8 @@ namespace reader { class ShuffleReader : public framework::DecoratedReader { public: - ShuffleReader(ReaderBase* reader, size_t buffer_size, size_t seed = 0) + ShuffleReader(const std::shared_ptr& reader, size_t buffer_size, + size_t seed = 0) : DecoratedReader(reader), buffer_size_(buffer_size), seed_(seed) { VLOG(10) << "Create shuffle reader of " << reader_; if (seed_ == 0) { diff --git a/paddle/fluid/operators/reader/create_threaded_reader_op.cc b/paddle/fluid/operators/reader/create_threaded_reader_op.cc index 1db70f3e96..3798015146 100644 --- a/paddle/fluid/operators/reader/create_threaded_reader_op.cc +++ b/paddle/fluid/operators/reader/create_threaded_reader_op.cc @@ -21,7 +21,8 @@ namespace reader { class ThreadedReader : public framework::DecoratedReader { public: - explicit ThreadedReader(ReaderBase* reader) : DecoratedReader(reader) {} + explicit ThreadedReader(const std::shared_ptr& reader) + : DecoratedReader(reader) {} void ReadNext(std::vector* out) override { std::lock_guard lock(mutex_); diff --git a/paddle/fluid/operators/reader/open_files_op.cc b/paddle/fluid/operators/reader/open_files_op.cc index 8c0dac65dd..31e5d81e55 100644 --- a/paddle/fluid/operators/reader/open_files_op.cc +++ b/paddle/fluid/operators/reader/open_files_op.cc @@ -26,7 +26,11 @@ class MultiFileReader : public framework::ReaderBase { MultiFileReader(const std::vector& file_names, const std::vector& dims, size_t thread_num, size_t buffer_size) - : file_names_(file_names), dims_(dims), buffer_size_(buffer_size) { + : buffer_size_(buffer_size) { + readers_.reserve(file_names.size()); + for (const std::string& f_name : file_names) { + readers_.emplace_back(CreateReaderByFileName(f_name, dims)); + } prefetchers_.resize(thread_num); StartNewScheduler(); } @@ -40,14 +44,13 @@ class MultiFileReader : public framework::ReaderBase { void StartNewScheduler(); void EndScheduler(); void ScheduleThreadFunc(); - void PrefetchThreadFunc(std::string file_name, size_t thread_idx); + void PrefetchThreadFunc(size_t reader_idx, size_t thread_idx); - std::vector file_names_; - std::vector dims_; + std::vector> readers_; std::thread scheduler_; std::vector prefetchers_; size_t buffer_size_; - reader::BlockingQueue* waiting_file_idx_; + reader::BlockingQueue* waiting_reader_idx_; reader::BlockingQueue* available_thread_idx_; reader::BlockingQueue>* buffer_; }; @@ -65,15 +68,15 @@ void MultiFileReader::ReInit() { void MultiFileReader::StartNewScheduler() { size_t thread_num = prefetchers_.size(); - waiting_file_idx_ = new reader::BlockingQueue(file_names_.size()); + waiting_reader_idx_ = new reader::BlockingQueue(readers_.size()); available_thread_idx_ = new reader::BlockingQueue(thread_num); buffer_ = new reader::BlockingQueue>( buffer_size_); - for (size_t i = 0; i < file_names_.size(); ++i) { - waiting_file_idx_->Send(i); + for (size_t i = 0; i < readers_.size(); ++i) { + waiting_reader_idx_->Send(i); } - waiting_file_idx_->Close(); + waiting_reader_idx_->Close(); for (size_t i = 0; i < thread_num; ++i) { available_thread_idx_->Send(i); } @@ -84,13 +87,13 @@ void MultiFileReader::StartNewScheduler() { void MultiFileReader::EndScheduler() { available_thread_idx_->Close(); buffer_->Close(); - waiting_file_idx_->Close(); + waiting_reader_idx_->Close(); if (scheduler_.joinable()) { scheduler_.join(); } delete buffer_; delete available_thread_idx_; - delete waiting_file_idx_; + delete waiting_reader_idx_; } void MultiFileReader::ScheduleThreadFunc() { @@ -102,12 +105,11 @@ void MultiFileReader::ScheduleThreadFunc() { if (prefetcher.joinable()) { prefetcher.join(); } - size_t file_idx; - if (waiting_file_idx_->Receive(&file_idx)) { + size_t reader_idx; + if (waiting_reader_idx_->Receive(&reader_idx)) { // Still have files to read. Start a new prefetch thread. - std::string file_name = file_names_[file_idx]; - prefetcher = std::thread([this, file_name, thread_idx] { - PrefetchThreadFunc(file_name, thread_idx); + prefetcher = std::thread([this, reader_idx, thread_idx] { + PrefetchThreadFunc(reader_idx, thread_idx); }); } else { // No more file to read. @@ -129,23 +131,22 @@ void MultiFileReader::ScheduleThreadFunc() { VLOG(5) << "MultiFileReader schedule thread terminates."; } -void MultiFileReader::PrefetchThreadFunc(std::string file_name, - size_t thread_idx) { - VLOG(5) << "The prefetch thread of file '" << file_name << "' starts."; - std::unique_ptr reader = - CreateReaderByFileName(file_name, dims_); +void MultiFileReader::PrefetchThreadFunc(size_t reader_idx, size_t thread_idx) { + VLOG(5) << "The prefetch thread of file idx '" << reader_idx << "' starts."; + std::unique_ptr& reader = readers_[reader_idx]; while (true) { std::vector ins; reader->ReadNext(&ins); if (ins.empty()) { + reader->ReInit(); break; } try { buffer_->Send(std::move(ins)); } catch (paddle::platform::EnforceNotMet e) { VLOG(5) << "WARNING: The buffer channel has been closed. The prefetch " - "thread of file '" - << file_name << "' will terminate."; + "thread of file idx '" + << reader_idx << "' will terminate."; break; } } @@ -154,7 +155,8 @@ void MultiFileReader::PrefetchThreadFunc(std::string file_name, VLOG(5) << "WARNING: The available_thread_idx_ channel has been closed. " "Fail to send thread_idx."; } - VLOG(5) << "The prefetch thread of file '" << file_name << "' terminates."; + VLOG(5) << "The prefetch thread of file idx '" << reader_idx + << "' terminates."; } class OpenFilesOp : public framework::OperatorBase { diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index d8ddb7b448..15dfb5469b 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -19,8 +19,7 @@ limitations under the License. */ #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" - -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { @@ -44,14 +43,15 @@ class RecvOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - auto rpc_client = detail::RPCClient::GetInstance(); + detail::RPCClient* rpc_client = + detail::RPCClient::GetInstance(); for (size_t i = 0; i < outs.size(); i++) { VLOG(3) << "getting " << outs[i] << " from " << epmap[i]; - rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]); + rpc_client->AsyncGetVar(epmap[i], ctx, scope, outs[i]); } if (sync_mode) { - PADDLE_ENFORCE(rpc_client->Wait()); + rpc_client->Wait(); } } }; @@ -77,9 +77,15 @@ This operator can get variables from server side. } }; +class RecvOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override {} +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(recv, ops::RecvOp, ops::RecvOpMaker); +REGISTER_OPERATOR(recv, ops::RecvOp, paddle::framework::EmptyGradOpMaker, + ops::RecvOpMaker, ops::RecvOpShapeInference); diff --git a/paddle/fluid/operators/reduce_max_op.cc b/paddle/fluid/operators/reduce_max_op.cc new file mode 100644 index 0000000000..95d3768e1f --- /dev/null +++ b/paddle/fluid/operators/reduce_max_op.cc @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_REDUCE_OP(reduce_max); +REGISTER_OP_CPU_KERNEL( + reduce_max, ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL( + reduce_max_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_max_op.cu b/paddle/fluid/operators/reduce_max_op.cu new file mode 100644 index 0000000000..0d86b3127e --- /dev/null +++ b/paddle/fluid/operators/reduce_max_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_max, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_max_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.cc b/paddle/fluid/operators/reduce_mean_op.cc new file mode 100644 index 0000000000..fc258c2496 --- /dev/null +++ b/paddle/fluid/operators/reduce_mean_op.cc @@ -0,0 +1,35 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_mean_op.h" + +REGISTER_REDUCE_OP(reduce_mean); +REGISTER_OP_CPU_KERNEL(reduce_mean, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_mean_grad, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.cu b/paddle/fluid/operators/reduce_mean_op.cu new file mode 100644 index 0000000000..960cb3235b --- /dev/null +++ b/paddle/fluid/operators/reduce_mean_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_mean_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_mean, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_mean_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.h b/paddle/fluid/operators/reduce_mean_op.h new file mode 100644 index 0000000000..1359679c47 --- /dev/null +++ b/paddle/fluid/operators/reduce_mean_op.h @@ -0,0 +1,39 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +struct MeanFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->mean(dim); + } +}; + +struct MeanGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + dx->device(place) = dy->broadcast(dim) / dx->constant(size); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_min_max_op.h b/paddle/fluid/operators/reduce_min_max_op.h new file mode 100644 index 0000000000..ec59f3e71c --- /dev/null +++ b/paddle/fluid/operators/reduce_min_max_op.h @@ -0,0 +1,50 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#pragma once + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +struct MaxFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->maximum(dim); + } +}; + +struct MinFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->minimum(dim); + } +}; + +struct MaxOrMinGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + auto equals = (*x) == y->broadcast(dim); + auto ones = dx->constant(1); + auto zeros = dx->constant(0); + // If there are multiple minimum or maximum elements, the subgradient of + // each is the set [0, 1], and we pass gradient to all of them here. + dx->device(place) = dy->broadcast(dim) * equals.select(ones, zeros); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_min_op.cc b/paddle/fluid/operators/reduce_min_op.cc new file mode 100644 index 0000000000..330a86d2e4 --- /dev/null +++ b/paddle/fluid/operators/reduce_min_op.cc @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_REDUCE_OP(reduce_min); +REGISTER_OP_CPU_KERNEL( + reduce_min, ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL( + reduce_min_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_min_op.cu b/paddle/fluid/operators/reduce_min_op.cu new file mode 100644 index 0000000000..da466f805e --- /dev/null +++ b/paddle/fluid/operators/reduce_min_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_min_max_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_min, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_min_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_op.cc b/paddle/fluid/operators/reduce_op.cc deleted file mode 100644 index e293fd5e41..0000000000 --- a/paddle/fluid/operators/reduce_op.cc +++ /dev/null @@ -1,186 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/reduce_op.h" - -#include -#include -#include - -namespace paddle { -namespace operators { - -using framework::Tensor; - -class ReduceOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of ReduceOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of ReduceOp should not be null."); - auto x_dims = ctx->GetInputDim("X"); - auto x_rank = x_dims.size(); - PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); - auto dims = ctx->Attrs().Get>("dim"); - for (size_t i = 0; i < dims.size(); ++i) { - if (dims[i] < 0) dims[i] = x_rank + dims[i]; - PADDLE_ENFORCE_LT( - dims[i], x_rank, - "The dim should be in the range [-rank(input), rank(input))."); - } - sort(dims.begin(), dims.end()); - bool reduce_all = ctx->Attrs().Get("reduce_all"); - bool keep_dim = ctx->Attrs().Get("keep_dim"); - if (reduce_all) { - if (keep_dim) - ctx->SetOutputDim( - "Out", framework::make_ddim(std::vector(x_rank, 1))); - else - ctx->SetOutputDim("Out", {1}); - } else { - auto dims_vector = vectorize(x_dims); - if (keep_dim) { - for (size_t i = 0; i < dims.size(); ++i) { - dims_vector[dims[i]] = 1; - } - } else { - const int kDelFlag = -2; - for (size_t i = 0; i < dims.size(); ++i) { - dims_vector[dims[i]] = kDelFlag; - } - dims_vector.erase( - remove(dims_vector.begin(), dims_vector.end(), kDelFlag), - dims_vector.end()); - } - auto out_dims = framework::make_ddim(dims_vector); - ctx->SetOutputDim("Out", out_dims); - if (dims[0] != 0) { - // Only pass LoD when not reducing on the first dim. - ctx->ShareLoD("X", /*->*/ "Out"); - } - } - } -}; - -class ReduceGradOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); - PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), - "Input(Out@GRAD) should not be null."); - auto x_dims = ctx->GetInputDim("X"); - auto x_rank = x_dims.size(); - PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); - auto dims = ctx->Attrs().Get>("dim"); - for (size_t i = 0; i < dims.size(); ++i) { - if (dims[i] < 0) dims[i] = x_rank + dims[i]; - PADDLE_ENFORCE_LT( - dims[i], x_rank, - "The dim should be in the range [-rank(input), rank(input))."); - } - sort(dims.begin(), dims.end()); - auto x_grad_name = framework::GradVarName("X"); - if (ctx->HasOutput(x_grad_name)) { - ctx->SetOutputDim(x_grad_name, x_dims); - ctx->ShareLoD("X", /*->*/ x_grad_name); - } - } -}; - -class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() final { - AddInput("X", - "(Tensor) The input tensor. Tensors with rank at most 6 are " - "supported."); - AddOutput("Out", "(Tensor) The result tensor."); - AddAttr>( - "dim", - "(list, default {0}) The dimensions to reduce. " - "Must be in the range [-rank(input), rank(input)). " - "If `dim[i] < 0`, the dims[i] to reduce is `rank + dims[i]`. " - "Note that reducing on the first dim will make the LoD info lost.") - .SetDefault({0}); - AddAttr("keep_dim", - "(bool, default false) " - "If true, retain the reduced dimension with length 1.") - .SetDefault(false); - AddAttr("reduce_all", - "(bool, default false) " - "If true, output a scalar reduced along all dimensions.") - .SetDefault(false); - AddComment(string::Sprintf(R"DOC( -%s Operator. - -This operator computes the %s of input tensor along the given dimension. -The result tensor has 1 fewer dimension than the input unless keep_dim is true. -If reduce_all is true, just reduce along all dimensions and output a scalar. - -)DOC", - GetOpType(), GetName())); - } - - protected: - virtual std::string GetName() const = 0; - virtual std::string GetOpType() const = 0; -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; - -#define REGISTER_REDUCE_OP(op_name) \ - class __##op_name##Maker__ : public ops::ReduceOpMaker { \ - protected: \ - virtual std::string GetName() const { return #op_name; } \ - virtual std::string GetOpType() const { return "Reduce " #op_name; } \ - }; \ - REGISTER_OPERATOR(reduce_##op_name, ops::ReduceOp, __##op_name##Maker__, \ - paddle::framework::DefaultGradOpDescMaker); \ - REGISTER_OPERATOR(reduce_##op_name##_grad, ops::ReduceGradOp) - -REGISTER_REDUCE_OP(sum); -REGISTER_REDUCE_OP(mean); -REGISTER_REDUCE_OP(max); -REGISTER_REDUCE_OP(min); -REGISTER_REDUCE_OP(prod); - -#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ - REGISTER_OP_CPU_KERNEL(reduce_type, \ - ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel); \ - REGISTER_OP_CPU_KERNEL( \ - reduce_type##_grad, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel); - -FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL); diff --git a/paddle/fluid/operators/reduce_op.cu b/paddle/fluid/operators/reduce_op.cu deleted file mode 100644 index ae29587f55..0000000000 --- a/paddle/fluid/operators/reduce_op.cu +++ /dev/null @@ -1,41 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#define EIGEN_USE_GPU -#include "paddle/fluid/operators/reduce_op.h" - -namespace ops = paddle::operators; - -#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor) \ - REGISTER_OP_CUDA_KERNEL( \ - reduce_type, ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel, \ - ops::ReduceKernel); \ - REGISTER_OP_CUDA_KERNEL( \ - reduce_type##_grad, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel, \ - ops::ReduceGradKernel); - -FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL); diff --git a/paddle/fluid/operators/reduce_op.h b/paddle/fluid/operators/reduce_op.h index cd19cc1460..72b6cf1773 100644 --- a/paddle/fluid/operators/reduce_op.h +++ b/paddle/fluid/operators/reduce_op.h @@ -14,105 +14,20 @@ limitations under the License. */ #pragma once +#include +#include #include -#include "glog/logging.h" -#include "paddle/fluid/framework/eigen.h" -#include "paddle/fluid/framework/op_registry.h" + +#include "paddle/fluid/operators/reduce_op_function.h" namespace paddle { namespace operators { -using Tensor = framework::Tensor; -using DDim = framework::DDim; -template -using EigenTensor = framework::EigenTensor; -template -using EigenScalar = framework::EigenScalar; -template -using EigenVector = framework::EigenVector; - -struct SumFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->sum(dim); - } -}; - -struct SumGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - dx->device(place) = dy->broadcast(dim); - } -}; - -struct MeanFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->mean(dim); - } -}; - -struct MeanGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - dx->device(place) = dy->broadcast(dim) / dx->constant(size); - } -}; - -struct MaxFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->maximum(dim); - } -}; - -struct MinFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->minimum(dim); - } -}; - -struct MaxOrMinGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - auto equals = (*x) == y->broadcast(dim); - auto ones = dx->constant(1); - auto zeros = dx->constant(0); - // If there are multiple minimum or maximum elements, the subgradient of - // each is the set [0, 1], and we pass gradient to all of them here. - dx->device(place) = dy->broadcast(dim) * equals.select(ones, zeros); - } -}; - -struct ProdFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { - y->device(place) = x->prod(dim); - } -}; - -struct ProdGradFunctor { - template - void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, - const Dim& dim, int size) { - dx->device(place) = dy->broadcast(dim) * y->broadcast(dim) * x->inverse(); - } -}; - -#define HANDLE_DIM(NDIM, RDIM) \ - if (ndim == NDIM && rdim == RDIM) { \ - ReduceCompute(context); \ +#define HANDLE_DIM(NDIM, RDIM) \ + if (ndim == NDIM && rdim == RDIM) { \ + ReduceFunctor( \ + context.template device_context(), *input, output, \ + dims, keep_dim); \ } template @@ -120,11 +35,15 @@ class ReduceKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { bool reduce_all = context.Attr("reduce_all"); + auto* input = context.Input("X"); + auto* output = context.Output("Out"); + output->mutable_data(context.GetPlace()); + + auto dims = context.Attr>("dim"); + bool keep_dim = context.Attr("keep_dim"); + if (reduce_all) { // Flatten and reduce 1-D tensor - auto* input = context.Input("X"); - auto* output = context.Output("Out"); - output->mutable_data(context.GetPlace()); auto x = EigenVector::Flatten(*input); auto out = EigenScalar::From(*output); auto& place = @@ -133,17 +52,18 @@ class ReduceKernel : public framework::OpKernel { Functor functor; functor(place, &x, &out, reduce_dim); } else { - int ndim = context.Input("X")->dims().size(); - int rdim = context.Attr>("dim").size(); - HANDLE_DIM(6, 5); - HANDLE_DIM(6, 4); - HANDLE_DIM(6, 3); - HANDLE_DIM(6, 2); - HANDLE_DIM(6, 1); - HANDLE_DIM(5, 4); - HANDLE_DIM(5, 3); - HANDLE_DIM(5, 2); - HANDLE_DIM(5, 1); + int ndim = input->dims().size(); + int rdim = dims.size(); + // comments for accelerating compiling temporarily. + // HANDLE_DIM(6, 5); + // HANDLE_DIM(6, 4); + // HANDLE_DIM(6, 3); + // HANDLE_DIM(6, 2); + // HANDLE_DIM(6, 1); + // HANDLE_DIM(5, 4); + // HANDLE_DIM(5, 3); + // HANDLE_DIM(5, 2); + // HANDLE_DIM(5, 1); HANDLE_DIM(4, 3); HANDLE_DIM(4, 2); HANDLE_DIM(4, 1); @@ -153,48 +73,6 @@ class ReduceKernel : public framework::OpKernel { HANDLE_DIM(1, 1); } } - - private: - template - void ReduceCompute(const framework::ExecutionContext& context) const { - auto* input = context.Input("X"); - auto* output = context.Output("Out"); - output->mutable_data(context.GetPlace()); - - auto x = EigenTensor::From(*input); - auto x_rank = static_cast(x.dimensions().size()); - auto dims = context.Attr>("dim"); - auto reduce_dim = Eigen::array(); - for (size_t i = 0; i < dims.size(); ++i) { - if (dims[i] < 0) dims[i] = x_rank + dims[i]; - reduce_dim[i] = dims[i]; - } - // construct the squeezed output tensor - bool keep_dim = context.Attr("keep_dim"); - DDim out_dims = output->dims(); - if (keep_dim && x_rank > 1) { - const int kDelFlag = -2; - auto dims_vector = vectorize(out_dims); - for (size_t i = 0; i < dims.size(); ++i) { - dims_vector[dims[i]] = kDelFlag; - } - dims_vector.erase( - remove(dims_vector.begin(), dims_vector.end(), kDelFlag), - dims_vector.end()); - out_dims = framework::make_ddim(dims_vector); - } - auto& place = - *context.template device_context().eigen_device(); - Functor functor; - - if (D == 1) { - auto out = EigenScalar::From(*output); - functor(place, &x, &out, reduce_dim); - } else { - auto out = EigenTensor::From(*output, out_dims); - functor(place, &x, &out, reduce_dim); - } - } }; template @@ -202,12 +80,15 @@ class ReduceGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { bool reduce_all = context.Attr("reduce_all"); + auto dims = context.Attr>("dim"); + + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Out"); + auto* input2 = context.Input(framework::GradVarName("Out")); + auto* output = context.Output(framework::GradVarName("X")); + output->mutable_data(context.GetPlace()); + if (reduce_all) { - auto* input0 = context.Input("X"); - auto* input1 = context.Input("Out"); - auto* input2 = context.Input(framework::GradVarName("Out")); - auto* output = context.Output(framework::GradVarName("X")); - output->mutable_data(context.GetPlace()); auto x = EigenVector::Flatten(*input0); auto x_reduce = EigenVector::From(*input1); auto x_reduce_grad = EigenVector::From(*input2); @@ -220,74 +101,172 @@ class ReduceGradKernel : public framework::OpKernel { functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim, broadcast_dim[0]); } else { - int rank = context.Input("X")->dims().size(); + int rank = input0->dims().size(); switch (rank) { case 1: - ReduceGradCompute<1>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 2: - ReduceGradCompute<2>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 3: - ReduceGradCompute<3>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 4: - ReduceGradCompute<4>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 5: - ReduceGradCompute<5>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; case 6: - ReduceGradCompute<6>(context); + ReduceGradFunctor( + context.template device_context(), *input0, + *input1, *input2, output, dims); break; } } } +}; - private: - template - void ReduceGradCompute(const framework::ExecutionContext& context) const { - auto* input0 = context.Input("X"); - auto* input1 = context.Input("Out"); - auto* input2 = context.Input(framework::GradVarName("Out")); - auto* output = context.Output(framework::GradVarName("X")); +class ReduceOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; - output->mutable_data(context.GetPlace()); - auto x = EigenTensor::From(*input0); - auto x_grad = EigenTensor::From(*output); - auto x_rank = static_cast(x.dimensions().size()); - auto dims = context.Attr>("dim"); - auto x_dims = input0->dims(); - auto reduced_dims_v = vectorize(x_dims); - Eigen::array broadcast_dim; - for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ReduceOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReduceOp should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + auto dims = ctx->Attrs().Get>("dim"); + for (size_t i = 0; i < dims.size(); ++i) { + if (dims[i] < 0) dims[i] = x_rank + dims[i]; + PADDLE_ENFORCE_LT( + dims[i], x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + } + sort(dims.begin(), dims.end()); + bool reduce_all = ctx->Attrs().Get("reduce_all"); + bool keep_dim = ctx->Attrs().Get("keep_dim"); + if (reduce_all) { + if (keep_dim) + ctx->SetOutputDim( + "Out", framework::make_ddim(std::vector(x_rank, 1))); + else + ctx->SetOutputDim("Out", {1}); + } else { + auto dims_vector = vectorize(x_dims); + if (keep_dim) { + for (size_t i = 0; i < dims.size(); ++i) { + dims_vector[dims[i]] = 1; + } + } else { + const int kDelFlag = -2; + for (size_t i = 0; i < dims.size(); ++i) { + dims_vector[dims[i]] = kDelFlag; + } + dims_vector.erase( + remove(dims_vector.begin(), dims_vector.end(), kDelFlag), + dims_vector.end()); + } + auto out_dims = framework::make_ddim(dims_vector); + ctx->SetOutputDim("Out", out_dims); + if (dims[0] != 0) { + // Only pass LoD when not reducing on the first dim. + ctx->ShareLoD("X", /*->*/ "Out"); + } + } + } +}; + +class ReduceGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; - int broad_cats_times = 1; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + auto x_dims = ctx->GetInputDim("X"); + auto x_rank = x_dims.size(); + PADDLE_ENFORCE_LE(x_rank, 6, "Tensors with rank at most 6 are supported."); + auto dims = ctx->Attrs().Get>("dim"); for (size_t i = 0; i < dims.size(); ++i) { if (dims[i] < 0) dims[i] = x_rank + dims[i]; - reduced_dims_v[dims[i]] = 1; - broadcast_dim[dims[i]] = x_dims[dims[i]]; - broad_cats_times *= x_dims[dims[i]]; + PADDLE_ENFORCE_LT( + dims[i], x_rank, + "The dim should be in the range [-rank(input), rank(input))."); + } + sort(dims.begin(), dims.end()); + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, x_dims); + ctx->ShareLoD("X", /*->*/ x_grad_name); } - auto reduced_dims = framework::make_ddim(reduced_dims_v); - auto x_reduce = EigenTensor::From(*input1, reduced_dims); - auto x_reduce_grad = EigenTensor::From(*input2, reduced_dims); + } +}; + +class ReduceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() final { + AddInput("X", + "(Tensor) The input tensor. Tensors with rank at most 6 are " + "supported."); + AddOutput("Out", "(Tensor) The result tensor."); + AddAttr>( + "dim", + "(list, default {0}) The dimensions to reduce. " + "Must be in the range [-rank(input), rank(input)). " + "If `dim[i] < 0`, the dims[i] to reduce is `rank + dims[i]`. " + "Note that reducing on the first dim will make the LoD info lost.") + .SetDefault({0}); + AddAttr("keep_dim", + "(bool, default false) " + "If true, retain the reduced dimension with length 1.") + .SetDefault(false); + AddAttr("reduce_all", + "(bool, default false) " + "If true, output a scalar reduced along all dimensions.") + .SetDefault(false); + AddComment(string::Sprintf(R"DOC( +%s Operator. - auto& place = - *context.template device_context().eigen_device(); +This operator computes the %s of input tensor along the given dimension. +The result tensor has 1 fewer dimension than the input unless keep_dim is true. +If reduce_all is true, just reduce along all dimensions and output a scalar. - Functor functor; - functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim, - broad_cats_times); +)DOC", + GetOpType(), GetName())); } + + protected: + virtual std::string GetName() const = 0; + virtual std::string GetOpType() const = 0; }; } // namespace operators } // namespace paddle -#define FOR_EACH_KERNEL_FUNCTOR(__macro) \ - __macro(reduce_sum, SumFunctor, SumGradFunctor); \ - __macro(reduce_mean, MeanFunctor, MeanGradFunctor); \ - __macro(reduce_max, MaxFunctor, MaxOrMinGradFunctor); \ - __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor); \ - __macro(reduce_prod, ProdFunctor, ProdGradFunctor); +namespace ops = paddle::operators; + +#define REGISTER_REDUCE_OP(op_name) \ + class __##op_name##Maker__ : public ops::ReduceOpMaker { \ + protected: \ + virtual std::string GetName() const { return #op_name; } \ + virtual std::string GetOpType() const { return "Reduce " #op_name; } \ + }; \ + REGISTER_OPERATOR(op_name, ops::ReduceOp, __##op_name##Maker__, \ + paddle::framework::DefaultGradOpDescMaker); \ + REGISTER_OPERATOR(op_name##_grad, ops::ReduceGradOp) diff --git a/paddle/fluid/operators/reduce_op_function.h b/paddle/fluid/operators/reduce_op_function.h new file mode 100644 index 0000000000..3da27bc8ac --- /dev/null +++ b/paddle/fluid/operators/reduce_op_function.h @@ -0,0 +1,109 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using DDim = framework::DDim; +template +using EigenTensor = framework::EigenTensor; +template +using EigenScalar = framework::EigenScalar; +template +using EigenVector = framework::EigenVector; + +template +void ReduceFunctor(const DeviceContext& context, const framework::Tensor& input, + framework::Tensor* output, const std::vector& dims, + bool keep_dim) { + auto x = EigenTensor::From(input); + auto x_rank = static_cast(x.dimensions().size()); + auto reduce_dim = Eigen::array(); + std::vector dims_ref = dims; + for (size_t i = 0; i < dims_ref.size(); ++i) { + if (dims_ref[i] < 0) dims_ref[i] = x_rank + dims_ref[i]; + reduce_dim[i] = dims_ref[i]; + } + // construct the squeezed output tensor + DDim out_dims = output->dims(); + if (keep_dim && x_rank > 1) { + const int kDelFlag = -2; + auto dims_vector = framework::vectorize(out_dims); + for (size_t i = 0; i < dims_ref.size(); ++i) { + dims_vector[dims_ref[i]] = kDelFlag; + } + dims_vector.erase(remove(dims_vector.begin(), dims_vector.end(), kDelFlag), + dims_vector.end()); + out_dims = framework::make_ddim(dims_vector); + } + auto& place = *context.eigen_device(); + Functor functor; + + if (D == 1) { + auto out = EigenScalar::From(*output); + functor(place, &x, &out, reduce_dim); + } else { + auto out = EigenTensor::From(*output, out_dims); + functor(place, &x, &out, reduce_dim); + } +} + +template +void ReduceGradFunctor(const DeviceContext& context, + const framework::Tensor& input0, + const framework::Tensor& input1, + const framework::Tensor& input2, + framework::Tensor* output, + const std::vector& dims) { + auto x = EigenTensor::From(input0); + auto x_grad = EigenTensor::From(*output); + auto x_rank = static_cast(x.dimensions().size()); + auto x_dims = input0.dims(); + auto reduced_dims_v = framework::vectorize(x_dims); + std::vector dims_ref = dims; + Eigen::array broadcast_dim; + for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1; + + int broad_cats_times = 1; + for (size_t i = 0; i < dims_ref.size(); ++i) { + if (dims_ref[i] < 0) { + dims_ref[i] = x_rank + dims_ref[i]; + } + reduced_dims_v[dims_ref[i]] = 1; + broadcast_dim[dims_ref[i]] = x_dims[dims_ref[i]]; + broad_cats_times *= x_dims[dims_ref[i]]; + } + auto reduced_dims = framework::make_ddim(reduced_dims_v); + auto x_reduce = EigenTensor::From(input1, reduced_dims); + auto x_reduce_grad = EigenTensor::From(input2, reduced_dims); + + auto& place = *context.eigen_device(); + + Functor functor; + functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim, + broad_cats_times); +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_prod_op.cc b/paddle/fluid/operators/reduce_prod_op.cc new file mode 100644 index 0000000000..713728b997 --- /dev/null +++ b/paddle/fluid/operators/reduce_prod_op.cc @@ -0,0 +1,35 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_prod_op.h" + +REGISTER_REDUCE_OP(reduce_prod); +REGISTER_OP_CPU_KERNEL(reduce_prod, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_prod_grad, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_prod_op.cu b/paddle/fluid/operators/reduce_prod_op.cu new file mode 100644 index 0000000000..d62e677d92 --- /dev/null +++ b/paddle/fluid/operators/reduce_prod_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_prod_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_prod, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_prod_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_prod_op.h b/paddle/fluid/operators/reduce_prod_op.h new file mode 100644 index 0000000000..97748113e0 --- /dev/null +++ b/paddle/fluid/operators/reduce_prod_op.h @@ -0,0 +1,39 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +struct ProdFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->prod(dim); + } +}; + +struct ProdGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + dx->device(place) = dy->broadcast(dim) * y->broadcast(dim) * x->inverse(); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reduce_sum_op.cc b/paddle/fluid/operators/reduce_sum_op.cc new file mode 100644 index 0000000000..c5b5398787 --- /dev/null +++ b/paddle/fluid/operators/reduce_sum_op.cc @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_sum_op.h" + +REGISTER_REDUCE_OP(reduce_sum); +REGISTER_OP_CPU_KERNEL( + reduce_sum, ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CPU_KERNEL(reduce_sum_grad, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_sum_op.cu b/paddle/fluid/operators/reduce_sum_op.cu new file mode 100644 index 0000000000..f2e16955a5 --- /dev/null +++ b/paddle/fluid/operators/reduce_sum_op.cu @@ -0,0 +1,34 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reduce_sum_op.h" + +REGISTER_OP_CUDA_KERNEL(reduce_sum, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel, + ops::ReduceKernel); +REGISTER_OP_CUDA_KERNEL( + reduce_sum_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_sum_op.h b/paddle/fluid/operators/reduce_sum_op.h new file mode 100644 index 0000000000..e67d7e1da5 --- /dev/null +++ b/paddle/fluid/operators/reduce_sum_op.h @@ -0,0 +1,39 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/operators/reduce_op.h" + +namespace paddle { +namespace operators { + +struct SumFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) { + y->device(place) = x->sum(dim); + } +}; + +struct SumGradFunctor { + template + void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy, + const Dim& dim, int size) { + dx->device(place) = dy->broadcast(dim); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reverse_op.cc b/paddle/fluid/operators/reverse_op.cc new file mode 100644 index 0000000000..a20f7d231f --- /dev/null +++ b/paddle/fluid/operators/reverse_op.cc @@ -0,0 +1,107 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reverse_op.h" +#include + +namespace paddle { +namespace operators { + +class ReverseOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null"); + const auto& x_dims = ctx->GetInputDim("X"); + const auto& axis = ctx->Attrs().Get>("axis"); + PADDLE_ENFORCE(!axis.empty(), "'axis' can not be empty."); + for (int a : axis) { + PADDLE_ENFORCE_LT(a, x_dims.size(), + "The axis must be less than input tensor's rank."); + } + ctx->SetOutputDim("Out", x_dims); + } +}; + +class ReverseOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "The LoDTensor to be flipped."); + AddOutput("Out", "The LoDTensor after flipping."); + AddAttr>( + "axis", "The axises that along which order of elements is reversed."); + AddComment(R"DOC( + Reverse Operator. + + Reverse the order of elements in the input LoDTensor along given axises. + + Case 1: + Given + X = [[1, 2, 3, 4, 5] + [6, 7, 8, 9, 10] + [11, 12, 13, 14, 15]], + and + axis = [0], + we get: + Out = [[11, 12, 13, 14, 15] + [6, 7, 8, 9, 10] + [1, 2, 3, 4, 5]]. + + Case 2: + Given + X = [[[1, 2, 3, 4] + [5, 6, 7, 8]] + [[9, 10, 11, 12] + [13, 14, 15, 16]]], + and + axis = [0, 2], + we get: + Out = [[[12, 11, 10, 9] + [16, 15, 14, 13]] + [[4, 3, 2, 1] + [8, 7, 6, 5]]], + )DOC"); + } +}; + +class ReverseGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + std::unique_ptr Apply() const override { + auto* grad_op = new framework::OpDesc(); + grad_op->SetType("reverse"); + grad_op->SetInput("X", OutputGrad("Out")); + grad_op->SetOutput("Out", InputGrad("X")); + grad_op->SetAttr("axis", GetAttr("axis")); + return std::unique_ptr(grad_op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(reverse, ops::ReverseOp, ops::ReverseOpMaker, + ops::ReverseGradMaker); +REGISTER_OPERATOR(reverse_grad, ops::ReverseOp); +REGISTER_OP_CPU_KERNEL( + reverse, ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel) diff --git a/paddle/fluid/operators/reverse_op.cu b/paddle/fluid/operators/reverse_op.cu new file mode 100644 index 0000000000..635c41529b --- /dev/null +++ b/paddle/fluid/operators/reverse_op.cu @@ -0,0 +1,24 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/operators/reverse_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + reverse, ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel, + ops::ReverseKernel) diff --git a/paddle/fluid/operators/reverse_op.h b/paddle/fluid/operators/reverse_op.h new file mode 100644 index 0000000000..9063cd59bb --- /dev/null +++ b/paddle/fluid/operators/reverse_op.h @@ -0,0 +1,87 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +template +struct ReverseFunctor { + void operator()(const DeviceContext& context, const framework::LoDTensor& in, + framework::LoDTensor* out, const std::vector& axis) { + Eigen::array reverse_axis; + for (int i = 0; i < Rank; ++i) { + reverse_axis[i] = false; + } + for (int a : axis) { + reverse_axis[a] = true; + } + + auto in_eigen = framework::EigenTensor::From(in); + auto out_eigen = framework::EigenTensor::From(*out); + auto* dev = context.eigen_device(); + + out_eigen.device(*dev) = in_eigen.reverse(reverse_axis); + } +}; + +template +class ReverseKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* x = context.Input("X"); + auto* out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + const auto& axis = context.Attr>("axis"); + int rank = x->dims().size(); + auto& dev_ctx = context.template device_context(); + + switch (rank) { + case 1: + ReverseFunctor functor1; + functor1(dev_ctx, *x, out, axis); + break; + case 2: + ReverseFunctor functor2; + functor2(dev_ctx, *x, out, axis); + break; + case 3: + ReverseFunctor functor3; + functor3(dev_ctx, *x, out, axis); + break; + case 4: + ReverseFunctor functor4; + functor4(dev_ctx, *x, out, axis); + break; + case 5: + ReverseFunctor functor5; + functor5(dev_ctx, *x, out, axis); + break; + case 6: + ReverseFunctor functor6; + functor6(dev_ctx, *x, out, axis); + break; + default: + PADDLE_THROW( + "Reserve operator doesn't supports tensors whose ranks are greater " + "than 6."); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 2c77ee2e27..c6c975a23c 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -19,8 +19,8 @@ limitations under the License. */ #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/macros.h" -#include "paddle/fluid/operators/detail/grpc_client.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { @@ -44,16 +44,19 @@ class SendBarrierOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - auto rpc_client = detail::RPCClient::GetInstance(); + detail::RPCClient* rpc_client = + detail::RPCClient::GetInstance(); + + VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode; // need to wait before sending send_barrier message - PADDLE_ENFORCE(rpc_client->Wait()); + rpc_client->Wait(); if (sync_mode) { for (auto& ep : eps) { VLOG(3) << "send barrier, ep: " << ep; rpc_client->AsyncSendBatchBarrier(ep); } - PADDLE_ENFORCE(rpc_client->Wait()); + rpc_client->Wait(); } } }; diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index a5150f242c..84ec366253 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -16,10 +16,9 @@ limitations under the License. */ #include #include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" #include "paddle/fluid/operators/send_recv_util.h" #include "paddle/fluid/platform/profiler.h" @@ -36,12 +35,9 @@ class SendOp : public framework::OperatorBase { void RunImpl(const framework::Scope& scope, const platform::Place& place) const override { auto ins = Inputs("X"); - auto outs = Outputs("Out"); - std::vector epmap = Attr>("epmap"); - std::vector endpoints = - Attr>("endpoints"); - bool sync_mode = Attr("sync_mode"); + std::vector epmap = Attr>("epmap"); + int sync_send = Attr("sync_mode"); platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); @@ -49,38 +45,21 @@ class SendOp : public framework::OperatorBase { // For profiling platform::RecordEvent record_event(Type(), &ctx); - auto rpc_client = detail::RPCClient::GetInstance(); + detail::RPCClient* rpc_client = + detail::RPCClient::GetInstance(); for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { VLOG(3) << "sending " << ins[i] << " to " << epmap[i]; - rpc_client->AsyncSendVariable(epmap[i], ctx, scope, ins[i]); + // TODO(Yancey1989): we need to use an IO threadpool which has + // a larger number of threads than the computing threadpool. + rpc_client->AsyncSendVar(epmap[i], ctx, scope, ins[i]); } else { VLOG(3) << "don't send no-initialied variable: " << ins[i]; } } - PADDLE_ENFORCE(rpc_client->Wait()); - - if (sync_mode) { - for (auto& ep : endpoints) { - VLOG(3) << "batch barrier, ep: " << ep; - rpc_client->AsyncSendBatchBarrier(ep); - } - PADDLE_ENFORCE(rpc_client->Wait()); - } - - if (outs.size() > 0) { - for (size_t i = 0; i < outs.size(); i++) { - VLOG(2) << "getting " << outs[i] << " from " << epmap[i]; - rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]); - } - PADDLE_ENFORCE(rpc_client->Wait()); - // tell pservers that current trainer have called fetch - for (auto& ep : endpoints) { - VLOG(2) << "send fetch barrier, ep: " << ep; - rpc_client->AsyncSendFetchBarrier(ep); - } - PADDLE_ENFORCE(rpc_client->Wait()); + if (sync_send) { + rpc_client->Wait(); } } }; @@ -88,26 +67,22 @@ class SendOp : public framework::OperatorBase { class SendOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() { - AddInput("X", "(Tensor) Input tensor to be sent").AsDuplicable(); - AddOutput("Out", "(Tensor) Output tensor to be received from server") + AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") .AsDuplicable(); AddComment(R"DOC( Send operator -This operator will send tensor to recv_op at the parameter server. +This operator will send variables to listen_and_serve op at the parameter server. )DOC"); - // TODO(typhoonzero): remove this attr generate de-duplicated vector from - // epmap when initializing. - AddAttr>("endpoints", - "(string vector, default 127.0.0.1:6164)" - "Server endpoints to send variables to.") - .SetDefault({}); + AddAttr("sync_mode", + "(int, default 0)" + "sync send or async send.") + .SetDefault(0); AddAttr>("epmap", "(string vector, default 127.0.0.1:6164)" "Server endpoints in the order of input " "variables for mapping") - .SetDefault({}); - AddAttr("sync_mode", "work in sync_mode or not").SetDefault(true); + .SetDefault({"127.0.0.1:6164"}); } }; diff --git a/paddle/fluid/operators/send_vars_op.cc b/paddle/fluid/operators/send_vars_op.cc deleted file mode 100644 index fe839dab69..0000000000 --- a/paddle/fluid/operators/send_vars_op.cc +++ /dev/null @@ -1,100 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include // NOLINT -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/detail/grpc_client.h" -#include "paddle/fluid/operators/send_recv_util.h" -#include "paddle/fluid/platform/profiler.h" - -namespace paddle { -namespace operators { - -class SendVarsOp : public framework::OperatorBase { - public: - SendVarsOp(const std::string& type, const framework::VariableNameMap& inputs, - const framework::VariableNameMap& outputs, - const framework::AttributeMap& attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - void RunImpl(const framework::Scope& scope, - const platform::Place& place) const override { - auto ins = Inputs("X"); - - std::vector epmap = Attr>("epmap"); - int sync_send = Attr("sync_send"); - - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(place); - - // For profiling - platform::RecordEvent record_event(Type(), &ctx); - - auto rpc_client = detail::RPCClient::GetInstance(); - - for (size_t i = 0; i < ins.size(); i++) { - if (NeedSend(scope, ins[i])) { - VLOG(3) << "sending " << ins[i] << " to " << epmap[i]; - // TODO(Yancey1989): we need to use an IO threadpool which has - // a larger number of threads than the computing threadpool. - rpc_client->AsyncSendVariable(epmap[i], ctx, scope, ins[i]); - } else { - VLOG(3) << "don't send no-initialied variable: " << ins[i]; - } - } - if (sync_send) { - rpc_client->Wait(); - } - } -}; - -class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() { - AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") - .AsDuplicable(); - AddComment(R"DOC( -Send operator - -This operator will send variables to listen_and_serve op at the parameter server. -)DOC"); - AddAttr("sync_send", - "(int, default 0)" - "sync send or async send.") - .SetDefault(0); - AddAttr>("epmap", - "(string vector, default 127.0.0.1:6164)" - "Server endpoints in the order of input " - "variables for mapping") - .SetDefault({"127.0.0.1:6164"}); - } -}; - -class SendVarsOpShapeInference : public framework::InferShapeBase { - public: - void operator()(framework::InferShapeContext* ctx) const override {} -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; - -REGISTER_OPERATOR(send_vars, ops::SendVarsOp, - paddle::framework::EmptyGradOpMaker, ops::SendVarsOpMaker, - ops::SendVarsOpShapeInference); diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index 7a2bdeac09..fef230e42d 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -74,7 +74,8 @@ class SGDOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Grad", "(Tensor or SelectedRows) Input gradient"); AddOutput("ParamOut", "(Tensor or SelectedRows, same with Param) " - "Output parameter, should share the same memory with Param"); + "Output parameter, should share the same memory with Param") + .Reuse("Param"); AddComment(R"DOC( SGD operator diff --git a/paddle/fluid/operators/sgd_op.h b/paddle/fluid/operators/sgd_op.h index f9e0596191..2685ce217e 100644 --- a/paddle/fluid/operators/sgd_op.h +++ b/paddle/fluid/operators/sgd_op.h @@ -114,7 +114,7 @@ class SGDOpKernel : public framework::OpKernel { int64_t id_index = param.Index(grad.rows()[i]); PADDLE_ENFORCE_GE(id_index, static_cast(0), "id should be in the table"); - for (size_t j = 0; j < grad_row_width; j++) { + for (int64_t j = 0; j < grad_row_width; j++) { out_data[id_index * grad_row_width + j] -= lr[0] * grad_data[i * grad_row_width + j]; } diff --git a/paddle/fluid/operators/shape_op.cc b/paddle/fluid/operators/shape_op.cc new file mode 100644 index 0000000000..c75fce7959 --- /dev/null +++ b/paddle/fluid/operators/shape_op.cc @@ -0,0 +1,54 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/shape_op.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class ShapeOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input (Input) of get_shape op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output (Out) of get_shape op should not be null."); + auto in_dim = ctx->GetInputDim("Input"); + ctx->SetOutputDim("Out", {in_dim.size()}); + } +}; + +class ShapeOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Input", "(Tensor), The input tensor."); + AddOutput("Out", "(Tensor), The shape of input tensor."); + AddComment(R"DOC( +Shape Operator. +Get the shape of input tensor. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(shape, ops::ShapeOp, ops::ShapeOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(shape, ops::ShapeKernel, ops::ShapeKernel, + ops::ShapeKernel, ops::ShapeKernel); diff --git a/paddle/fluid/operators/shape_op.cu b/paddle/fluid/operators/shape_op.cu new file mode 100644 index 0000000000..7736a2a1e1 --- /dev/null +++ b/paddle/fluid/operators/shape_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/shape_op.h" + +REGISTER_OP_CUDA_KERNEL(shape, paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel); diff --git a/paddle/fluid/operators/shape_op.h b/paddle/fluid/operators/shape_op.h new file mode 100644 index 0000000000..3be86b66a5 --- /dev/null +++ b/paddle/fluid/operators/shape_op.h @@ -0,0 +1,38 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class ShapeKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in_t = ctx.Input("Input"); + auto* out_t = ctx.Output("Out"); + auto out_data = out_t->mutable_data(platform::CPUPlace()); + auto in_dims = in_t->dims(); + for (int i = 0; i < in_dims.size(); ++i) { + out_data[i] = in_dims[i]; + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/slice_op.cc b/paddle/fluid/operators/slice_op.cc new file mode 100644 index 0000000000..61bb445e8b --- /dev/null +++ b/paddle/fluid/operators/slice_op.cc @@ -0,0 +1,130 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/slice_op.h" +#include +#include + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +class SliceOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input (Input) of slice op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output (Out) of slice op should not be null."); + + auto in_dims = ctx->GetInputDim("Input"); + PADDLE_ENFORCE(in_dims.size() < 7, + "The rank of input should be less than 7."); + framework::DDim out_dims(in_dims); + auto axes = ctx->Attrs().Get>("axes"); + auto starts = ctx->Attrs().Get>("starts"); + auto ends = ctx->Attrs().Get>("ends"); + + PADDLE_ENFORCE_EQ(starts.size(), ends.size()); + PADDLE_ENFORCE_EQ(starts.size(), axes.size()); + int dim_value, start, end; + for (size_t i = 0; i < axes.size(); ++i) { + dim_value = out_dims[axes[i]]; + start = starts[i] < 0 ? (starts[i] + dim_value) : starts[i]; + end = ends[i] < 0 ? (ends[i] + dim_value) : ends[i]; + start = std::max(start, 0); + end = std::max(end, 0); + start = std::min(start, dim_value); + end = std::min(end, dim_value); + start = std::min(start, end); + out_dims[axes[i]] = end - start; + } + ctx->SetOutputDim("Out", out_dims); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Input")->type()), + ctx.GetPlace()); + } +}; + +class SliceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Input", "Tensor of data to extract slices from."); + AddOutput("Out", "Sliced data tensor."); + + AddAttr>( + "axes", + "(list) Axes that `starts` and `ends` apply to. It's optional." + "If not present, will be treated as [0, 1, ..., len(`starts`) - 1]."); + AddAttr>( + "starts", + "(list) Starting indices of corresponding axis in `axes`"); + AddAttr>( + "ends", + "(list) Starting indices of corresponding axis in `axes`."); + + AddComment(R"DOC( +Slice Operator. + +Produces a slice of the input tensor along multiple axes. Similar to numpy: +https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html +Slice uses `axes`, `starts` and `ends` attributes to specify the start and +end dimension for each axis in the list of axes, it uses this information +to slice the input data tensor. If a negative value is passed for any of +the start or end indices, it represents number of elements before the end +of that dimension. If the value passed to start or end is larger than +the n (the number of elements in this dimension), it represents n. +For slicing to the end of a dimension with unknown size, it is recommended +to pass in INT_MAX. If axes are omitted, they are set to [0, ..., ndim-1]. + + Example 1: + Given: + data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] + axes = [0, 1] + starts = [1, 0] + ends = [2, 3] + Then: + result = [ [5, 6, 7], ] + + Example 2: + Given: + data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] + starts = [0, 1] + ends = [-1, 1000] + Then: + result = [ [2, 3, 4], ] +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(slice, ops::SliceOp, ops::SliceOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL( + slice, ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel); diff --git a/paddle/fluid/operators/slice_op.cu b/paddle/fluid/operators/slice_op.cu new file mode 100644 index 0000000000..8c1767c70b --- /dev/null +++ b/paddle/fluid/operators/slice_op.cu @@ -0,0 +1,22 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/slice_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + slice, ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel, + ops::SliceKernel); diff --git a/paddle/fluid/operators/slice_op.h b/paddle/fluid/operators/slice_op.h new file mode 100644 index 0000000000..ba231aee17 --- /dev/null +++ b/paddle/fluid/operators/slice_op.h @@ -0,0 +1,88 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class SliceKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + int rank = ctx.Input("Input")->dims().size(); + switch (rank) { + case 1: + SliceCompute<1>(ctx); + break; + case 2: + SliceCompute<2>(ctx); + break; + case 3: + SliceCompute<3>(ctx); + break; + case 4: + SliceCompute<4>(ctx); + break; + case 5: + SliceCompute<5>(ctx); + break; + case 6: + SliceCompute<6>(ctx); + break; + } + } + + private: + template + void SliceCompute(const framework::ExecutionContext& context) const { + auto& place = + *context.template device_context().eigen_device(); + auto in = context.Input("Input"); + auto out = context.Output("Out"); + out->mutable_data(context.GetPlace()); + auto out_dims = out->dims(); + auto in_dims = in->dims(); + auto axes = context.Attr>("axes"); + auto starts = context.Attr>("starts"); + + auto offsets = Eigen::array(); + auto extents = Eigen::array(); + for (size_t i = 0; i < D; ++i) { + offsets[i] = 0; + extents[i] = out_dims[i]; + } + int start; + for (size_t i = 0; i < axes.size(); ++i) { + start = starts[i]; + if (start < 0) { + start = (start + in_dims[axes[i]]); + } + start = std::max(start, 0); + offsets[axes[i]] = start; + } + auto in_t = + framework::EigenTensor::From( + *in); + auto out_t = + framework::EigenTensor::From( + *out); + out_t.device(place) = in_t.slice(offsets, extents); + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index cc256aa627..847b3cbd1b 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -49,6 +49,9 @@ class SoftmaxOp : public framework::OperatorWithKernel { const framework::ExecutionContext& ctx) const override { // choose cudnn kernel if the runtime supported. framework::LibraryType library_{framework::LibraryType::kPlain}; + std::string data_format = ctx.Attr("data_format"); + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; @@ -58,6 +61,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { if (library_ == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { library_ = framework::LibraryType::kMKLDNN; + layout_ = framework::DataLayout::kMKLDNN; } #endif @@ -68,9 +72,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { "float16 can only be used on GPU place"); } - std::string data_format = ctx.Attr("data_format"); - return framework::OpKernelType(input_data_type, ctx.GetPlace(), - framework::StringToDataLayout(data_format), + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_, library_); } }; @@ -81,7 +83,8 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("X", "The input tensor of softmax. " "2-D with shape [batch_size, input_feature_dimensions]."); - AddOutput("Out", "The normalized values with the same shape as X."); + AddOutput("Out", "The normalized values with the same shape as X.") + .Reuse("X"); AddAttr( "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") @@ -142,6 +145,7 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { const framework::ExecutionContext& ctx) const override { // choose cudnn kernel if the runtime supported. framework::LibraryType library_{framework::LibraryType::kPlain}; + #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { library_ = framework::LibraryType::kCUDNN; diff --git a/paddle/fluid/operators/split_op.cc b/paddle/fluid/operators/split_op.cc index 5e2b2a9945..d661b276bc 100644 --- a/paddle/fluid/operators/split_op.cc +++ b/paddle/fluid/operators/split_op.cc @@ -115,4 +115,7 @@ USE_CPU_ONLY_OP(concat); REGISTER_OPERATOR(split, ops::SplitOp, ops::SplitOpMaker, ops::SplitGradMaker); REGISTER_OP_CPU_KERNEL(split, - ops::SplitOpKernel); + ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel); diff --git a/paddle/fluid/operators/split_op.cu.cc b/paddle/fluid/operators/split_op.cu.cc index efa378af85..18e0904681 100644 --- a/paddle/fluid/operators/split_op.cu.cc +++ b/paddle/fluid/operators/split_op.cu.cc @@ -15,4 +15,7 @@ limitations under the License. */ #include "paddle/fluid/operators/split_op.h" namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( - split, ops::SplitOpKernel); + split, ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel); diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index bcc5e22d4a..863baba9ea 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -115,7 +115,7 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "(vector) The input tensors of sum operator.") .AsDuplicable(); - AddOutput("Out", "(Tensor) The output tensor of sum operator."); + AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X"); AddComment(R"DOC( Sum operator. diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc index 83e768b4dc..4b1208c437 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.cc +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -17,22 +17,93 @@ #include "paddle/fluid/operators/tensorrt_engine_op.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/engine.h" #include "paddle/fluid/inference/utils/singleton.h" namespace paddle { namespace operators { +using inference::Singleton; +using inference::tensorrt::TRT_EngineManager; + +using FluidDT = framework::proto::VarType_Type; +using TRT_DT = nvinfer1::DataType; + +namespace { + +TRT_DT FluidDataType2TRT(FluidDT type) { + switch (type) { + case FluidDT::VarType_Type_FP32: + return TRT_DT::kFLOAT; + case FluidDT::VarType_Type_INT32: + return TRT_DT::kINT32; + default: + return TRT_DT::kINT32; + } + PADDLE_THROW("unkown type"); + return TRT_DT::kINT32; +} + +nvinfer1::Dims Vec2TRT_Dims(const std::vector &shape) { + PADDLE_ENFORCE_GT(shape.size(), 1UL, + "TensorRT' tensor input requires at least 2 dimensions"); + PADDLE_ENFORCE_LE(shape.size(), 4UL, + "TensorRT' tensor input requires at most 4 dimensions"); + + switch (shape.size()) { + case 2: + return nvinfer1::Dims2(shape[0], shape[1]); + case 3: + return nvinfer1::Dims3(shape[0], shape[1], shape[2]); + case 4: + return nvinfer1::Dims4(shape[0], shape[1], shape[2], shape[3]); + default: + return nvinfer1::Dims(); + } + return nvinfer1::Dims(); +} + +} // namespace + template void paddle::operators::TensorRTEngineKernel::Prepare( const framework::ExecutionContext &context) const { + VLOG(4) << "Prepare engine"; // Get the ProgramDesc and pass to convert. - const auto &block = context.Attr("subgraph"); + framework::proto::BlockDesc block_desc; + block_desc.ParseFromString(context.Attr("subgraph")); max_batch_ = context.Attr("max_batch"); auto max_workspace = context.Attr("max_workspace"); - engine_.reset(new inference::tensorrt::TensorRTEngine( - max_batch_, max_workspace, nullptr)); + engine_ = Singleton::Global().Create( + max_batch_, max_workspace, &stream_); + engine_->InitNetwork(); + + framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); + // Add inputs + VLOG(4) << "declare inputs"; + for (auto &input : context.Inputs("Xs")) { + VLOG(4) << "declare input " << input; + auto *var = block.FindVar(input); + PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR, + "TensorRT engine only takes LoDTensor as input"); + auto shape = var->GetShape(); + engine_->DeclareInput( + input, FluidDataType2TRT( + var->Proto()->type().lod_tensor().tensor().data_type()), + Vec2TRT_Dims(var->GetShape())); + } + + // TODO(Superjomn) parameters should be passed after analysised from outside. inference::Singleton::Global().ConvertBlock( - block, engine_.get()); + block_desc, {}, context.scope(), engine_); + + // Add outputs + VLOG(4) << "declare outputs"; + for (auto &output : context.Outputs("Ys")) { + VLOG(4) << "declare output " << output; + engine_->DeclareOutput(output); + } + engine_->FreezeNetwork(); } @@ -41,7 +112,9 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("Xs", "A list of inputs.").AsDuplicable(); AddOutput("Ys", "A list of outputs").AsDuplicable(); - AddAttr("subgraph", "the subgraph"); + AddAttr("subgraph", "the subgraph."); + AddAttr("max_batch", "the maximum batch size."); + AddAttr("max_workspace", "the maximum batch size."); AddComment("TensorRT engine operator."); } }; diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index fe273d386c..4b089601ff 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -32,9 +32,12 @@ class TensorRTEngineOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { + auto input0 = ctx.Inputs("Xs").front(); framework::OpKernelType kt = framework::OpKernelType( - framework::ToDataType( - ctx.Input("pre_ids")->type()), + framework::ToDataType(ctx.scope() + .FindVar(input0) + ->GetMutable() + ->type()), platform::CPUPlace()); return kt; } @@ -50,17 +53,16 @@ class TensorRTEngineKernel : public framework::OpKernel { auto input_names = context.op().Inputs("Xs"); PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs"); // Try to determine a batch_size - auto* tensor0 = context.Input(input_names.front()); - PADDLE_ENFORCE_NOT_NULL(tensor0); - int batch_size = tensor0->dims()[0]; + auto& tensor0 = inference::analysis::GetFromScope( + context.scope(), input_names.front()); + int batch_size = tensor0.dims()[0]; PADDLE_ENFORCE_LE(batch_size, max_batch_); // Convert input tensor from fluid to engine. for (const auto& x : context.Inputs("Xs")) { // convert input and copy to TRT engine's buffer - auto* v = context.scope().FindVar(x); - PADDLE_ENFORCE_NOT_NULL(v, "no variable called %s", x); - auto& t = v->Get(); + auto& t = inference::analysis::GetFromScope( + context.scope(), x); if (platform::is_cpu_place(t.place())) { engine_->SetInputFromCPU(x, static_cast(t.data()), t.memory_size()); @@ -86,13 +88,18 @@ class TensorRTEngineKernel : public framework::OpKernel { fluid_t->Resize(framework::make_ddim(ddim)); auto size = inference::analysis::AccuDims(dims.d, dims.nbDims); if (platform::is_cpu_place(fluid_t->place())) { + // TODO(Superjomn) change this float to dtype size. engine_->GetOutputInCPU( - y, fluid_t->mutable_data(platform::CPUPlace()), size); + y, fluid_t->mutable_data(platform::CPUPlace()), + size * sizeof(float)); } else { engine_->GetOutputInGPU( - y, fluid_t->mutable_data(platform::CUDAPlace()), size); + y, fluid_t->mutable_data(platform::CUDAPlace()), + size * sizeof(float)); } } + + cudaStreamSynchronize(stream_); } protected: @@ -100,7 +107,8 @@ class TensorRTEngineKernel : public framework::OpKernel { void Prepare(const framework::ExecutionContext& context) const; private: - mutable std::unique_ptr engine_; + mutable cudaStream_t stream_; + mutable inference::tensorrt::TensorRTEngine* engine_{nullptr}; mutable int max_batch_{0}; }; diff --git a/paddle/fluid/operators/tensorrt_engine_op_test.cc b/paddle/fluid/operators/tensorrt_engine_op_test.cc new file mode 100644 index 0000000000..6f383de259 --- /dev/null +++ b/paddle/fluid/operators/tensorrt_engine_op_test.cc @@ -0,0 +1,152 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_desc.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +USE_CPU_ONLY_OP(tensorrt_engine); + +namespace paddle { +namespace operators { + +namespace { +void CreateCPUTensor(framework::Scope* scope, const std::string& name, + const std::vector& shape) { + auto* var = scope->Var(name); + auto* tensor = var->GetMutable(); + auto dims = framework::make_ddim(shape); + tensor->Resize(dims); + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + inference::tensorrt::RandomizeTensor(tensor, place, ctx); +} + +void AddTensorToBlockDesc(framework::proto::BlockDesc* block, + const std::string& name, + const std::vector& shape) { + using framework::proto::VarType; + auto* var = block->add_vars(); + framework::VarDesc desc(name); + desc.SetType(VarType::LOD_TENSOR); + desc.SetDataType(VarType::FP32); + desc.SetShape(shape); + *var = *desc.Proto(); +} + +template +void SetAttr(framework::proto::OpDesc* op, const std::string& name, + const T& data); + +template <> +void SetAttr(framework::proto::OpDesc* op, const std::string& name, + const std::string& data) { + auto* attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::STRING); + attr->set_s(data); +} +template <> +void SetAttr(framework::proto::OpDesc* op, const std::string& name, + const int& data) { + auto* attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::INT); + attr->set_i(data); +} +template <> +void SetAttr(framework::proto::OpDesc* op, const std::string& name, + const int64_t& data) { + auto* attr = op->add_attrs(); + attr->set_name(name); + attr->set_type(paddle::framework::proto::AttrType::LONG); + attr->set_l(data); +} + +} // namespace + +TEST(TensorRTEngineOp, manual) { + framework::ProgramDesc program; + auto* block_ = program.Proto()->add_blocks(); + block_->set_idx(0); + block_->set_parent_idx(-1); + + LOG(INFO) << "create block desc"; + framework::BlockDesc block_desc(&program, block_); + LOG(INFO) << "create mul op"; + auto* mul = block_desc.AppendOp(); + mul->SetType("mul"); + mul->SetInput("X", std::vector({"x"})); // 2 x 4 + mul->SetInput("Y", std::vector({"y"})); // 4 x 6 + mul->SetOutput("Out", std::vector({"z"})); // 2 x 6 + + LOG(INFO) << "create fc op"; + auto* fc = block_desc.AppendOp(); + fc->SetType("mul"); + fc->SetInput("X", std::vector({"z"})); + fc->SetInput("Y", std::vector({"y0"})); // 6 x 8 + fc->SetOutput("Out", std::vector({"z0"})); // 2 x 8 + + // Set inputs' variable shape in BlockDesc + AddTensorToBlockDesc(block_, "x", std::vector({2, 4})); + AddTensorToBlockDesc(block_, "y", std::vector({4, 6})); + AddTensorToBlockDesc(block_, "y0", std::vector({6, 8})); + AddTensorToBlockDesc(block_, "z", std::vector({2, 6})); + + // It is wired, need to copy manually. + *block_->add_ops() = *mul->Proto(); + *block_->add_ops() = *fc->Proto(); + + ASSERT_EQ(block_->ops_size(), 2); + + LOG(INFO) << "create tensorrt desc"; + framework::OpDesc engine_op_desc(nullptr); + engine_op_desc.SetType("tensorrt_engine"); + engine_op_desc.SetInput("Xs", std::vector({"x", "y", "y0"})); + engine_op_desc.SetOutput("Ys", std::vector({"z0"})); + SetAttr(engine_op_desc.Proto(), "subgraph", + block_->SerializeAsString()); + SetAttr(engine_op_desc.Proto(), "max_batch", 30); + SetAttr(engine_op_desc.Proto(), "max_workspace", 1 << 10); + + LOG(INFO) << "create engine op"; + auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto()); + + framework::Scope scope; + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + // Prepare variables. + CreateCPUTensor(&scope, "x", std::vector({2, 4})); + CreateCPUTensor(&scope, "y", std::vector({4, 6})); + CreateCPUTensor(&scope, "z", std::vector({2, 6})); + + CreateCPUTensor(&scope, "y0", std::vector({6, 8})); + CreateCPUTensor(&scope, "z0", std::vector({2, 8})); + + // Execute them. + LOG(INFO) << "engine_op run"; + engine_op->Run(scope, place); +} + +} // namespace operators +} // namespace paddle + +USE_TRT_CONVERTER(mul) +USE_TRT_CONVERTER(fc) diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc index 719f039a0f..5015b10055 100644 --- a/paddle/fluid/operators/test_send_nccl_id.cc +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -20,13 +20,18 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/macros.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/nccl_helper.h" #include "paddle/fluid/string/printf.h" +#ifdef PADDLE_WITH_GRPC +#include "paddle/fluid/operators/send_recv_util.h" +#endif + USE_NO_KERNEL_OP(listen_and_serv); namespace f = paddle::framework; @@ -35,42 +40,43 @@ namespace m = paddle::operators::math; namespace detail = paddle::operators::detail; namespace string = paddle::string; -std::unique_ptr rpc_service; +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; -void StartServer(std::atomic* initialized) { +void StartServer() { f::Scope scope; p::CPUPlace place; scope.Var(NCCL_ID_VARNAME); p::DeviceContextPool& pool = p::DeviceContextPool::Instance(); auto& dev_ctx = *pool.Get(p::CPUPlace()); - rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", true)); - f::ProgramDesc empty_program; f::Executor executor(dev_ctx.GetPlace()); - rpc_service->SetScope(&scope); - rpc_service->SetDevCtx(&dev_ctx); - rpc_service->SetProgram(&empty_program); - rpc_service->SetExecutor(&executor); + g_req_handler->SetScope(&scope); + g_req_handler->SetDevCtx(&dev_ctx); + g_req_handler->SetProgram(&empty_program); + g_req_handler->SetExecutor(&executor); + + g_rpc_service->RegisterRPC(detail::kRequestSend, g_req_handler.get()); + g_req_handler->SetRPCServer(g_rpc_service.get()); std::thread server_thread( - std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, rpc_service.get())); - *initialized = true; - rpc_service->SetCond(0); - auto recv = rpc_service->Get(); + std::bind(&detail::RPCServer::StartServer, g_rpc_service.get())); + + g_rpc_service->SetCond(detail::kRequestSend); + g_rpc_service->WaitBarrier(detail::kRequestSend); + LOG(INFO) << "got nccl id and stop server..."; - rpc_service->ShutDown(); + g_rpc_service->ShutDown(); server_thread.join(); } -TEST(SendNcclId, DISABLED_Normal) { - std::atomic initialized{false}; - std::thread server_thread(StartServer, &initialized); - while (!initialized) { - } - // wait server to start - // sleep(2); - rpc_service->WaitServerReady(); +TEST(SendNcclId, RPCServer) { + g_req_handler.reset(new detail::RequestSendHandler(true)); + g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); + + std::thread server_thread(StartServer); + g_rpc_service->WaitServerReady(); f::Scope scope; p::CPUPlace place; @@ -78,17 +84,22 @@ TEST(SendNcclId, DISABLED_Normal) { auto& dev_ctx = *pool.Get(p::CPUPlace()); auto var = scope.Var(NCCL_ID_VARNAME); - // var->SetType(f::proto::VarType_Type_RAW); auto id = var->GetMutable(); p::dynload::ncclGetUniqueId(id); - int port = rpc_service->GetSelectedPort(); + int port = g_rpc_service->GetSelectedPort(); + std::string ep = string::Sprintf("127.0.0.1:%d", port); - detail::RPCClient client; - client.AsyncSendVariable(ep, dev_ctx, scope, NCCL_ID_VARNAME); - client.Wait(); + detail::RPCClient* client = detail::RPCClient::GetInstance(); + + LOG(INFO) << "connect to server" << ep; + client->AsyncSendVar(ep, dev_ctx, scope, NCCL_ID_VARNAME); + client->Wait(); + client->AsyncSendBatchBarrier(ep); + client->Wait(); + server_thread.join(); - auto* ptr = rpc_service.release(); - delete ptr; + g_rpc_service.reset(nullptr); + g_req_handler.reset(nullptr); } diff --git a/paddle/fluid/operators/top_k_op.cc b/paddle/fluid/operators/top_k_op.cc index c17d1afc30..4a8ac441cf 100644 --- a/paddle/fluid/operators/top_k_op.cc +++ b/paddle/fluid/operators/top_k_op.cc @@ -50,7 +50,7 @@ class TopkOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "(Tensor) The input of Topk op"); - AddOutput("Out", "(Tensor) The output tensor of Topk op"); + AddOutput("Out", "(Tensor) The output tensor of Topk op").Reuse("X"); AddOutput("Indices", "(Tensor) The indices of Topk elements of input"); AddComment(R"DOC( Top K operator diff --git a/paddle/fluid/platform/assert.h b/paddle/fluid/platform/assert.h index 123d3598f4..2ce9b31bb8 100644 --- a/paddle/fluid/platform/assert.h +++ b/paddle/fluid/platform/assert.h @@ -17,7 +17,7 @@ limitations under the License. */ #define STRINGIFY(x) #x #define TOSTRING(x) STRINGIFY(x) -#if defined(__APPLE__) && defined(__CUDA_ARCH__) && !defined(NDEBUG) +#if defined(__CUDA_ARCH__) #include #define PADDLE_ASSERT(e) \ do { \ @@ -38,6 +38,9 @@ limitations under the License. */ } while (0) #else #include -#define PADDLE_ASSERT(e) assert(e) +// For cuda, the assertions can affect performance and it is therefore +// recommended to disable them in production code +// https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#assertion +#define PADDLE_ASSERT(e) assert((e)) #define PADDLE_ASSERT_MSG(e, m) assert((e) && (m)) #endif diff --git a/paddle/fluid/platform/cpu_info.cc b/paddle/fluid/platform/cpu_info.cc index 4fc9aae8e3..40dc7c9a0b 100644 --- a/paddle/fluid/platform/cpu_info.cc +++ b/paddle/fluid/platform/cpu_info.cc @@ -21,12 +21,17 @@ limitations under the License. */ #include #endif +#include #include "gflags/gflags.h" DEFINE_double(fraction_of_cpu_memory_to_use, 1, "Default use 100% of CPU memory for PaddlePaddle," "reserve the rest for page tables, etc"); +DEFINE_uint64( + initial_cpu_memory_in_mb, 500, + "Default initial 500MB of CPU memory for PaddlePaddle, in MD unit."); + DEFINE_double( fraction_of_cuda_pinned_memory_to_use, 0.5, "Default use 50% of CPU memory as the pinned_memory for PaddlePaddle," @@ -54,7 +59,10 @@ inline size_t CpuTotalPhysicalMemory() { size_t CpuMaxAllocSize() { // For distributed systems, it requires configuring and limiting // the fraction of memory to use. - return FLAGS_fraction_of_cpu_memory_to_use * CpuTotalPhysicalMemory(); + return std::min( + static_cast(FLAGS_fraction_of_cpu_memory_to_use * + CpuTotalPhysicalMemory()), + static_cast(FLAGS_initial_cpu_memory_in_mb * 1 << 20)); } size_t CpuMinChunkSize() { diff --git a/paddle/fluid/platform/cudnn_helper.h b/paddle/fluid/platform/cudnn_helper.h index c0d399d078..6ea4f8b7cb 100644 --- a/paddle/fluid/platform/cudnn_helper.h +++ b/paddle/fluid/platform/cudnn_helper.h @@ -22,6 +22,8 @@ limitations under the License. */ #include "paddle/fluid/platform/float16.h" #include "paddle/fluid/platform/macros.h" +DECLARE_bool(cudnn_deterministic); + namespace paddle { namespace platform { @@ -76,8 +78,44 @@ enum class DataLayout { // Not use enum class PoolingMode { kMaximum, kAverage, + kMaximumDeterministic, }; +#if CUDNN_VERSION < 6000 +#pragma message "CUDNN version under 6.0 is supported at best effort." +#pragma message "We strongly encourage you to move to 6.0 and above." +#pragma message "This message is intended to annoy you enough to update." +#pragma message \ + "please see https://docs.nvidia.com/deeplearning/sdk/cudnn-release-notes/" + +inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { + switch (mode) { + case PoolingMode::kMaximumDeterministic: + return CUDNN_POOLING_MAX; + case PoolingMode::kAverage: + return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kMaximum: + return CUDNN_POOLING_MAX; + default: + PADDLE_THROW("Unexpected pooling mode."); + } +} +#else + +inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { + switch (mode) { + case PoolingMode::kMaximumDeterministic: + return CUDNN_POOLING_MAX_DETERMINISTIC; + case PoolingMode::kAverage: + return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kMaximum: + return CUDNN_POOLING_MAX; + default: + PADDLE_THROW("Unexpected pooling mode."); + } +} +#endif // CUDNN_VERSION < 6000 + template class CudnnDataType; @@ -293,9 +331,7 @@ class ScopedPoolingDescriptor { PADDLE_ENFORCE_EQ(kernel.size(), pads.size()); PADDLE_ENFORCE_EQ(kernel.size(), strides.size()); PADDLE_ENFORCE(dynload::cudnnSetPoolingNdDescriptor( - desc_, (mode == PoolingMode::kMaximum - ? CUDNN_POOLING_MAX - : CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING), + desc_, (GetPoolingMode(mode)), CUDNN_PROPAGATE_NAN, // Always propagate nans. kernel.size(), kernel.data(), pads.data(), strides.data())); return desc_; diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index 1f733d71bd..6c50ab2685 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -175,7 +175,6 @@ CUDADeviceContext::~CUDADeviceContext() { Place CUDADeviceContext::GetPlace() const { return place_; } void CUDADeviceContext::Wait() const { - std::lock_guard guard(mutex_); PADDLE_ENFORCE(cudaStreamSynchronize(stream_)); PADDLE_ENFORCE(cudaGetLastError()); } diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index a9c1984616..292ffef1ae 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -11,6 +11,7 @@ limitations under the License. */ #pragma once #include +#include // NOLINT #include #include #include @@ -100,7 +101,7 @@ class CUDADeviceContext : public DeviceContext { template void RecordEvent(cudaEvent_t ev, Callback callback) { - std::lock_guard guard(mutex_); + std::lock_guard guard(mtx_); callback(); PADDLE_ENFORCE(cudaEventRecord(ev, stream_)); } @@ -110,8 +111,6 @@ class CUDADeviceContext : public DeviceContext { std::unique_ptr eigen_device_; std::unique_ptr eigen_stream_; - - mutable std::recursive_mutex mutex_; cudaStream_t stream_; cudnnHandle_t cudnn_handle_; cublasHandle_t cublas_handle_; @@ -119,6 +118,8 @@ class CUDADeviceContext : public DeviceContext { int compute_capability; int multi_process; int max_threads_per_mp; + + std::mutex mtx_; }; template <> diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index 1a9be044e0..d9e2afadaf 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -322,7 +322,6 @@ class DeviceTracerImpl : public DeviceTracer { DisableActivity(); dynload::cuptiUnsubscribe(subscriber_); CUPTI_CALL(dynload::cuptiGetTimestamp(&end_ns_)); - PADDLE_ENFORCE(dynload::cuptiFinalize()); enabled_ = false; } diff --git a/paddle/fluid/platform/dynload/cublas.h b/paddle/fluid/platform/dynload/cublas.h index 81acaff87d..25bcda7eed 100644 --- a/paddle/fluid/platform/dynload/cublas.h +++ b/paddle/fluid/platform/dynload/cublas.h @@ -45,7 +45,7 @@ extern void *cublas_dso_handle; std::call_once(cublas_dso_flag, []() { \ cublas_dso_handle = paddle::platform::dynload::GetCublasDsoHandle(); \ }); \ - void *p_##__name = dlsym(cublas_dso_handle, #__name); \ + static void *p_##__name = dlsym(cublas_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h index 34d83e3956..77e46fa768 100644 --- a/paddle/fluid/platform/dynload/cudnn.h +++ b/paddle/fluid/platform/dynload/cudnn.h @@ -39,7 +39,7 @@ extern void EnforceCUDNNLoaded(const char* fn_name); cudnn_dso_handle = paddle::platform::dynload::GetCUDNNDsoHandle(); \ }); \ EnforceCUDNNLoaded(#__name); \ - void* p_##__name = dlsym(cudnn_dso_handle, #__name); \ + static void* p_##__name = dlsym(cudnn_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/cupti.h b/paddle/fluid/platform/dynload/cupti.h index e64de7c20f..e8f4a82ef1 100644 --- a/paddle/fluid/platform/dynload/cupti.h +++ b/paddle/fluid/platform/dynload/cupti.h @@ -45,7 +45,7 @@ extern void *cupti_dso_handle; std::call_once(cupti_dso_flag, []() { \ cupti_dso_handle = paddle::platform::dynload::GetCUPTIDsoHandle(); \ }); \ - void *p_##__name = dlsym(cupti_dso_handle, #__name); \ + static void *p_##__name = dlsym(cupti_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ @@ -72,7 +72,6 @@ extern void *cupti_dso_handle; __macro(cuptiGetResultString); \ __macro(cuptiActivityGetNumDroppedRecords); \ __macro(cuptiActivityFlushAll); \ - __macro(cuptiFinalize); \ __macro(cuptiSubscribe); \ __macro(cuptiUnsubscribe); \ __macro(cuptiEnableCallback); \ diff --git a/paddle/fluid/platform/dynload/curand.h b/paddle/fluid/platform/dynload/curand.h index 46ad4379d5..5b9e0820e0 100644 --- a/paddle/fluid/platform/dynload/curand.h +++ b/paddle/fluid/platform/dynload/curand.h @@ -34,7 +34,7 @@ extern void *curand_dso_handle; std::call_once(curand_dso_flag, []() { \ curand_dso_handle = paddle::platform::dynload::GetCurandDsoHandle(); \ }); \ - void *p_##__name = dlsym(curand_dso_handle, #__name); \ + static void *p_##__name = dlsym(curand_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/nccl.h b/paddle/fluid/platform/dynload/nccl.h index 37902ae20c..575516f818 100644 --- a/paddle/fluid/platform/dynload/nccl.h +++ b/paddle/fluid/platform/dynload/nccl.h @@ -37,7 +37,7 @@ extern void* nccl_dso_handle; std::call_once(nccl_dso_flag, []() { \ nccl_dso_handle = paddle::platform::dynload::GetNCCLDsoHandle(); \ }); \ - void* p_##__name = dlsym(nccl_dso_handle, #__name); \ + static void* p_##__name = dlsym(nccl_dso_handle, #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/dynload/tensorrt.h b/paddle/fluid/platform/dynload/tensorrt.h index f584a49da0..5d67658b94 100644 --- a/paddle/fluid/platform/dynload/tensorrt.h +++ b/paddle/fluid/platform/dynload/tensorrt.h @@ -40,7 +40,7 @@ extern void* tensorrt_dso_handle; paddle::platform::dynload::GetTensorRtDsoHandle(); \ PADDLE_ENFORCE(tensorrt_dso_handle, "load tensorrt so failed"); \ }); \ - void* p_##__name = dlsym(tensorrt_dso_handle, #__name); \ + static void* p_##__name = dlsym(tensorrt_dso_handle, #__name); \ PADDLE_ENFORCE(p_##__name, "load %s failed", #__name); \ return reinterpret_cast(p_##__name)(args...); \ } \ diff --git a/paddle/fluid/platform/dynload/warpctc.h b/paddle/fluid/platform/dynload/warpctc.h index 7c70649d21..d157c1fda7 100644 --- a/paddle/fluid/platform/dynload/warpctc.h +++ b/paddle/fluid/platform/dynload/warpctc.h @@ -40,7 +40,7 @@ extern void* warpctc_dso_handle; std::call_once(warpctc_dso_flag, []() { \ warpctc_dso_handle = paddle::platform::dynload::GetWarpCTCDsoHandle(); \ }); \ - void* p_##_name = dlsym(warpctc_dso_handle, #__name); \ + static void* p_##_name = dlsym(warpctc_dso_handle, #__name); \ return reinterpret_cast(p_##_name)(args...); \ } \ }; \ diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index f1187620d8..de711b7d23 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/place.h" namespace paddle { namespace platform { @@ -86,5 +87,17 @@ inline mkldnn::memory::data_type MKLDNNGetDataType() { return mkldnn::memory::f32; } +inline void Reorder(const mkldnn::memory& src, const mkldnn::memory& dst) { + auto reorder_prim = mkldnn::reorder(src, dst); + std::vector pipeline; + pipeline.push_back(reorder_prim); + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); +} + +inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) { + return static_cast( + memory.get_primitive_desc().desc().data.format); +} + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 09367889a9..cc46c88fd1 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -15,6 +15,7 @@ #pragma once #include +#include #include // NOLINT #include #include @@ -40,6 +41,11 @@ inline ncclDataType_t ToNCCLDataType(std::type_index type) { } } +// NOTE(minqiyang): according to the ncclGroupEnd documentations: +// https://docs.nvidia.com/deeplearning/sdk/nccl-api/ncclapidoc.html, +// ncclGroupEnd will wait for all communicators to be initialized, which will +// cause blocking problem when a runtime_error was thrown, so try only guard +// NCCL actions when use it. class NCCLGroupGuard { public: static std::mutex &NCCLMutex() { diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 3d8d64e4c2..01de9d7041 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -127,6 +127,7 @@ double Event::CpuElapsedMs(const Event& e) const { double Event::CudaElapsedMs(const Event& e) const { #ifdef PADDLE_WITH_CUDA + if (!has_cuda_) return 0.0; PADDLE_ENFORCE(e.has_cuda() && has_cuda()); PADDLE_ENFORCE(e.device() == device()); PADDLE_ENFORCE(cudaEventSynchronize(event_)); diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 3af8941be6..bd5c613f8c 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -413,6 +413,9 @@ All parameter, weight, gradient are variables in Paddle. py::class_(m, "Executor") .def(py::init()) +#ifdef PADDLE_WITH_DISTRIBUTE + .def("complete", &Executor::Complete) +#endif .def("run", (void (Executor::*)(const ProgramDesc &, Scope *, int, bool, bool)) & Executor::Run); @@ -509,16 +512,24 @@ All parameter, weight, gradient are variables in Paddle. self.num_threads_ = num_threads; }) .def_property( - "use_event", - [](const ExecutionStrategy &self) { return self.use_event_; }, - [](ExecutionStrategy &self, bool use_event) { - self.use_event_ = use_event; + "use_cuda", + [](const ExecutionStrategy &self) { return self.use_cuda_; }, + [](ExecutionStrategy &self, bool use_cuda) { + self.use_cuda_ = use_cuda; }) .def_property( "allow_op_delay", [](const ExecutionStrategy &self) { return self.allow_op_delay_; }, [](ExecutionStrategy &self, bool allow_op_delay) { self.allow_op_delay_ = allow_op_delay; + }) + .def_property( + "num_iteration_per_drop_scope", + [](const ExecutionStrategy &self) { + return self.num_iteration_per_drop_scope_; + }, + [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) { + self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope; }); py::class_ build_strategy(pe, "BuildStrategy"); @@ -545,6 +556,12 @@ All parameter, weight, gradient are variables in Paddle. [](BuildStrategy &self, BuildStrategy::GradientScaleStrategy strategy) { self.gradient_scale_ = strategy; + }) + .def_property( + "debug_graphviz_path", + [](const BuildStrategy &self) { return self.debug_graphviz_path_; }, + [](BuildStrategy &self, const std::string &path) { + self.debug_graphviz_path_ = path; }); pe.def(py::init &, diff --git a/paddle/fluid/recordio/chunk.cc b/paddle/fluid/recordio/chunk.cc index 82d9aa601c..6c65d9160c 100644 --- a/paddle/fluid/recordio/chunk.cc +++ b/paddle/fluid/recordio/chunk.cc @@ -119,40 +119,56 @@ bool Chunk::Write(std::ostream& os, Compressor ct) const { } bool Chunk::Parse(std::istream& sin) { - Header hdr; - bool ok = hdr.Parse(sin); + ChunkParser parser(sin); + if (!parser.Init()) { + return false; + } + Clear(); + while (parser.HasNext()) { + Add(parser.Next()); + } + return true; +} + +ChunkParser::ChunkParser(std::istream& sin) : in_(sin) {} +bool ChunkParser::Init() { + pos_ = 0; + bool ok = header_.Parse(in_); if (!ok) { return ok; } - auto beg_pos = sin.tellg(); - uint32_t crc = Crc32Stream(sin, hdr.CompressSize()); - PADDLE_ENFORCE_EQ(hdr.Checksum(), crc); - Clear(); - sin.seekg(beg_pos, sin.beg); - std::unique_ptr compressed_stream; - switch (hdr.CompressType()) { + auto beg_pos = in_.tellg(); + uint32_t crc = Crc32Stream(in_, header_.CompressSize()); + PADDLE_ENFORCE_EQ(header_.Checksum(), crc); + in_.seekg(beg_pos, in_.beg); + + switch (header_.CompressType()) { case Compressor::kNoCompress: break; case Compressor::kSnappy: - compressed_stream.reset(new snappy::iSnappyStream(sin)); + compressed_stream_.reset(new snappy::iSnappyStream(in_)); break; default: PADDLE_THROW("Not implemented"); } + return true; +} - std::istream& stream = compressed_stream ? *compressed_stream : sin; +bool ChunkParser::HasNext() const { return pos_ < header_.NumRecords(); } - for (uint32_t i = 0; i < hdr.NumRecords(); ++i) { - uint32_t rec_len; - stream.read(reinterpret_cast(&rec_len), sizeof(uint32_t)); - std::string buf; - buf.resize(rec_len); - stream.read(&buf[0], rec_len); - PADDLE_ENFORCE_EQ(rec_len, stream.gcount()); - Add(buf); +std::string ChunkParser::Next() { + if (!HasNext()) { + return ""; } - return true; + ++pos_; + std::istream& stream = compressed_stream_ ? *compressed_stream_ : in_; + uint32_t rec_len; + stream.read(reinterpret_cast(&rec_len), sizeof(uint32_t)); + std::string buf; + buf.resize(rec_len); + stream.read(&buf[0], rec_len); + PADDLE_ENFORCE_EQ(rec_len, stream.gcount()); + return buf; } - } // namespace recordio } // namespace paddle diff --git a/paddle/fluid/recordio/chunk.h b/paddle/fluid/recordio/chunk.h index 71a1556a33..cfb954a591 100644 --- a/paddle/fluid/recordio/chunk.h +++ b/paddle/fluid/recordio/chunk.h @@ -13,6 +13,7 @@ // limitations under the License. #pragma once +#include #include #include @@ -53,9 +54,20 @@ class Chunk { DISABLE_COPY_AND_ASSIGN(Chunk); }; -size_t CompressData(const char* in, size_t in_length, Compressor ct, char* out); +class ChunkParser { + public: + explicit ChunkParser(std::istream& sin); + + bool Init(); + std::string Next(); + bool HasNext() const; -void DeflateData(const char* in, size_t in_length, Compressor ct, char* out); + private: + Header header_; + uint32_t pos_{0}; + std::istream& in_; + std::unique_ptr compressed_stream_; +}; } // namespace recordio } // namespace paddle diff --git a/paddle/fluid/recordio/scanner.cc b/paddle/fluid/recordio/scanner.cc index 88b4d4001b..06a13e6c5b 100644 --- a/paddle/fluid/recordio/scanner.cc +++ b/paddle/fluid/recordio/scanner.cc @@ -22,35 +22,33 @@ namespace paddle { namespace recordio { Scanner::Scanner(std::unique_ptr &&stream) - : stream_(std::move(stream)) { + : stream_(std::move(stream)), parser_(*stream_) { Reset(); } -Scanner::Scanner(const std::string &filename) { - stream_.reset(new std::ifstream(filename)); +Scanner::Scanner(const std::string &filename) + : stream_(new std::ifstream(filename)), parser_(*stream_) { Reset(); } void Scanner::Reset() { stream_->clear(); stream_->seekg(0, std::ios::beg); - ParseNextChunk(); + parser_.Init(); } std::string Scanner::Next() { - PADDLE_ENFORCE(!eof_, "StopIteration"); - auto rec = cur_chunk_.Record(offset_++); - if (offset_ == cur_chunk_.NumRecords()) { - ParseNextChunk(); + if (stream_->eof()) { + return ""; } - return rec; -} -void Scanner::ParseNextChunk() { - eof_ = !cur_chunk_.Parse(*stream_); - offset_ = 0; + auto res = parser_.Next(); + if (!parser_.HasNext() && HasNext()) { + parser_.Init(); + } + return res; } -bool Scanner::HasNext() const { return !eof_; } +bool Scanner::HasNext() const { return !stream_->eof(); } } // namespace recordio } // namespace paddle diff --git a/paddle/fluid/recordio/scanner.h b/paddle/fluid/recordio/scanner.h index 34f1b0c78d..0d885dd87a 100644 --- a/paddle/fluid/recordio/scanner.h +++ b/paddle/fluid/recordio/scanner.h @@ -37,11 +37,7 @@ class Scanner { private: std::unique_ptr stream_; - Chunk cur_chunk_; - size_t offset_; - bool eof_; - - void ParseNextChunk(); + ChunkParser parser_; }; } // namespace recordio } // namespace paddle diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h index f3d8b1a39e..854e4baa39 100644 --- a/paddle/math/MathFunctions.h +++ b/paddle/math/MathFunctions.h @@ -12,8 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#ifndef MATHFUNCTIONS_H_ -#define MATHFUNCTIONS_H_ +#pragma once #ifdef PADDLE_WITH_MKLML #include @@ -21,7 +20,7 @@ limitations under the License. */ #include #endif -#if defined(PADDLE_USE_VECLIB) +#ifdef PADDLE_USE_VECLIB extern "C" { #include #include @@ -30,8 +29,10 @@ extern "C" { #ifdef PADDLE_USE_OPENBLAS #include +#ifdef LAPACK_FOUND #include #endif +#endif #ifndef LAPACK_FOUND extern "C" { @@ -126,5 +127,3 @@ template void vTanh(const int n, const T* a, T* r); } // namespace paddle - -#endif // MATHFUNCTIONS_H_ diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index 8eeea1805d..e8b3053267 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -132,7 +132,8 @@ EOF -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \ -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \ - -DWITH_CONTRIB=${WITH_CONTRIB:-ON} + -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \ + -DWITH_ANAKIN=ON } function abort(){ @@ -145,19 +146,17 @@ function check_style() { trap 'abort' 0 set -e - # install glide - curl https://glide.sh/get | bash - eval "$(GIMME_GO_VERSION=1.8.3 gimme)" + if [ -x "$(command -v gimme)" ]; then + eval "$(GIMME_GO_VERSION=1.8.3 gimme)" + fi # set up go environment for running gometalinter mkdir -p $GOPATH/src/github.com/PaddlePaddle/ ln -sf ${PADDLE_ROOT} $GOPATH/src/github.com/PaddlePaddle/Paddle - cd $GOPATH/src/github.com/PaddlePaddle/Paddle/go; glide install; cd - + mkdir -p ./build/go + cp go/glide.* build/go + cd build/go; glide install; cd - - go get github.com/alecthomas/gometalinter - gometalinter --install - - cd ${PADDLE_ROOT} export PATH=/usr/bin:$PATH pre-commit install clang-format --version @@ -183,6 +182,7 @@ function build() { ============================================ EOF make clean + make -j `nproc` make install -j `nproc` } @@ -449,7 +449,7 @@ EOF # run paddle version to install python packages first RUN apt-get update &&\ ${NCCL_DEPS}\ - apt-get install -y wget python-pip dmidecode python-tk && easy_install -U pip && \ + apt-get install -y wget python-pip python-opencv libgtk2.0-dev dmidecode python-tk && easy_install -U pip && \ pip install /*.whl; apt-get install -f -y && \ apt-get clean -y && \ rm -f /*.whl && \ diff --git a/paddle/testing/paddle_gtest_main.cc b/paddle/testing/paddle_gtest_main.cc index 586ec48477..507479c862 100644 --- a/paddle/testing/paddle_gtest_main.cc +++ b/paddle/testing/paddle_gtest_main.cc @@ -30,7 +30,7 @@ int main(int argc, char** argv) { new_argv.push_back( strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory")); #else - new_argv.push_back(strdup("--tryfromenv=use_pinned_memory")); + new_argv.push_back(strdup("--tryfromenv=use_pinned_memory,use_mkldnn")); #endif int new_argc = static_cast(new_argv.size()); char** new_argv_address = new_argv.data(); diff --git a/python/paddle/batch.py b/python/paddle/batch.py index 317cf037c6..3c6a53db3c 100644 --- a/python/paddle/batch.py +++ b/python/paddle/batch.py @@ -15,7 +15,7 @@ __all__ = ['batch'] -def batch(reader, batch_size): +def batch(reader, batch_size, drop_last=True): """ Create a batched reader. @@ -23,6 +23,8 @@ def batch(reader, batch_size): :type reader: callable :param batch_size: size of each mini-batch :type batch_size: int + :param drop_last: drop the last batch, if the size of last batch is not equal to batch_size. + :type drop_last: bool :return: the batched reader. :rtype: callable """ @@ -35,7 +37,7 @@ def batch(reader, batch_size): if len(b) == batch_size: yield b b = [] - if b: + if drop_last == False and len(b) != 0: yield b return batch_reader diff --git a/python/paddle/dataset/flowers.py b/python/paddle/dataset/flowers.py index f082e33be3..527044b415 100644 --- a/python/paddle/dataset/flowers.py +++ b/python/paddle/dataset/flowers.py @@ -119,7 +119,8 @@ def reader_creator(data_file, yield sample, int(label) - 1 if use_xmap: - return xmap_readers(mapper, reader, cpu_count(), buffered_size) + cpu_num = int(os.environ.get('CPU_NUM', cpu_count())) + return xmap_readers(mapper, reader, cpu_num, buffered_size) else: return map_readers(mapper, reader) diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index 859605d005..bd985ad733 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -26,6 +26,7 @@ from trainer import BeginEpochEvent from trainer import EndEpochEvent from trainer import BeginStepEvent from trainer import EndStepEvent +from trainer import CheckpointConfig import inferencer from inferencer import Inferencer @@ -44,8 +45,8 @@ import transpiler from param_attr import ParamAttr, WeightNormParamAttr from data_feeder import DataFeeder from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace -from transpiler import DistributeTranspiler, SimpleDistributeTranspiler, \ - InferenceTranspiler, memory_optimize, release_memory +from transpiler import DistributeTranspiler, InferenceTranspiler, \ + memory_optimize, release_memory from concurrency import (Go, make_channel, channel_send, channel_recv, channel_close, Select) from lod_tensor import create_lod_tensor, create_random_int_lodtensor @@ -116,11 +117,11 @@ def __bootstrap__(): read_env_flags = [ 'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir', - 'eager_delete_scope' + 'eager_delete_scope', 'use_mkldnn' ] if core.is_compiled_with_cuda(): read_env_flags += [ - 'fraction_of_gpu_memory_to_use', 'cudnn_algo_use_autotune' + 'fraction_of_gpu_memory_to_use', 'cudnn_deterministic' ] core.init_gflags([sys.argv[0]] + ["--tryfromenv=" + ",".join(read_env_flags)]) diff --git a/python/paddle/fluid/data_feeder.py b/python/paddle/fluid/data_feeder.py index 7940dabcfb..e2013137b1 100644 --- a/python/paddle/fluid/data_feeder.py +++ b/python/paddle/fluid/data_feeder.py @@ -15,6 +15,7 @@ from __future__ import print_function import core import numpy +import os import six.moves as six import multiprocessing @@ -150,7 +151,9 @@ class DataFeeder(object): elif isinstance(self.place, core.CUDAPlace): return core.get_cuda_device_count() else: - return multiprocessing.cpu_count() + cpu_num = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + return cpu_num def decorate_reader(self, reader, diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index 93aa5f908e..33d8f70941 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -170,6 +170,8 @@ def get_program_cache_key(feed, fetch_list): return var.desc.name() elif isinstance(var, str): return var + elif isinstance(var, basestring): + return str(var) else: raise TypeError(str(var) + " should be Variable or str") diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 33b5caa0ea..f6438c82ac 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -72,6 +72,8 @@ def convert_np_dtype_to_dtype_(np_dtype): return core.VarDesc.VarType.INT64 elif dtype == np.bool: return core.VarDesc.VarType.BOOL + elif dtype == np.uint16: + return core.VarDesc.VarType.INT16 elif dtype == np.uint8: return core.VarDesc.VarType.UINT8 else: @@ -361,6 +363,13 @@ class OpProtoHolder(object): raise ValueError("Operator \"%s\" has not been registered." % type) return self.op_proto_map[type] + @staticmethod + def generated_op_attr_names(): + return { + core.op_proto_and_checker_maker.kOpRoleAttrName(), + core.op_proto_and_checker_maker.kOpRoleVarAttrName() + } + class Operator(object): """ @@ -368,6 +377,13 @@ class Operator(object): Block. Users can use the build in instructions to describe their neural network. """ + OP_WITHOUT_KERNEL_SET = { + 'feed', 'fetch', 'save', 'load', 'recurrent', 'go', + 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv', + 'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine', + 'ncclInit', 'channel_create', 'channel_close', 'channel_send', + 'channel_recv', 'select', 'gen_nccl_id' + } def __init__(self, block, @@ -504,17 +520,13 @@ class Operator(object): else: self.desc.set_attr(attr_name, self.attrs[attr_name]) self.desc.check_attrs() - no_kernel_op_set = { - 'feed', 'fetch', 'save', 'load', 'recurrent', 'go', - 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', - 'recv', 'listen_and_serv', 'parallel_do', 'save_combine', - 'load_combine', 'ncclInit', 'channel_create', 'channel_close', - 'channel_send', 'channel_recv', 'select', 'gen_nccl_id' - } - if type not in no_kernel_op_set: + if self.has_kernel(type): self.desc.infer_var_type(self.block.desc) self.desc.infer_shape(self.block.desc) + def has_kernel(self, op_type): + return op_type not in self.OP_WITHOUT_KERNEL_SET + def to_string(self, throw_on_error): """ To debug string. @@ -742,7 +754,9 @@ class Block(object): def var(self, name): if not isinstance(name, basestring): - raise TypeError() + raise TypeError( + "var require string as parameter, but get %s instead." % + (type(name))) v = self.vars.get(name, None) if v is None: raise ValueError("var %s not in this block" % name) diff --git a/python/paddle/fluid/inferencer.py b/python/paddle/fluid/inferencer.py index 9f242cf29a..6baac00905 100644 --- a/python/paddle/fluid/inferencer.py +++ b/python/paddle/fluid/inferencer.py @@ -56,6 +56,8 @@ class Inferencer(object): else: self.exe = executor.Executor(self.place) + self.inference_program = self.inference_program.clone(for_test=True) + def infer(self, inputs, return_numpy=True): """ :param inputs: a map of {"input_name": input_var} that will be feed into the inference program diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index 8e58e5eb79..6323c9899e 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -24,7 +24,8 @@ __all__ = [ 'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params', 'load_persistables', 'save_inference_model', 'load_inference_model', 'get_inference_program', 'save_checkpoint', 'load_checkpoint', - 'clean_checkpoint' + 'clean_checkpoint', 'load_persist_vars_without_grad', + 'save_persist_vars_without_grad', 'get_latest_checkpoint_serial' ] @@ -457,95 +458,161 @@ def get_parameter_value_by_name(name, executor, program=None): SUCCESS_MARK_FILENAME = "_SUCCESS" CHECKPOINT_PREFIX = "checkpoint" +MODEL_DIR = "__model__" +TRAINER_PREFIX = "trainer" CHECKPOINT_SEPARATOR = "_" def save_checkpoint(executor, - checkpoint_dir=None, - max_num_checkpoints=3, - save_interval_secs=600, - main_program=None): + checkpoint_dir, + trainer_id, + trainer_args=None, + main_program=None, + max_num_checkpoints=3): """ Save Checkpoint will save persistable LodTensor variables from main_program in checkpoint directory, the directory named by serial number from 0 to (n -1), save_checkpoint use LRU strategy to keep numbers of checkpoint directory, the numbers of checkpoint directory are max_num_checkpoints at most, The interval between two saved checkpoints must greater than save_interval_secs. - :param executor - :param checkpoint_dir - :param max_num_checkpoints - :param save_interval_secs - :param main_program + :param executor executor for save the value + :param checkpoint_dir the checkpoint directory + :param trainer_id currect trainer id, if id is equal to 0, the trainer is chief + :param main_program will save all variables in program + :param max_num_checkpoints will keep numbers of checkpoint serials not bigger than max_num_checkpoints """ if checkpoint_dir is None: - checkpoint_dir = os.getcwd() + raise ValueError("'checkpoint_dir' should not be None") + + if trainer_args: + assert isinstance(trainer_args, dict) if not os.path.isdir(checkpoint_dir): os.makedirs(checkpoint_dir) - serial = _get_lastest_checkpoint_dir(checkpoint_dir) - if serial >= 0 and not _interval_secs_exceed( - _get_serial_dir(serial, checkpoint_dir), save_interval_secs): - return + serial = get_latest_checkpoint_serial(checkpoint_dir) + 1 + cur_dir = _get_serial_dir(checkpoint_dir, serial) - serial += 1 - cur_dir = _get_serial_dir(serial, checkpoint_dir) + save_trainer_args(cur_dir, trainer_id, trainer_args) - save_vars( - executor, - dirname=cur_dir, - main_program=main_program, - vars=None, - predicate=_is_checkpoint_var, - filename=None) - _write_success(cur_dir) - _lru_delete(checkpoint_dir, max_num_checkpoints) + if trainer_id == 0: + save_persist_vars_without_grad(executor, cur_dir, main_program) + + _scroll_delete(checkpoint_dir, max_num_checkpoints) -def load_checkpoint(executor, checkpoint_dir=None, main_program=None): +def load_checkpoint(executor, checkpoint_dir, serial, main_program): """ Load checkpoint from a directory by executor, it will find the most recent saved checkpoint file and load it auto. - :param executor - :param checkpoint_dir - :param main_program + :param executor executor for load the value + :param checkpoint_dir the checkpoint directory + :param serial the serial folder in checkpoint directory will be load + :param main_program will load all variables in program """ if checkpoint_dir is None: - checkpoint_dir = os.getcwd() + raise ValueError("'checkpoint_dir' should not be None") - serial = _get_lastest_checkpoint_dir(checkpoint_dir) + if serial is None or serial < 0: + raise ValueError("'serial' should not be None or <0 ") - if serial < 0: - return + if main_program is None: + raise ValueError('main_program should not be None.') - cur_dir = _get_serial_dir(serial, checkpoint_dir) - - load_vars( - executor, - dirname=cur_dir, - main_program=main_program, - predicate=_is_checkpoint_var, - filename=None) + cur_dir = _get_serial_dir(checkpoint_dir, serial) + load_persist_vars_without_grad(executor, cur_dir, main_program, True) def clean_checkpoint(checkpoint_dir, delete_dir=False): """ clean the checkpoint dir, when the train exits normally, the trainer will call clean_checkpoint to delete checkpoint directory saved before. delete_dir only works when the directory is empty, otherwise, OSError is raised. + + :param checkpoint_dir + :param delete_dir """ + if checkpoint_dir is None: - checkpoint_dir = os.getcwd() - _lru_delete(checkpoint_dir, max_num_checkpoints=0) + raise ValueError("'checkpoint_dir' should not be None") + _scroll_delete(checkpoint_dir, max_num_checkpoints=0) if delete_dir and not os.listdir(checkpoint_dir): os.rmdir(checkpoint_dir) -def _get_serial_dir(serial, checkpoint_dir): - serial_folder = CHECKPOINT_PREFIX + CHECKPOINT_SEPARATOR + str(serial) - return os.path.join(checkpoint_dir, serial_folder) +def load_persist_vars_without_grad(executor, + dirname, + program, + has_model_dir=False): + """ + load_persist_vars_without_grad will load variables from a directory by an executor, + the variable named end with "@GRAD" will not be loaded. + + :param executor executor for load the value + :param dirname the checkpoint directory + :param program will load all variables in program + :param has_model_dir if has_model_dir is True, will load variables from sub directory named __model__ + """ + + if has_model_dir: + dirname = _get_model_dir(dirname) + + load_vars( + executor, + dirname=dirname, + main_program=program, + predicate=_is_checkpoint_var, + filename=None) + + +def save_persist_vars_without_grad(executor, dirname, program): + """ + save_persist_vars_without_grad will save variables to a directory by an executor, + the variable named end with "@GRAD" will not be saved. + + :param executor executor for load the value + :param dirname the checkpoint directory + :param program will load all variables in program + """ + cur_dir = _get_model_dir(dirname) + save_vars( + executor, + dirname=cur_dir, + main_program=program, + vars=None, + predicate=_is_checkpoint_var, + filename=None) + _write_success(cur_dir) + + +def save_trainer_args(dirname, trainer_id, trainer_args): + assert isinstance(trainer_args, dict) + + cur_dir = _get_trainer_dir(dirname, trainer_id) + + for name, value in trainer_args.iteritems(): + args_file = os.path.join(cur_dir, name) + with open(args_file, 'w') as f: + f.write(str(value)) + _write_success(cur_dir) + + +def load_trainer_args(checkpoint_dir, serial, trainer_id, trainer_args): + assert isinstance(trainer_args, list) + + cur_dir = _get_serial_dir(checkpoint_dir, serial) + cur_dir = _get_trainer_dir(cur_dir, trainer_id) + + ret_values = [] + + for arg in trainer_args: + cur_file = os.path.join(cur_dir, arg) + with open(cur_file, 'r') as f: + contents = f.read() + ret_values.append(contents.strip()) + return ret_values def _is_checkpoint_var(var): @@ -559,36 +626,74 @@ def _is_checkpoint_var(var): var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \ var.desc.type() == core.VarDesc.VarType.RAW: return False + # @GRAD are named for gradient variables, checkpoint will not save it. + if "@GRAD" in var.name: + return False + # .trainer_ are named for distribute train variables, checkpoint will not save it. + if ".trainer_" in var.name: + return False - if var.name.endswith("@GRAD"): + # .block is named for distribute train variables, checkpoint will not save it. + if ".block" in var.name: return False return var.persistable -def _interval_secs_exceed(dirname, save_interval_secs): - dir_time = os.path.getmtime(dirname) - if save_interval_secs > (time.time() - dir_time): - return False - return True +def _get_dir_serial(dirname): + _, serial = dirname.split(CHECKPOINT_SEPARATOR) + + try: + serial_num = int(serial) + except ValueError: + serial_num = -1 + return serial_num + + +def _get_serial_dir(dirname, serial): + serial_folder = CHECKPOINT_PREFIX + CHECKPOINT_SEPARATOR + str(serial) + serial_dir = os.path.join(dirname, serial_folder) + + if not os.path.isdir(serial_dir): + os.makedirs(serial_dir) + + return serial_dir + +def _get_model_dir(dirname): + model_dir = os.path.join(dirname, MODEL_DIR) -def _lru_delete(dirname, max_num_checkpoints=3): + if not os.path.isdir(model_dir): + os.makedirs(model_dir) + + return model_dir + + +def _get_trainer_dir(dirname, trainer_id): + trainer_folder = TRAINER_PREFIX + CHECKPOINT_SEPARATOR + str(trainer_id) + trainer_dir = os.path.join(dirname, trainer_folder) + + if not os.path.isdir(trainer_dir): + os.makedirs(trainer_dir) + + return trainer_dir + + +def _scroll_delete(dirname, max_num_checkpoints=3): dirs = os.listdir(dirname) - serials = [] + serial_map = {} for serial in dirs: - try: - serials.append(int(serial)) - except ValueError: - continue + serial_num = _get_dir_serial(serial) + serial_map[serial_num] = serial - if len(serials) <= max_num_checkpoints: + if len(serial_map.keys()) <= max_num_checkpoints: return + serials = serial_map.keys() serials.sort(reverse=True) serials = serials[max_num_checkpoints:] for serial in serials: - cur_dir = os.path.join(dirname, str(serial)) + cur_dir = _get_serial_dir(dirname, serial) shutil.rmtree(cur_dir) @@ -604,33 +709,30 @@ def _write_success(dirname): f.write(now) -def _get_lastest_checkpoint_dir(checkpoint_dir): +def get_latest_checkpoint_serial(checkpoint_dir): """ get the latest file in checkpoint directory, the _SUCCESS file must exist in the directory :param checkpoint_dir """ - if not checkpoint_dir.strip(): + if not checkpoint_dir: return -1 def has_success(checkpoint_dir, cur_dir): """ is _SUCCESS in this dir """ - _, serial = cur_dir.split(CHECKPOINT_SEPARATOR) - - try: - int(serial) - except ValueError: - return -1 - if not os.path.isdir(os.path.join(checkpoint_dir, cur_dir)): + serial = _get_dir_serial(cur_dir) + if serial == -1 or not os.path.isdir( + os.path.join(checkpoint_dir, cur_dir)): return -1 success_path = os.path.join( - _get_serial_dir(serial, checkpoint_dir), SUCCESS_MARK_FILENAME) + _get_serial_dir(checkpoint_dir, serial), MODEL_DIR, + SUCCESS_MARK_FILENAME) if os.path.isfile(success_path): - return int(serial) + return serial if not os.path.isdir(checkpoint_dir): return -1 diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index d1ea9f1485..4db085e9f5 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -13,7 +13,7 @@ # limitations under the License. import contextlib -from layer_function_generator import autodoc +from layer_function_generator import autodoc, templatedoc from tensor import assign, fill_constant from .. import core from ..framework import Program, Variable, Operator @@ -721,26 +721,22 @@ def lod_rank_table(x, level=0): return table +@templatedoc() def max_sequence_len(rank_table): - """Max Sequence Len Operator. Given a LoDRankTable object, this layer - returns the max length of a batch of sequences. In fact, a LoDRankTable - object contains a list of tuples() and - the list is already sorted by sequence length in descending order, so the - operator just returns the sequence length of the first tuple element. + """ + ${comment} + + >>> import paddle.fluid as fluid + >>> x = fluid.layers.data(name='x', shape=[10], dtype='float32', + >>> lod_level=1) + >>> rank_table = layers.lod_rank_table(x=x, level=0) + >>> max_seq_len = layers.max_sequence_len(rank_table) Args: - rank_table (Variable): Input variable which is a LoDRankTable object. + rank_table(${rank_table_type}): ${rank_table_comment}. Returns: - Variable: The max length of sequence. - - Examples: - .. code-block:: python - - x = fluid.layers.data(name='x', shape=[10], - dtype='float32', lod_level=1) - rank_table = layers.lod_rank_table(x=x, level=0) - max_seq_len = layers.max_sequence_len(rank_table) + ${out_comment}. """ helper = LayerHelper("max_seqence_len", **locals()) res = helper.create_tmp_variable(dtype="int64") @@ -1213,6 +1209,34 @@ class IfElseBlockGuard(object): class IfElse(object): + """ + if-else control flow. + + Args: + cond (Variable): condition used to compare. + name (str, default None): The name of this layer. + + Examples: + .. code-block:: python + + limit = fluid.layers.fill_constant_batch_size_like( + input=label, dtype='int64', shape=[1], value=5.0) + cond = fluid.layers.less_than(x=label, y=limit) + ie = fluid.layers.IfElse(cond) + with ie.true_block(): + true_image = ie.input(image) + hidden = fluid.layers.fc(input=true_image, size=100, act='tanh') + prob = fluid.layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + + with ie.false_block(): + false_image = ie.input(image) + hidden = fluid.layers.fc( + input=false_image, size=200, act='tanh') + prob = fluid.layers.fc(input=hidden, size=10, act='softmax') + ie.output(prob) + prob = ie() + """ OUT_IF_ELSE_BLOCKS = 0 IN_IF_ELSE_TRUE_BLOCKS = 1 IN_IF_ELSE_FALSE_BLOCKS = 2 diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 8758ac9f94..f3aeb6cd75 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -19,11 +19,12 @@ from ..unique_name import generate as unique_name from control_flow import BlockGuard from ..layer_helper import LayerHelper from ..executor import global_scope +from layer_function_generator import generate_layer_fn, templatedoc __all__ = [ 'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file', 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer', - 'random_data_generator', 'Preprocessor' + 'random_data_generator', 'Preprocessor', 'load' ] @@ -434,7 +435,7 @@ def open_files(filenames, shapes, lod_levels, dtypes, - thread_num, + thread_num=1, buffer_size=None, pass_num=1, for_parallel=True): @@ -586,6 +587,26 @@ def read_file(file_obj): class Preprocessor(object): + """ + A block for data pre-processing in reader. + + Args: + reader (Variable): A reader variable. + name (str, default None): The name of the reader. + + Examples: + .. code-block:: python + + preprocessor = fluid.layers.io.Preprocessor(reader=reader) + with preprocessor.block(): + img, lbl = preprocessor.inputs() + img_out = img / 2 + lbl_out = lbl + 1 + preprocessor.outputs(img_out, lbl_out) + + data_file = fluid.layers.io.double_buffer(preprocessor()) + + """ BEFORE_SUB_BLOCK = 0 IN_SUB_BLOCK = 1 AFTER_SUB_BLOCK = 2 @@ -662,3 +683,29 @@ class Preprocessor(object): "sink_var_names": self.sink_var_names }) return monkey_patch_reader_methods(self.reader) + + +@templatedoc() +def load(out, file_path, load_as_fp16=None): + """ + ${comment} + + >>> import paddle.fluid as fluid + >>> tmp_tensor = fluid.layers.create_tensor(dtype='float32') + >>> fluid.layers.load(tmp_tensor, "./tmp_tensor.bin") + + Args: + out(${out_type}): ${out_comment}. + + file_path(${file_path_type}): ${file_path_comment}. + + load_as_fp16(${load_as_fp16_type}): ${load_as_fp16_comment}. + + Returns: + None + """ + helper = LayerHelper("load", **locals()) + attrs = {"file_path": file_path} + if load_as_fp16 is not None: + attrs['load_as_fp16'] = load_as_fp16 + helper.append_op(type="load", inputs={}, output={"Out": out}, args=attrs) diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 295d1b7190..cb60a3aec9 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -15,16 +15,13 @@ import re import cStringIO import functools import warnings +import string from ..proto import framework_pb2 from ..framework import OpProtoHolder, Variable from ..layer_helper import LayerHelper -__all__ = [ - 'deprecated', - 'generate_layer_fn', - 'autodoc', -] +__all__ = ['deprecated', 'generate_layer_fn', 'autodoc', 'templatedoc'] def _convert_(name): @@ -43,6 +40,10 @@ def _convert_(name): return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower() +def _type_to_str_(tp): + return framework_pb2.AttrType.Name(tp) + + def _generate_doc_string_(op_proto): """ Generate docstring by OpProto @@ -54,9 +55,6 @@ def _generate_doc_string_(op_proto): str: the document string """ - def _type_to_str_(tp): - return framework_pb2.AttrType.Name(tp) - if not isinstance(op_proto, framework_pb2.OpProto): raise TypeError("OpProto should be `framework_pb2.OpProto`") @@ -75,7 +73,11 @@ def _generate_doc_string_(op_proto): buf.write(str(each_input.dispensable)) buf.write('\n') + skip_attrs = OpProtoHolder.generated_op_attr_names() + for each_attr in op_proto.attrs: + if each_attr.name in skip_attrs: + continue buf.write(' ') buf.write(each_attr.name) buf.write(' (') @@ -220,3 +222,67 @@ def autodoc(comment=""): return func return __impl__ + + +_inline_math_single_dollar = re.compile(r"\$([^\$]+)\$") + + +def templatedoc(op_type=None): + """ + Decorator of layer function. It will use the docstring from the layer + function as the template. The template arguments are: + + * ${comment}: The operator comment written in CPP. + * ${{name}_comment}: The comment of ${name} written with AddAttr, AddOutput, + and AddInput. The ${name} is Python snake style. i.e., xxx_xxx. + * ${{name}_type}: The type of ${name}. + + Returns: + Decorated function. + """ + + def trim_ending_dot(msg): + return msg.rstrip('.') + + def escape_inline_math(msg): + return _inline_math_single_dollar.sub(repl=r':math:`\1`', string=msg) + + def __impl__(func): + if op_type is None: + op_type_name = func.__name__ + else: + op_type_name = op_type + op_proto = OpProtoHolder.instance().get_op_proto(op_type_name) + tmpl = string.Template(func.__doc__) + + comment_lines = op_proto.comment.split("\n") + comment = "" + for line in comment_lines: + line = line.strip() + if len(line) != 0: + comment += escape_inline_math(line) + comment += " " + elif len(comment) != 0: + comment += "\n \n " + + args = {"comment": trim_ending_dot(comment)} + for each_input in op_proto.inputs: + input_name = _convert_(each_input.name) + args["{0}_comment".format(input_name)] = trim_ending_dot( + each_input.comment) + args["{0}_type".format(input_name)] = "Variable" + for each_attr in op_proto.attrs: + input_name = _convert_(each_attr.name) + args["{0}_comment".format(input_name)] = trim_ending_dot( + each_attr.comment) + args["{0}_type".format(input_name)] = _type_to_str_(each_attr.type) + + for each_opt in op_proto.outputs: + output_name = _convert_(each_opt.name) + args["{0}_comment".format(output_name)] = trim_ending_dot( + each_opt.comment) + args["{0}_type".format(output_name)] = "Variable" + func.__doc__ = tmpl.substitute(args) + return func + + return __impl__ diff --git a/python/paddle/fluid/layers/learning_rate_scheduler.py b/python/paddle/fluid/layers/learning_rate_scheduler.py index d13c54daa5..716cc7824e 100644 --- a/python/paddle/fluid/layers/learning_rate_scheduler.py +++ b/python/paddle/fluid/layers/learning_rate_scheduler.py @@ -11,6 +11,14 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. +""" +When training a model, it's often useful to decay the +learning rate during training process, this is called +learning_rate_decay. There are many strategies to do +this, this module will provide some classical method. +User can also implement their own learning_rate_decay +strategy according to this module. +""" import control_flow import nn @@ -22,14 +30,6 @@ __all__ = [ 'exponential_decay', 'natural_exp_decay', 'inverse_time_decay', 'polynomial_decay', 'piecewise_decay', 'noam_decay' ] -""" -When training a model, it's often useful to decay the -learning rate during training process, this is called -learning_rate_decay. There are many strategies to do -this, this module will provide some classical method. -User can also implement their own learning_rate_decay -strategy according to this module. -""" def _decay_step_counter(begin=0): @@ -41,18 +41,20 @@ def _decay_step_counter(begin=0): def noam_decay(d_model, warmup_steps): - """Apply decay to learning rate. - ```python - lr_value = np.power(d_model, -0.5) * np.min([ - np.power(current_steps, -0.5), - np.power(warmup_steps, -1.5) * current_steps - ]) - ``` + """ + Noam decay method. The numpy implementation of noam decay as follows. + + >>> import numpy as np + >>> lr_value = np.power(d_model, -0.5) * np.min([ + >>> np.power(current_steps, -0.5), + >>> np.power(warmup_steps, -1.5) * current_steps]) + + Please reference `attention is all you need + `_. Args: d_model(Variable): The dimensionality of input and output of model. - Reference: attention is all you need - https://arxiv.org/pdf/1706.03762.pdf + warmup_steps(Variable): A super parameter. Returns: diff --git a/python/paddle/fluid/layers/metric.py b/python/paddle/fluid/layers/metric.py index cab2eb5551..a1c64ce277 100644 --- a/python/paddle/fluid/layers/metric.py +++ b/python/paddle/fluid/layers/metric.py @@ -64,10 +64,6 @@ def auc(input, label, curve='ROC', num_thresholds=200): topk_indices = helper.create_tmp_variable(dtype="int64") topk_out, topk_indices = nn.topk(input, k=k) auc_out = helper.create_tmp_variable(dtype="float32") - if correct is None: - correct = helper.create_tmp_variable(dtype="int64") - if total is None: - total = helper.create_tmp_variable(dtype="int64") helper.append_op( type="accuracy", inputs={ diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 004dcf7382..c3ff9b7725 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -12,16 +12,17 @@ # See the License for the specific language governing permissions and # limitations under the License. """ -All layers just related to the neural network. +All layers just related to the neural network. """ from ..layer_helper import LayerHelper from ..initializer import Normal, Constant from ..framework import Variable from ..param_attr import ParamAttr -from layer_function_generator import autodoc +from layer_function_generator import autodoc, templatedoc from tensor import concat import utils +import random __all__ = [ 'fc', @@ -38,13 +39,16 @@ __all__ = [ 'chunk_eval', 'sequence_conv', 'conv2d', + 'conv3d', 'sequence_pool', 'sequence_softmax', 'softmax', 'pool2d', + 'pool3d', 'batch_norm', 'beam_search_decode', 'conv2d_transpose', + 'conv3d_transpose', 'sequence_expand', 'lstm_unit', 'reduce_sum', @@ -82,8 +86,12 @@ __all__ = [ 'label_smooth', 'roi_pool', 'dice_loss', - 'upsampling_bilinear2d', + 'image_resize', + 'image_resize_short', + 'resize_bilinear', + 'gather', 'random_crop', + 'mean_iou', ] @@ -92,7 +100,6 @@ def fc(input, num_flatten_dims=1, param_attr=None, bias_attr=None, - use_cudnn=False, use_mkldnn=False, act=None, is_test=False, @@ -219,6 +226,7 @@ def embedding(input, have two elements which indicate the size of the dictionary of embeddings and the size of each embedding vector respectively. is_sparse(bool): The flag indicating whether to use sparse update. + is_distributed (bool): Whether to run lookup table from remote parameter server. padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup. Otherwise the given :attr:`padding_idx` indicates padding the output with zeros whenever lookup encounters it in :attr:`input`. If @@ -258,9 +266,10 @@ def embedding(input, return tmp -# TODO(qijun): expose H0 and C0 def dynamic_lstm(input, size, + h_0=None, + c_0=None, param_attr=None, bias_attr=None, use_peepholes=True, @@ -321,6 +330,13 @@ def dynamic_lstm(input, (T X 4D), where T is the total time steps in this mini-batch, D is the hidden size. size(int): 4 * hidden size. + h_0(Variable): The initial hidden state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size and D is the hidden size. + c_0(Variable): The initial cell state is an optional input, default is zero. + This is a tensor with shape (N x D), where N is the + batch size. `h_0` and `c_0` can be NULL but only at the same time. + param_attr(ParamAttr|None): The parameter attribute for the learnable hidden-hidden weights. @@ -384,12 +400,20 @@ def dynamic_lstm(input, cell = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype) batch_cell_pre_act = helper.create_tmp_variable(dtype) + inputs = {'Input': input, 'Weight': weight, 'Bias': bias} + batch_size = input.shape[0] + if h_0: + assert h_0.shape == (batch_size, size), \ + 'The shape of h0 should be (batch_size, %d)' % size + inputs['H0'] = h_0 + if c_0: + assert c_0.shape == (batch_size, size), \ + 'The shape of c0 should be (batch_size, %d)' % size + inputs['C0'] = c_0 helper.append_op( type='lstm', - inputs={'Input': input, - 'Weight': weight, - 'Bias': bias}, + inputs=inputs, outputs={ 'Hidden': hidden, 'Cell': cell, @@ -651,8 +675,9 @@ def dynamic_gru(input, :attr:`False`. gate_activation(str): The activation for update gate and reset gate. Choices = ["sigmoid", "tanh", "relu", "identity"], default "sigmoid". - activation(str): The activation for candidate hidden state. + candidate_activation(str): The activation for candidate hidden state. Choices = ["sigmoid", "tanh", "relu", "identity"], default "tanh". + h_0 (Variable): The hidden output of the first time step. Returns: Variable: The hidden state of GRU. The shape is :math:`(T \\times D)`, \ @@ -673,11 +698,13 @@ def dynamic_gru(input, attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) bias = helper.create_parameter( attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True) + batch_size = input.shape[0] inputs = {'Input': input, 'Weight': weight, 'Bias': bias} if h_0 != None: assert h_0.shape == ( - size, size), 'The shape of h0 should be(%d, %d)' % (size, size) - inputs['h0'] = h_0 + batch_size, size + ), 'The shape of h0 should be(batch_size, %d)' % size + inputs['H0'] = h_0 hidden = helper.create_tmp_variable(dtype) batch_gate = helper.create_tmp_variable(dtype) @@ -799,7 +826,22 @@ def gru_unit(input, return updated_hidden, reset_hidden_pre, gate +@templatedoc() def linear_chain_crf(input, label, param_attr=None): + """ + Linear Chain CRF. + + ${comment} + + Args: + input(${emission_type}): ${emission_comment} + label(${label_type}): ${label_comment} + param_attr(ParamAttr): The attribute of the learnable parameter. + + Returns: + ${log_likelihood_comment} + + """ helper = LayerHelper('linear_chain_crf', **locals()) size = input.shape[1] transition = helper.create_parameter( @@ -825,7 +867,19 @@ def linear_chain_crf(input, label, param_attr=None): return log_likelihood +@templatedoc() def crf_decoding(input, param_attr, label=None): + """ + ${comment} + + Args: + input(${emission_type}): ${emission_comment} + param_attr(ParamAttr): The parameter attribute for training. + label(${label_type}): ${label_comment} + + Returns: + ${viterbi_path_comment} + """ helper = LayerHelper('crf_decoding', **locals()) transition = helper.get_parameter(param_attr.name) viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype()) @@ -843,6 +897,13 @@ def cos_sim(X, Y): """ This function performs the cosine similarity between two tensors X and Y and returns that as the output. + + Args: + X (Variable): The input X. + Y (Variable): The input Y. + + Returns: + Variable: the output of cosine(X, Y). """ helper = LayerHelper('cos_sim', **locals()) out = helper.create_tmp_variable(dtype=X.dtype) @@ -869,15 +930,15 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None): unchanged. Args: - x(variable): The input tensor. - dropout_prob(float): Probability of setting units to zero. - is_test(bool): A flag indicating whether it is in test phrase or not. - seed(int): A Python integer used to create random seeds. If this - parameter is set to None, a random seed is used. - NOTE: If an integer seed is given, always the same output - units will be dropped. DO NOT use a fixed seed in training. - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + x (Variable): The input tensor. + dropout_prob (float): Probability of setting units to zero. + is_test (bool): A flag indicating whether it is in test phrase or not. + seed (int): A Python integer used to create random seeds. If this + parameter is set to None, a random seed is used. + NOTE: If an integer seed is given, always the same output + units will be dropped. DO NOT use a fixed seed in training. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: A tensor variable. @@ -999,8 +1060,8 @@ def square_error_cost(input, label): * :math:`Out`: Output value, same shape with :math:`X`. Args: - input(Variable): Input tensor, has predictions. - label(Variable): Label tensor, has target labels. + input (Variable): Input tensor, has predictions. + label (Variable): Label tensor, has target labels. Returns: Variable: The tensor variable storing the element-wise squared error \ @@ -1029,6 +1090,7 @@ def square_error_cost(input, label): return square_out +@templatedoc() def chunk_eval(input, label, chunk_scheme, @@ -1037,6 +1099,18 @@ def chunk_eval(input, """ This function computes and outputs the precision, recall and F1-score of chunk detection. + + Args: + input (Variable): prediction output of the network. + label (Variable): label of the test data set. + chunk_scheme (str): ${chunk_scheme_comment} + num_chunk_types (int): ${num_chunk_types_comment} + excluded_chunk_types (list): ${excluded_chunk_types_comment} + + Returns: + tuple: tuple containing: (precision, recall, f1_score, + num_infer_chunks, num_label_chunks, + num_correct_chunks) """ helper = LayerHelper("chunk_eval", **locals()) @@ -1069,6 +1143,7 @@ def chunk_eval(input, num_correct_chunks) +@templatedoc() def sequence_conv(input, num_filters, filter_size=3, @@ -1081,6 +1156,19 @@ def sequence_conv(input, This function creates the op for sequence_conv, using the inputs and other convolutional configurations for the filters and stride as given in the input parameters to the function. + + Args: + input (Variable): ${x_comment} + num_filters (int): number of filters. + filter_size (int): the filter size (H and W). + filter_stride (int): stride of the filter. + padding (bool): if True, add paddings. + bias_attr (ParamAttr|None): attributes for bias + param_attr (ParamAttr|None): attributes for parameter + act (str): the activation type + + Returns: + Variable: output of sequence_conv """ # FIXME(dzh) : want to unify the argument of python layer @@ -1180,48 +1268,49 @@ def conv2d(input, - Input: - Input shape: $(N, C_{in}, H_{in}, W_{in})$ + Input shape: :math:`(N, C_{in}, H_{in}, W_{in})` - Filter shape: $(C_{out}, C_{in}, H_f, W_f)$ + Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)` - Output: - Output shape: $(N, C_{out}, H_{out}, W_{out})$ + Output shape: :math:`(N, C_{out}, H_{out}, W_{out})` Where .. math:: - H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\ - W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1 - - Args: - input(Variable): The input image with [N, C, H, W] format. - num_filters(int): The number of filter. It is as same as the output - image channel. - filter_size(int|tuple|None): The filter size. If filter_size is a tuple, - it must contain two integers, (filter_size_H, filter_size_W). - Otherwise, the filter will be a square. - stride(int|tuple): The stride size. If stride is a tuple, it must - contain two integers, (stride_H, stride_W). Otherwise, the - stride_H = stride_W = stride. Default: stride = 1. - padding(int|tuple): The padding size. If padding is a tuple, it must - contain two integers, (padding_H, padding_W). Otherwise, the - padding_H = padding_W = padding. Default: padding = 0. - dilation(int|tuple): The dilation size. If dilation is a tuple, it must - contain two integers, (dilation_H, dilation_W). Otherwise, the - dilation_H = dilation_W = dilation. Default: dilation = 1. - groups(int): The groups number of the Conv2d Layer. According to grouped - convolution in Alex Krizhevsky's Deep CNN paper: when group=2, - the first half of the filters is only connected to the first half - of the input channels, while the second half of the filters is only - connected to the second half of the input channels. Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv2d Layer. Default: None - bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None - use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn - library is installed. Default: True - act(str): Activation type. Default: None - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\ + W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1 + + Args: + input (Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of filter. It is as same as the output + image channel. + filter_size (int|tuple|None): The filter size. If filter_size is a tuple, + it must contain two integers, (filter_size_H, filter_size_W). + Otherwise, the filter will be a square. + stride (int|tuple): The stride size. If stride is a tuple, it must + contain two integers, (stride_H, stride_W). Otherwise, the + stride_H = stride_W = stride. Default: stride = 1. + padding (int|tuple): The padding size. If padding is a tuple, it must + contain two integers, (padding_H, padding_W). Otherwise, the + padding_H = padding_W = padding. Default: padding = 0. + dilation (int|tuple): The dilation size. If dilation is a tuple, it must + contain two integers, (dilation_H, dilation_W). Otherwise, the + dilation_H = dilation_W = dilation. Default: dilation = 1. + groups (int): The groups number of the Conv2d Layer. According to grouped + convolution in Alex Krizhevsky's Deep CNN paper: when group=2, + the first half of the filters is only connected to the first half + of the input channels, while the second half of the filters is only + connected to the second half of the input channels. Default: groups=1 + param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None + bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not. + act (str): Activation type. Default: None + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The tensor variable storing the convolution and \ @@ -1239,8 +1328,6 @@ def conv2d(input, conv2d = fluid.layers.conv2d( input=data, num_filters=2, filter_size=3, act="relu") """ - if stride is None: - stride = [1, 1] num_channels = input.shape[1] @@ -1303,6 +1390,171 @@ def conv2d(input, return helper.append_activation(pre_act) +def conv3d(input, + num_filters, + filter_size, + stride=1, + padding=0, + dilation=1, + groups=None, + param_attr=None, + bias_attr=None, + use_cudnn=True, + use_mkldnn=False, + act=None, + name=None): + """ + **Convlution3D Layer** + + The convolution3D layer calculates the output based on the input, filter + and strides, paddings, dilations, groups parameters. Input(Input) and + Output(Output) are in NCDHW format. Where N is batch size C is the number of + channels, D is the depth of the feature, H is the height of the feature, + and W is the width of the feature. Convlution3D is similar with Convlution2D + but adds one dimension(depth). If bias attribution and activation type are + provided, bias is added to the output of the convolution, and the + corresponding activation function is applied to the final result. + + For each input :math:`X`, the equation is: + + .. math:: + + Out = \sigma (W \\ast X + b) + + In the above equation: + + * :math:`X`: Input value, a tensor with NCDHW format. + * :math:`W`: Filter value, a tensor with MCDHW format. + * :math:`\\ast`: Convolution operation. + * :math:`b`: Bias value, a 2-D tensor with shape [M, 1]. + * :math:`\\sigma`: Activation function. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be + different. + + Example: + + - Input: + + Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})` + + Filter shape: :math:`(C_{out}, C_{in}, D_f, H_f, W_f)` + + - Output: + Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})` + + Where + + .. math:: + + D_{out}&= \\frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{strides[0]} + 1 \\\\ + H_{out}&= \\frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{strides[1]} + 1 \\\\ + W_{out}&= \\frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{strides[2]} + 1 + + Args: + input (Variable): The input image with [N, C, D, H, W] format. + num_filters(int): The number of filter. It is as same as the output + image channel. + filter_size (int|tuple|None): The filter size. If filter_size is a tuple, + it must contain three integers, (filter_size_D, filter_size_H, filter_size_W). + Otherwise, the filter will be a square. + stride (int|tuple): The stride size. If stride is a tuple, it must + contain three integers, (stride_D, stride_H, stride_W). Otherwise, the + stride_D = stride_H = stride_W = stride. Default: stride = 1. + padding (int|tuple): The padding size. If padding is a tuple, it must + contain three integers, (padding_D, padding_H, padding_W). Otherwise, the + padding_D = padding_H = padding_W = padding. Default: padding = 0. + dilation (int|tuple): The dilation size. If dilation is a tuple, it must + contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the + dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1. + groups (int): The groups number of the Conv3d Layer. According to grouped + convolution in Alex Krizhevsky's Deep CNN paper: when group=2, + the first half of the filters is only connected to the first half + of the input channels, while the second half of the filters is only + connected to the second half of the input channels. Default: groups=1 + param_attr (ParamAttr): The parameters to the Conv3d Layer. Default: None + bias_attr (ParamAttr): Bias parameter for the Conv3d layer. Default: None + use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + use_mkldnn (bool): Use mkldnn kernels or not. + act (str): Activation type. Default: None + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The tensor variable storing the convolution and \ + non-linearity activation result. + + Raises: + ValueError: If the shapes of input, filter_size, stride, padding and + groups mismatch. + + Examples: + .. code-block:: python + + data = fluid.layers.data( + name='data', shape=[3, 12, 32, 32], dtype='float32') + conv2d = fluid.layers.conv3d( + input=data, num_filters=2, filter_size=3, act="relu") + """ + + l_type = 'conv3d' + + helper = LayerHelper(l_type, **locals()) + dtype = helper.input_dtype() + + num_channels = input.shape[1] + + if groups is None: + num_filter_channels = num_channels + else: + if num_channels % groups != 0: + raise ValueError("num_channels must be divisible by groups.") + num_filter_channels = num_channels / groups + + filter_size = utils.convert_to_list(filter_size, 3, 'filter_size') + stride = utils.convert_to_list(stride, 3, 'stride') + padding = utils.convert_to_list(padding, 3, 'padding') + dilation = utils.convert_to_list(dilation, 3, 'dilation') + + if not isinstance(use_cudnn, bool): + raise ValueError("use_cudnn should be True or False") + + input_shape = input.shape + filter_shape = [num_filters, num_filter_channels] + filter_size + + def _get_default_param_initializer(): + std = (2.0 / (filter_size[0]**3 * num_channels))**0.5 + return Normal(0.0, std, 0) + + filter_param = helper.create_parameter( + attr=helper.param_attr, + shape=filter_shape, + dtype=dtype, + default_initializer=_get_default_param_initializer()) + + pre_bias = helper.create_tmp_variable(dtype) + + helper.append_op( + type=l_type, + inputs={ + 'Input': input, + 'Filter': filter_param, + }, + outputs={"Output": pre_bias}, + attrs={ + 'strides': stride, + 'paddings': padding, + 'dilations': dilation, + 'groups': groups, + 'use_cudnn': use_cudnn, + 'use_mkldnn': use_mkldnn + }) + + pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) + + return helper.append_activation(pre_act) + + def sequence_pool(input, pool_type): """ This function add the operator for sequence pooling. @@ -1379,7 +1631,7 @@ def sequence_pool(input, pool_type): def sequence_first_step(input): """ - This funciton get the first step of sequence. + This function gets the first step of sequence. .. code-block:: text @@ -1412,7 +1664,7 @@ def sequence_first_step(input): def sequence_last_step(input): """ - This funciton get the last step of sequence. + This function gets the last step of sequence. .. code-block:: text @@ -1456,6 +1708,22 @@ def pool2d(input, """ This function adds the operator for pooling in 2 dimensions, using the pooling configurations mentioned in input parameters. + + Args: + input (Variable): ${input_comment} + pool_size (int): ${ksize_comment} + pool_type (str): ${pooling_type_comment} + pool_stride (int): stride of the pooling layer. + pool_padding (int): padding size. + global_pooling (bool): ${global_pooling_comment} + use_cudnn (bool): ${use_cudnn_comment} + ceil_mode (bool): ${ceil_mode_comment} + use_mkldnn (bool): ${use_mkldnn_comment} + name (str): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: output of pool2d layer. """ if pool_type not in ["max", "avg"]: raise ValueError( @@ -1474,12 +1742,84 @@ def pool2d(input, if not isinstance(use_cudnn, bool): raise ValueError("use_cudnn should be True or False") - helper = LayerHelper('pool2d', **locals()) + l_type = 'pool2d' + + helper = LayerHelper(l_type, **locals()) + dtype = helper.input_dtype() + pool_out = helper.create_tmp_variable(dtype) + + helper.append_op( + type=l_type, + inputs={"X": input}, + outputs={"Out": pool_out}, + attrs={ + "pooling_type": pool_type, + "ksize": pool_size, + "global_pooling": global_pooling, + "strides": pool_stride, + "paddings": pool_padding, + "use_cudnn": use_cudnn, + "ceil_mode": ceil_mode, + "use_mkldnn": use_mkldnn + }) + + return pool_out + + +def pool3d(input, + pool_size=-1, + pool_type="max", + pool_stride=1, + pool_padding=0, + global_pooling=False, + use_cudnn=True, + ceil_mode=False, + use_mkldnn=False, + name=None): + """ + This function adds the operator for pooling in 3-dimensions, using the + pooling configurations mentioned in input parameters. + + Args: + input (Variable): ${input_comment} + pool_size (int): ${ksize_comment} + pool_type (str): ${pooling_type_comment} + pool_stride (int): stride of the pooling layer. + pool_padding (int): padding size. + global_pooling (bool): ${global_pooling_comment} + use_cudnn (bool): ${use_cudnn_comment} + ceil_mode (bool): ${ceil_mode_comment} + use_mkldnn (bool): ${use_mkldnn_comment} + name (str): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: output of pool3d layer. + """ + if pool_type not in ["max", "avg"]: + raise ValueError( + "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.", + str(pool_type)) + + if global_pooling is False and pool_size == -1: + raise ValueError( + "When the global_pooling is False, pool_size must be passed " + "and be a valid value. Received pool_size: " + str(pool_size)) + + pool_size = utils.convert_to_list(pool_size, 3, 'pool_size') + pool_padding = utils.convert_to_list(pool_padding, 3, 'pool_padding') + pool_stride = utils.convert_to_list(pool_stride, 3, 'pool_stride') + + if not isinstance(use_cudnn, bool): + raise ValueError("use_cudnn should be True or False") + + l_type = "pool3d" + helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() pool_out = helper.create_tmp_variable(dtype) helper.append_op( - type="pool2d", + type=l_type, inputs={"X": input}, outputs={"Out": pool_out}, attrs={ @@ -1513,6 +1853,25 @@ def batch_norm(input, """ This function helps create an operator to implement the BatchNorm layer using the configurations from the input parameters. + + Args: + input (Variable): the input variable. + act (str): activation type + is_test (bool): whether to run batch_norm as test mode. + momentum (float): momentum + epsilon (float): epsilon, default 1e-05 + param_attr (ParamAttr|None): attributes for parameter + bias_attr (ParamAttr|None): attributes for bias + data_layout (str): data layout, default NCHW + in_place (bool): if True, do not create tmp variable + use_mkldnn (bool): ${use_mkldnn_comment} + name (str): The name of this layer. It is optional. + moving_mean_name (str): The name of moving mean variable name, optional. + moving_variance_name (str): The name of moving variance name, optional. + do_model_average_for_mean_and_var (bool): + + Returns: + Variable: output of batch_norm layer. """ helper = LayerHelper('batch_norm', **locals()) dtype = helper.input_dtype() @@ -1640,6 +1999,7 @@ def layer_norm(input, bias_attr(ParamAttr|None): The parameter attribute for the learnable bias :math:`b`. act(str): Activation to be applied to the output of layer normalizaiton. + name (str): The name of this layer. It is optional. Returns: Variable: A tensor variable with the same shape as the input. @@ -1691,6 +2051,17 @@ def layer_norm(input, def beam_search_decode(ids, scores, name=None): + """ + ${beam_search_decode} + + Args: + ids (Variable): ${ids_comment} + scores (Variable): ${scores_comment} + name (str): The name of this layer. It is optional. + + Returns: + tuple: a tuple of two output variable: sentence_ids, sentence_scores + """ helper = LayerHelper('beam_search_decode', **locals()) sentence_ids = helper.create_tmp_variable(dtype=ids.dtype) sentence_scores = helper.create_tmp_variable(dtype=ids.dtype) @@ -1766,46 +2137,46 @@ def conv2d_transpose(input, W_{out} &= (W_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (W_f - 1) + 1 Args: - input(Variable): The input image with [N, C, H, W] format. - num_filters(int): The number of the filter. It is as same as the output - image channel. - output_size(int|tuple|None): The output image size. If output size is a - tuple, it must contain two integers, (image_H, image_W). This - parameter only works when filter_size is None. - filter_size(int|tuple|None): The filter size. If filter_size is a tuple, - it must contain two integers, (filter_size_H, filter_size_W). - Otherwise, the filter will be a square. None if use output size to - calculate filter_size. - padding(int|tuple): The padding size. If padding is a tuple, it must - contain two integers, (padding_H, padding_W). Otherwise, the - padding_H = padding_W = padding. Default: padding = 0. - stride(int|tuple): The stride size. If stride is a tuple, it must - contain two integers, (stride_H, stride_W). Otherwise, the - stride_H = stride_W = stride. Default: stride = 1. - dilation(int|tuple): The dilation size. If dilation is a tuple, it must - contain two integers, (dilation_H, dilation_W). Otherwise, the - dilation_H = dilation_W = dilation. Default: dilation = 1. - groups(int): The groups number of the Conv2d transpose layer. Inspired by - grouped convolution in Alex Krizhevsky's Deep CNN paper, in which - when group=2, the first half of the filters is only connected to the - first half of the input channels, while the second half of the - filters is only connected to the second half of the input channels. - Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. - Default: None - bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None - use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn - library is installed. Default: True - act(str): Activation type. Default: None - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. - - Returns: - Variable: The tensor variable storing the convolution transpose result. + input(Variable): The input image with [N, C, H, W] format. + num_filters(int): The number of the filter. It is as same as the output + image channel. + output_size(int|tuple|None): The output image size. If output size is a + tuple, it must contain two integers, (image_H, image_W). This + parameter only works when filter_size is None. + filter_size(int|tuple|None): The filter size. If filter_size is a tuple, + it must contain two integers, (filter_size_H, filter_size_W). + Otherwise, the filter will be a square. None if use output size to + calculate filter_size. + padding(int|tuple): The padding size. If padding is a tuple, it must + contain two integers, (padding_H, padding_W). Otherwise, the + padding_H = padding_W = padding. Default: padding = 0. + stride(int|tuple): The stride size. If stride is a tuple, it must + contain two integers, (stride_H, stride_W). Otherwise, the + stride_H = stride_W = stride. Default: stride = 1. + dilation(int|tuple): The dilation size. If dilation is a tuple, it must + contain two integers, (dilation_H, dilation_W). Otherwise, the + dilation_H = dilation_W = dilation. Default: dilation = 1. + groups(int): The groups number of the Conv2d transpose layer. Inspired by + grouped convolution in Alex Krizhevsky's Deep CNN paper, in which + when group=2, the first half of the filters is only connected to the + first half of the input channels, while the second half of the + filters is only connected to the second half of the input channels. + Default: groups=1 + param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. + Default: None + bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None + use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + act(str): Activation type. Default: None + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The tensor variable storing the convolution transpose result. Raises: - ValueError: If the shapes of input, filter_size, stride, padding and - groups mismatch. + ValueError: If the shapes of input, filter_size, stride, padding and + groups mismatch. Examples: .. code-block:: python @@ -1869,6 +2240,173 @@ def conv2d_transpose(input, return out +def conv3d_transpose(input, + num_filters, + output_size=None, + filter_size=None, + padding=0, + stride=1, + dilation=1, + groups=None, + param_attr=None, + bias_attr=None, + use_cudnn=True, + act=None, + name=None): + """ + **Convlution3D transpose layer** + + The convolution3D transpose layer calculates the output based on the input, + filter, and dilations, strides, paddings. Input(Input) and output(Output) + are in NCDHW format. Where N is batch size, C is the number of channels, + D is the depth of the feature, H is the height of the feature, and W + is the width of the feature. Parameters(dilations, strides, paddings) are + two elements. These two elements represent height and width, respectively. + The details of convolution transpose layer, please refer to the following + explanation and references `therein `_. + + For each input :math:`X`, the equation is: + + .. math:: + + Out = W \\ast X + + In the above equation: + + * :math:`X`: Input value, a tensor with NCDHW format. + * :math:`W`: Filter value, a tensor with MCDHW format. + * :math:`\\ast` : Convolution transpose operation. + * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be + different. + + Example: + + - Input: + + Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$ + + Filter shape: $(C_{in}, C_{out}, D_f, H_f, W_f)$ + + - Output: + + Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$ + + Where + + .. math:: + + D_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\\\ + H_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\\\ + W_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1 + + Args: + input(Variable): The input image with [N, C, D, H, W] format. + num_filters(int): The number of the filter. It is as same as the output + image channel. + output_size(int|tuple|None): The output image size. If output size is a + tuple, it must contain three integers, (image_D, image_H, image_W). This + parameter only works when filter_size is None. + filter_size(int|tuple|None): The filter size. If filter_size is a tuple, + it must contain three integers, (filter_size_D, filter_size_H, filter_size_W). + Otherwise, the filter will be a square. None if use output size to + calculate filter_size. + padding(int|tuple): The padding size. If padding is a tuple, it must + contain three integers, (padding_D, padding_H, padding_W). Otherwise, the + padding_D = padding_H = padding_W = padding. Default: padding = 0. + stride(int|tuple): The stride size. If stride is a tuple, it must + contain three integers, (stride_D, stride_H, stride_W). Otherwise, the + stride_D = stride_H = stride_W = stride. Default: stride = 1. + dilation(int|tuple): The dilation size. If dilation is a tuple, it must + contain three integers, (dilation_D, dilation_H, dilation_W). Otherwise, the + dilation_D = dilation_H = dilation_W = dilation. Default: dilation = 1. + groups(int): The groups number of the Conv3d transpose layer. Inspired by + grouped convolution in Alex Krizhevsky's Deep CNN paper, in which + when group=2, the first half of the filters is only connected to the + first half of the input channels, while the second half of the + filters is only connected to the second half of the input channels. + Default: groups=1 + param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer. + Default: None + bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None + use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn + library is installed. Default: True + act(str): Activation type. Default: None + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The tensor variable storing the convolution transpose result. + + Raises: + ValueError: If the shapes of input, filter_size, stride, padding and + groups mismatch. + + Examples: + .. code-block:: python + + data = fluid.layers.data( + name='data', shape=[3, 12, 32, 32], dtype='float32') + conv2d_transpose = fluid.layers.conv3d_transpose( + input=data, num_filters=2, filter_size=3) + """ + l_type = "conv3d_transpose" + helper = LayerHelper(l_type, **locals()) + if not isinstance(input, Variable): + raise TypeError("Input of conv3d_transpose must be Variable") + input_channel = input.shape[1] + + padding = utils.convert_to_list(padding, 3, 'padding') + stride = utils.convert_to_list(stride, 3, 'stride') + dilation = utils.convert_to_list(dilation, 3, 'dilation') + + if not isinstance(use_cudnn, bool): + raise ValueError("use_cudnn should be True or False") + + if filter_size is None: + if output_size is None: + raise ValueError("output_size must be set when filter_size is None") + if isinstance(output_size, int): + output_size = [output_size, output_size] + + d_in = input.shape[2] + h_in = input.shape[3] + w_in = input.shape[4] + + filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 * + padding[0] - 1) / dilation[0] + 1 + filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 * + padding[1] - 1) / dilation[1] + 1 + filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 * + padding[2] - 1) / dilation[2] + 1 + filter_size = [filter_size_d, filter_size_h, filter_size_w] + else: + filter_size = utils.convert_to_list(filter_size, 3, + 'conv3d_transpose.filter_size') + + groups = 1 if groups is None else groups + filter_shape = [input_channel, num_filters / groups] + filter_size + img_filter = helper.create_parameter( + dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) + + pre_bias = helper.create_tmp_variable(dtype=input.dtype) + helper.append_op( + type=l_type, + inputs={'Input': [input], + 'Filter': [img_filter]}, + outputs={'Output': pre_bias}, + attrs={ + 'strides': stride, + 'paddings': padding, + 'dilations': dilation, + 'groups': groups, + 'use_cudnn': use_cudnn + }) + + pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) + out = helper.append_activation(pre_act) + return out + + def sequence_expand(x, y, ref_level=-1, name=None): """Sequence Expand Layer. This layer will expand the input variable **x** according to specified level lod of **y**. Please note that lod level of @@ -1942,6 +2480,17 @@ def sequence_expand(x, y, ref_level=-1, name=None): def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0): ''' This function implements the beam search algorithm. + + Args: + pre_ids (Variable): ${pre_ids_comment} + ids (Variable): ${ids_comment} + scores (Variable): ${scores_comment} + beam_size (int): ${beam_size_comment} + end_id (int): ${end_id_comment} + level (int): ${level_comment} + + Returns: + tuple: a tuple of beam_search output variables: selected_ids, selected_scores ''' helper = LayerHelper('beam_search', **locals()) score_type = scores.dtype @@ -2437,19 +2986,21 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): The l2 normalize layer normalizes `x` along dimension `axis` using an L2 norm. For a 1-D tensor (`dim` is fixed to 0), this layer computes - output = x / sqrt(max(sum(x**2), epsilon)) + .. math:: + y = \frac{x}{ \sqrt{\sum {x^2} + epsion }} For `x` with more dimensions, this layer independently normalizes each 1-D slice along dimension `axis`. Args: - x(Variable|list): The input tensor to l2_normalize layer. - axis(int): Dimension along which to normalize the input. - epsilon(float): A lower bound value for `x`'s l2 norm. sqrt(epsilon) will - be used as the divisor if the l2 norm of `x` is less than - sqrt(epsilon). - name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + x(Variable|list): The input tensor to l2_normalize layer. + axis(int): The axis on which to apply normalization. If `axis < 0`, + the dimension to normalization is rank(X) + axis. -1 is the + last dimension. + epsilon(float): The epsilon value is used to avoid division by zero, + the defalut value is 1e-10. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: @@ -2468,46 +3019,17 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): axis = 0 helper = LayerHelper("l2_normalize", **locals()) - square = helper.create_tmp_variable(dtype=x.dtype) - helper.append_op(type="square", inputs={"X": x}, outputs={"Out": square}) - - reduced_sum = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_tmp_variable(dtype=x.dtype) + norm = helper.create_tmp_variable(dtype=x.dtype) helper.append_op( - type="reduce_sum", - inputs={"X": square}, - outputs={"Out": reduced_sum}, + type="norm", + inputs={"X": x}, + outputs={"Out": out, + "Norm": norm}, attrs={ - "dim": [1] if axis is None else [axis], - "keep_dim": True, - "reduce_all": False + "axis": 1 if axis is None else axis, + "epsilon": epsilon, }) - - # TODO(caoying) A lower bound value epsilon for the norm is needed to - # imporve the numeric stability of reciprocal. This requires a maximum_op. - rsquare = helper.create_tmp_variable(dtype=x.dtype) - helper.append_op( - type="reciprocal", inputs={"X": reduced_sum}, outputs={"Out": rsquare}) - - # TODO(caoying) the current elementwise_mul operator does not support a - # general broadcast rule which broadcasts input(Y) to have the same - # dimension with Input(X) starting from a specified dimension. So this - # exanpsion is requred. Once a general broadcast rule is spported, this - # expanding canbe removed. - rsquare_expanded = helper.create_tmp_variable(dtype=x.dtype) - expand_times = [1] * len(x.shape) - expand_times[axis] = int(x.shape[axis]) - helper.append_op( - type="expand", - inputs={"X": rsquare}, - outputs={"Out": rsquare_expanded}, - attrs={"expand_times": expand_times}) - - out = helper.create_tmp_variable(dtype=x.dtype) - helper.append_op( - type="elementwise_mul", - inputs={"X": x, - "Y": rsquare_expanded}, - outputs={"Out": out}) return out @@ -2691,16 +3213,13 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None, the edit distance will be divided by the length of reference string. Args: - input(Variable): The indices for hypothesis strings. - label(Variable): The indices for reference strings. - normalized(bool): Indicated whether to normalize the edit distance by the length of reference string. - ignored_tokens(list of int): Tokens that should be removed before calculating edit distance. + name (str): The name of this layer. It is optional. Returns: Variable: sequence-to-sequence edit distance in shape [batch_size, 1]. @@ -2790,10 +3309,10 @@ def ctc_greedy_decoder(input, blank, name=None): where Lp is the sum of all input sequences' length and num_classes is the true number of classes. (not including the blank label). - blank(int): the blank label index of Connectionist Temporal Classification (CTC) loss, which is in thehalf-opened interval [0, num_classes + 1). + name (str): The name of this layer. It is optional. Returns: Variable: CTC greedy decode result. If all the sequences in result were @@ -2830,23 +3349,23 @@ def warpctc(input, label, blank=0, norm_by_times=False): input tensor. Args: - input(Variable): (LodTensor, default: LoDTensor), - the unscaled probabilities of variable-length sequences, - which is a 2-D Tensor with LoD information. - It's shape is [Lp, num_classes + 1], where Lp is the sum of all input - sequences' length and num_classes is the true number of classes. - (not including the blank label). - label(Variable): (LodTensor, default: LoDTensor), the ground truth - of variable-length sequence, which is a 2-D Tensor with LoD - information. It is of the shape [Lg, 1], where Lg is th sum of - all labels' length. - blank: (int, default: 0), the blank label index of Connectionist - Temporal Classification (CTC) loss, which is in the - half-opened interval [0, num_classes + 1). - norm_by_times: (bool, default: false), whether to normalize - the gradients by the number of time-step, which is also the - sequence's length. There is no need to normalize the gradients - if warpctc layer was follewed by a mean_op. + input(Variable): (LodTensor, default: LoDTensor), + the unscaled probabilities of variable-length sequences, + which is a 2-D Tensor with LoD information. + It's shape is [Lp, num_classes + 1], where Lp is the sum of all input + sequences' length and num_classes is the true number of classes. + (not including the blank label). + label(Variable): (LodTensor, default: LoDTensor), the ground truth + of variable-length sequence, which is a 2-D Tensor with LoD + information. It is of the shape [Lg, 1], where Lg is th sum of + all labels' length. + blank (int): default 0, the blank label index of Connectionist + Temporal Classification (CTC) loss, which is in the + half-opened interval [0, num_classes + 1). + norm_by_times (bool): default false, whether to normalize + the gradients by the number of time-step, which is also the + sequence's length. There is no need to normalize the gradients + if warpctc layer was follewed by a mean_op. Returns: Variable: The Connectionist Temporal Classification (CTC) loss, @@ -2905,9 +3424,9 @@ def sequence_reshape(input, new_dim): no remainder for each sequence. Args: - input (Variable): (LodTensor, default: LoDTensor), a 2-D LoDTensor - with shape being [N, M] where M for dimension. - new_dim (int): New dimension which the input LoDTensor is reshaped to. + input (Variable): (LodTensor, default: LoDTensor), a 2-D LoDTensor + with shape being [N, M] where M for dimension. + new_dim (int): New dimension which the input LoDTensor is reshaped to. Returns: Variable: Reshaped LoDTensor according to new dimension. @@ -2929,7 +3448,10 @@ def sequence_reshape(input, new_dim): return out -@autodoc() +# FIXME(wuyi): let docstring_checker.py understand @autodoc. +# For now, the comments in c++ use types like Tensor, but in python side +# the type is often "Variable", and arguments may vary. +@templatedoc(op_type="nce") def nce(input, label, num_total_classes, @@ -2937,6 +3459,21 @@ def nce(input, param_attr=None, bias_attr=None, num_neg_samples=None): + """ + ${comment} + + Args: + input (Variable): input variable. + label (Variable): label. + num_total_classes (int):${num_total_classes_comment} + sample_weight (int): ${sample_weight_comment} + param_attr (ParamAttr|None): attributes for parameter + bias_attr (ParamAttr|None): attributes for bias + num_neg_samples (int): ${num_neg_samples_comment} + + Returns: + Variable: output of nce layer. + """ helper = LayerHelper('nce', **locals()) assert isinstance(input, Variable) dim = input.shape[1] @@ -3062,8 +3599,9 @@ def transpose(x, perm, name=None): perm[i]-th dimension of `input`. Args: - input (Variable): (Tensor), A Tensor. - perm (list): A permutation of the dimensions of `input`. + x (Variable): The input Tensor. + perm (list): A permutation of the dimensions of `input`. + name (str): The name of this layer. It is optional. Returns: Variable: A transposed Tensor. @@ -3296,9 +3834,9 @@ def multiplex(inputs, index): row of the matrix, then `O[i]` is equal to :math:`I_{ID[i]}[i]`. Args: - inputs (list): A list of variables to gather from. All variables have the + inputs (list): A list of variables to gather from. All variables have the same shape and the rank is at least 2. - index (Variable): Tensor, index variable which is a 2-D tensor + index (Variable): Tensor, index variable which is a 2-D tensor with shape [M, 1] where M is the batch size. Returns: @@ -3497,7 +4035,8 @@ def autoincreased_step_counter(counter_name=None, begin=1, step=1): begin(int): The first value of this counter. step(int): The increment step between each execution. - Returns(Variable): The global run counter. + Returns: + Variable: The global run counter. """ helper = LayerHelper('global_step_counter') if counter_name is None: @@ -3558,7 +4097,7 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): the corresponding dimension of x. Args: - input(variable): The input tensor. + x(variable): The input tensor. shape(list): The new shape. At most one dimension of the new shape can be -1. actual_shape(variable): An optional input. If provided, reshape @@ -3570,8 +4109,10 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): inplace(bool): If this flag is set true, a new output tensor is created whose data is copied from input x, otherwise the output shares data with input without copying. + name (str): The name of this layer. It is optional. - Returns(variable): The output tensor. + Returns: + Variable: The output tensor. Examples: .. code-block:: python @@ -3958,7 +4499,6 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): def dice_loss(input, label, epsilon=0.00001): """ - **Dice loss Layer** Dice loss for comparing the similarity of two batch of data, usually is used for binary image segmentation i.e. labels are binary. The dice loss can be defined as below equation: @@ -3998,30 +4538,35 @@ def dice_loss(input, label, epsilon=0.00001): return reduce_mean(dice_score) -def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): +def image_resize(input, + out_shape=None, + scale=None, + name=None, + resample='BILINEAR'): """ - The mathematical meaning of upsampling_bilinear2d is also called - Bilinear interpolation. - Bilinear interpolation is an extension of linear interpolation for - interpolating functions of two variables (e.g. H-direction and - W-direction in this layer) on a rectilinear 2D grid. + Resize a batch of images. + + The input must be a tensor of the shape (num_batches, channels, in_h, in_w), + and the resizing only applies on the last two dimensions(hight and width). - For details, please refer to Wikipedia: - https://en.wikipedia.org/wiki/Bilinear_interpolation + Supporting resample methods: + 'BILINEAR' : Bilinear interpolation Args: - input (Variable): The input tensor of bilinear interpolation, + input (Variable): The input tensor of image resize layer, This is a 4-D tensor of the shape (num_batches, channels, in_h, in_w). - out_shape(list|tuple|None): Output shape of bilinear interpolation + out_shape(list|tuple|Variable|None): Output shape of image resize layer, the shape is (out_h, out_w). Default: None - scale(int|None): The multiplier for the input height or width. + scale(float|None): The multiplier for the input height or width. At least one of out_shape or scale must be set. And out_shape has a higher priority than scale. Default: None name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. + resample(str): The resample method. It can only be 'BILINEAR' currently. + Default: 'BILINEAR' Returns: out (Variable): The output is a 4-D tensor of the shape @@ -4030,8 +4575,12 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): Examples: .. code-block:: python - out = fluid.layers.bilinear_interp(input, out_shape=[12, 12]) + out = fluid.layers.image_resize(input, out_shape=[12, 12]) """ + resample_methods = {'BILINEAR': 'bilinear_interp'} + if resample not in resample_methods: + raise ValueError( + "The 'resample' of image_resize can only be 'BILINEAR' currently.") if out_shape is None and scale is None: raise ValueError("One of out_shape and scale must not be None") helper = LayerHelper('bilinear_interp', **locals()) @@ -4040,31 +4589,164 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): def _is_list_or_turple_(data): return (isinstance(data, list) or isinstance(data, tuple)) + out_h = 0 + out_w = 0 + inputs = {"X": input} if out_shape is not None: - if not (_is_list_or_turple_(out_shape) and len(out_shape) == 2): - raise ValueError('out_shape should be a list or tuple ', - 'with length 2, (out_h, out_w).') - out_shape = list(map(int, out_shape)) - out_h = out_shape[0] - out_w = out_shape[1] + if not (_is_list_or_turple_(out_shape) and + len(out_shape) == 2) and not isinstance(out_shape, Variable): + raise ValueError('out_shape should be a list or tuple or variable') + if _is_list_or_turple_(out_shape): + out_shape = list(map(int, out_shape)) + out_h = out_shape[0] + out_w = out_shape[1] + else: + inputs['OutSize'] = out_shape else: out_h = int(input.shape[2] * scale) out_w = int(input.shape[3] * scale) out = helper.create_tmp_variable(dtype) helper.append_op( - type="bilinear_interp", - inputs={"X": input}, + type=resample_methods[resample], + inputs=inputs, outputs={"Out": out}, attrs={"out_h": out_h, "out_w": out_w}) return out -def random_crop(input, shape, seed=1): +@templatedoc(op_type="bilinear_interp") +def resize_bilinear(input, out_shape=None, scale=None, name=None): + """ + ${comment} + + Args: + input(${x_type}): ${x_comment}. + + out_shape(${out_size_type}): ${out_size_comment}. + + scale(float|None): The multiplier for the input height or width. At + least one of out_shape or scale must be set. And out_shape has + a higher priority than scale. Default: None. + + name(str|None): The output variable name. + + Returns: + ${out_comment}. + """ + + return image_resize(input, out_shape, scale, name, 'BILINEAR') + + +def image_resize_short(input, out_short_len, resample='BILINEAR'): + """ + Resize a batch of images. The short edge of input images will be + resized to the given 'out_short_len'. The long edge of input images + will be resized proportionately to make images' length-width ratio + constant. + + Args: + input (Variable): The input tensor of image resize layer, + This is a 4-D tensor of the shape + (num_batches, channels, in_h, in_w). + out_short_len(int): The length of output images' short edge. + resample (str): resample method, default: BILINEAR. + + Returns: + out (Variable): The output is a 4-D tensor of the shape + (num_batches, channls, out_h, out_w). + """ + in_shape = input.shape + if len(in_shape) != 4: + raise ValueError( + "The rank of input must be 4 (num_batches, channels, in_h, in_w).") + hw = in_shape[2:4] + short_idx = hw.index(min(hw)) + long_idx = 1 - short_idx + out_shape = list(hw) + out_shape[short_idx] = out_short_len + out_shape[long_idx] = int( + float(out_shape[long_idx]) * (float(out_short_len) / float(hw[ + short_idx])) + 0.5) + return image_resize(input=input, out_shape=out_shape, resample=resample) + + +def gather(input, index): + """ + Output is obtained by gathering entries of the outer-most dimension + of X indexed by `index` and concatenate them together. + + .. math:: + + Out = X[Index] + + + .. code-block:: text + + + Given: + + X = [[1, 2], + [3, 4], + [5, 6]] + + Index = [1, 2] + + Then: + + Out = [[3, 4], + [5, 6]] + + Args: + input (Variable): The source input with rank>=1. + index (Variable): The index input with rank=1. + + Returns: + output (Variable): The output is a tensor with the same rank as input. + + Examples: + + .. code-block:: python + + output = fluid.layers.gather(x, index) + """ + helper = LayerHelper('gather', **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + helper.append_op( + type="gather", + inputs={"X": input, + "Index": index}, + outputs={"Out": out}) + return out + + +@templatedoc() +def random_crop(x, shape, seed=None): + """ + ${comment} + + Examples: + >>> img = fluid.layers.data("img", [3, 256, 256]) + >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224]) + + Args: + x(${x_type}): ${x_comment} + shape(${shape_type}): ${shape_comment} + seed(int|${seed_type}|None): ${seed_comment} By default, the seed will + get from `random.randint(-65536, 65535)`. + + Returns: + ${out_comment} + + """ helper = LayerHelper("random_crop", **locals()) dtype = helper.input_dtype() out = helper.create_tmp_variable(dtype) + if seed is None: + seed = random.randint(-65536, 65535) + if isinstance(seed, int): seed_value = seed seed = helper.create_tmp_variable(dtype="int64") @@ -4082,9 +4764,59 @@ def random_crop(input, shape, seed=1): seed_out = helper.create_tmp_variable(dtype="int64") helper.append_op( type="random_crop", - inputs={"X": input, + inputs={"X": x, "Seed": seed}, outputs={"Out": out, "SeedOut": seed_out}, attrs={"shape": shape}) return out + + +def mean_iou(input, label, num_classes): + """ + Mean Intersection-Over-Union is a common evaluation metric for + semantic image segmentation, which first computes the IOU for each + semantic class and then computes the average over classes. + IOU is defined as follows: + + .. math:: + + IOU = true_positive / (true_positive + false_positive + false_negative). + + The predictions are accumulated in a confusion matrix and mean-IOU + is then calculated from it. + + + Args: + input (Variable): A Tensor of prediction results for semantic labels with type int32 or int64. + label (Variable): A Tensor of ground truth labels with type int32 or int64. + Its shape should be the same as input. + + Returns: + mean_iou (Variable): A Tensor representing the mean intersection-over-union with shape [1]. + out_wrong(Variable): A Tensor with shape [num_classes]. The wrong numbers of each class. + out_correct(Variable): A Tensor with shape [num_classes]. The correct numbers of each class. + + + Examples: + + .. code-block:: python + + iou, wrongs, corrects = fluid.layers.mean_iou(predict, label, num_classes) + """ + helper = LayerHelper('mean_iou', **locals()) + dtype = helper.input_dtype() + out_mean_iou = helper.create_tmp_variable(dtype='float32') + out_wrong = helper.create_tmp_variable(dtype='int32') + out_correct = helper.create_tmp_variable(dtype='int32') + helper.append_op( + type="mean_iou", + inputs={"predictions": input, + "labels": label}, + outputs={ + "out_mean_iou": out_mean_iou, + "out_wrong": out_wrong, + "out_correct": out_correct + }, + attrs={"num_classes": num_classes}) + return out_mean_iou, out_wrong, out_correct diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index a9fe25744c..98f169e8f0 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -71,6 +71,10 @@ __all__ = [ 'cumsum', 'scatter', 'sum', + 'slice', + 'polygon_box_transform', + 'shape', + 'maxout', ] + __activations__ for _OP in set(__all__): diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index be34cc81a5..62b01d595a 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -18,6 +18,7 @@ from ..framework import convert_np_dtype_to_dtype_ from ..framework import Variable from ..initializer import Constant, force_init_on_cpu from ..core import VarDesc +from layer_function_generator import templatedoc import numpy __all__ = [ @@ -30,6 +31,8 @@ __all__ = [ 'assign', 'fill_constant_batch_size_like', 'fill_constant', + 'argmin', + 'argmax', 'ones', 'zeros', ] @@ -266,6 +269,7 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None): return out +@templatedoc() def fill_constant_batch_size_like(input, shape, dtype, @@ -273,30 +277,28 @@ def fill_constant_batch_size_like(input, input_dim_idx=0, output_dim_idx=0): """ - **fill_constant_batch_size_like** - - This function creates a tensor of specified *shape*, *dtype* and batch size, - and initializes this with a constant supplied in *value*. The batch size is - obtained from the `input` tensor. + ${comment} It also sets *stop_gradient* to True. + >>> data = fluid.layers.fill_constant_batch_size_like( + >>> input=like, shape=[1], value=0, dtype='int64') + Args: - input(Variable): Tensor whose dimensions will be used to get batch size - shape(tuple|list|None): Shape of output tensor - dtype(np.dtype|core.VarDesc.VarType|str): Data type of output tensor - value(float): Constant value to initialize the output tensor - input_dim_idx(int): Index of input's batch size dimension - output_dim_idx(int): Index of output's batch size dimension + input(${input_type}): ${input_comment}. - Returns: - Variable: The tensor variable storing the output + shape(${shape_type}): ${shape_comment}. - Examples: - .. code-block:: python + dtype(${dtype_type}): ${dtype_comment}. + + value(${value_type}): ${value_comment}. - data = fluid.layers.fill_constant_batch_size_like( - input=like, shape=[1], value=0, dtype='int64') + input_dim_idx(${input_dim_idx_type}): ${input_dim_idx_comment}. + + output_dim_idx(${output_dim_idx_type}): ${output_dim_idx_comment}. + + Returns: + ${out_comment}. """ helper = LayerHelper("fill_constant_batch_size_like", **locals()) out = helper.create_tmp_variable(dtype=dtype) @@ -315,6 +317,68 @@ def fill_constant_batch_size_like(input, return out +def argmin(x, axis=0): + """ + **argmin** + + This function computes the indices of the min elements + of the input tensor's element along the provided axis. + + Args: + x(Variable): The input to compute the indices of + the min elements. + axis(int): Axis to compute indices along. + + Returns: + Variable: The tensor variable storing the output + + Examples: + .. code-block:: python + + out = fluid.layers.argmin(x=in, axis=0) + out = fluid.layers.argmin(x=in, axis=-1) + """ + helper = LayerHelper("arg_min", **locals()) + out = helper.create_tmp_variable(VarDesc.VarType.INT64) + helper.append_op( + type='arg_min', + inputs={'X': x}, + outputs={'Out': [out]}, + attrs={'axis': axis}) + return out + + +def argmax(x, axis=0): + """ + **argmax** + + This function computes the indices of the max elements + of the input tensor's element along the provided axis. + + Args: + x(Variable): The input to compute the indices of + the max elements. + axis(int): Axis to compute indices along. + + Returns: + Variable: The tensor variable storing the output + + Examples: + .. code-block:: python + + out = fluid.layers.argmax(x=in, axis=0) + out = fluid.layers.argmax(x=in, axis=-1) + """ + helper = LayerHelper("arg_max", **locals()) + out = helper.create_tmp_variable(VarDesc.VarType.INT64) + helper.append_op( + type='arg_max', + inputs={'X': x}, + outputs={'Out': [out]}, + attrs={'axis': axis}) + return out + + def ones(shape, dtype, force_cpu=False): """ **ones** @@ -363,6 +427,40 @@ def zeros(shape, dtype, force_cpu=False): return fill_constant(value=0.0, **locals()) +def reverse(x, axis): + """ + **reverse** + + This function reverse the input 'x' along given axises. + + Args: + x(Vairbale): the input to be reversed. + axis(int|tuple|list): Axis that along which order of elements + is reversed. If it is a tuple or a list, reversing + will be apply on each axis in the tuple or list. + + Returns: + Variable: The reversed tensor. + + Examples: + .. code-block:: python + + out = fluid.layers.reverse(x=in, axis=0) + # or: + out = fluid.layers.reverse(x=in, axis=[0,1]) + """ + if isinstance(axis, int): + axis = [axis] + helper = LayerHelper("reverse", **locals()) + out = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type='reverse', + inputs={'Input': x}, + outputs={'Out': [out]}, + attrs={'axis': axis}) + return out + + def save(x, file_path, overwrite=True): """ Saves a variable as a file. @@ -403,22 +501,6 @@ def save_combine(x, file_path, overwrite=True): "overwrite": overwrite}) -def load(out, file_path): - """ - Loads a variable from a given file. - - Args: - out(variable): The variable to be read from the disk file. - file_path(str): The path of the disk file. - """ - helper = LayerHelper("load", **locals()) - helper.append_op( - type="load", - inputs={}, - output={"Out": out}, - args={"file_path": file_path}) - - def load_combine(out, file_path): """ Loads a list of vairables from a single file. diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 3117dfe00c..0fdc9a0352 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -18,6 +18,7 @@ import framework import executor import warnings import sys +import os __all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy'] @@ -101,7 +102,9 @@ class ParallelExecutor(object): p.set_place(self._act_places[-1]) self._places.append(p) else: - for i in xrange(multiprocessing.cpu_count()): + cpu_num = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + for i in xrange(cpu_num): p = core.Place() self._act_places.append(core.CPUPlace()) p.set_place(self._act_places[-1]) @@ -110,19 +113,17 @@ class ParallelExecutor(object): if exec_strategy is None: exec_strategy = ExecutionStrategy() - if use_cuda: - exec_strategy.use_event = True - else: - exec_strategy.use_event = False + exec_strategy.use_cuda = use_cuda if exec_strategy.num_threads == 0: if use_cuda: # Experiments on se-resnext shows that too many threads hurt # performance. Worth tunning for other models in the future. - exec_strategy.num_threads = len(self._places) * 2 + exec_strategy.num_threads = len(self._places) * 4 else: - exec_strategy.num_threads = min( - len(self._places) * 2, multiprocessing.cpu_count()) + cpu_num = int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) + exec_strategy.num_threads = cpu_num if build_strategy is None: build_strategy = BuildStrategy() diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py index 5accaacd53..8d48e9abef 100644 --- a/python/paddle/fluid/recordio_writer.py +++ b/python/paddle/fluid/recordio_writer.py @@ -12,10 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +import os import core import contextlib - -__all__ = ['convert_reader_to_recordio_file'] +__all__ = [ + 'convert_reader_to_recordio_file', 'convert_reader_to_recordio_files' +] @contextlib.contextmanager @@ -46,3 +48,36 @@ def convert_reader_to_recordio_file( writer.complete_append_tensor() counter += 1 return counter + + +def convert_reader_to_recordio_files( + filename, + batch_per_file, + reader_creator, + feeder, + compressor=core.RecordIOWriter.Compressor.Snappy, + max_num_records=1000, + feed_order=None): + if feed_order is None: + feed_order = feeder.feed_names + f_name, f_ext = os.path.splitext(filename) + assert (f_ext == ".recordio") + + lines = [] + f_idx = 0 + counter = 0 + for idx, batch in enumerate(reader_creator()): + lines.append(batch) + if idx >= batch_per_file and idx % batch_per_file == 0: + filename = "%s-%05d%s" % (f_name, f_idx, f_ext) + with create_recordio_writer(filename, compressor, + max_num_records) as writer: + for l in lines: + res = feeder.feed(l) + for each in feed_order: + writer.append_tensor(res[each]) + writer.complete_append_tensor() + counter += 1 + lines = [] + f_idx += 1 + return counter diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py index de3906fc6a..ad28c9eff5 100644 --- a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py @@ -38,7 +38,7 @@ def inference_program(): return y_predict -def linear(): +def train_program(): y = fluid.layers.data(name='y', shape=[1], dtype='float32') y_predict = inference_program() @@ -48,13 +48,15 @@ def linear(): return avg_loss +def optimizer_func(): + return fluid.optimizer.SGD(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, - place=place, - optimizer=fluid.optimizer.SGD(learning_rate=0.001)) + train_func=train_program, place=place, optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndStepEvent): @@ -102,7 +104,7 @@ def main(use_cuda): # Directory for saving the trained model params_dirname = "fit_a_line.inference.model" - train(use_cuda, linear, params_dirname) + train(use_cuda, train_program, params_dirname) infer(use_cuda, inference_program, params_dirname) diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py index 63dc1b6ce3..8e222d2690 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py @@ -85,6 +85,10 @@ def train_network(): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): BATCH_SIZE = 128 EPOCH_NUM = 1 @@ -92,10 +96,11 @@ def train(use_cuda, train_program, params_dirname): train_reader = paddle.batch( paddle.reader.shuffle( cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), - batch_size=BATCH_SIZE) + batch_size=BATCH_SIZE, + drop_last=False) test_reader = paddle.batch( - paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) + paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False) def event_handler(event): if isinstance(event, fluid.EndStepEvent): @@ -111,9 +116,7 @@ def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, - optimizer=fluid.optimizer.Adam(learning_rate=0.001), - place=place) + train_func=train_program, optimizer_func=optimizer_func, place=place) trainer.train( reader=train_reader, diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py index 0bf8f265a1..dbc7bc06c9 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py @@ -64,15 +64,20 @@ def train_network(): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): BATCH_SIZE = 128 train_reader = paddle.batch( paddle.reader.shuffle( cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10), - batch_size=BATCH_SIZE) + batch_size=BATCH_SIZE, + drop_last=False) test_reader = paddle.batch( - paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE) + paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE, drop_last=False) def event_handler(event): if isinstance(event, fluid.EndStepEvent): @@ -88,9 +93,7 @@ def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() trainer = fluid.Trainer( - train_func=train_program, - place=place, - optimizer=fluid.optimizer.Adam(learning_rate=0.001)) + train_func=train_program, place=place, optimizer_func=optimizer_func) trainer.train( reader=train_reader, diff --git a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py index 8cce398ff3..0ccb3a39e0 100755 --- a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py +++ b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py @@ -141,12 +141,16 @@ def train_program(): return [avg_cost] +def optimize_func(): + return fluid.optimizer.SGD(learning_rate=fluid.layers.exponential_decay( + learning_rate=0.01, decay_steps=100000, decay_rate=0.5, staircase=True)) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.SGD(learning_rate=0.01) trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer=optimizer) + train_func=train_program, place=place, optimizer_func=optimize_func) feed_order = [ 'word_data', 'ctx_n2_data', 'ctx_n1_data', 'ctx_0_data', 'ctx_p1_data', @@ -245,7 +249,7 @@ def infer(use_cuda, inference_program, params_dirname): }, return_numpy=False) - print("infer results: ", np.array(results[0])) + print("infer results: ", np.array(results[0]).shape) def main(use_cuda): diff --git a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py index d4b723d3e6..c4b37df3a0 100644 --- a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py @@ -158,6 +158,13 @@ def train_program(is_sparse): return avg_cost +def optimizer_func(): + return fluid.optimizer.Adagrad( + learning_rate=1e-4, + regularization=fluid.regularizer.L2DecayRegularizer( + regularization_coeff=0.1)) + + def train(use_cuda, is_sparse, is_local=True): EPOCH_NUM = 1 @@ -182,11 +189,8 @@ def train(use_cuda, is_sparse, is_local=True): trainer = fluid.Trainer( train_func=partial(train_program, is_sparse), - optimizer=fluid.optimizer.Adagrad( - learning_rate=1e-4, - regularization=fluid.regularizer.L2DecayRegularizer( - regularization_coeff=0.1)), - place=place) + place=place, + optimizer_func=optimizer_func) trainer.train( reader=train_reader, diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py index 03439cbd37..9a09db25dc 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py @@ -57,14 +57,17 @@ def train_program(): return [avg_cost, acc] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adam(learning_rate=0.001) trainer = fluid.Trainer( train_func=train_program, place=place, - optimizer=optimizer, + optimizer_func=optimizer_func, parallel=True) def event_handler(event): diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py index 89bbd21bea..b2b544e791 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py @@ -44,12 +44,15 @@ def train_program(): return [avg_cost, acc] +def optimizer_func(): + return fluid.optimizer.Adam(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adam(learning_rate=0.001) trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer=optimizer) + train_func=train_program, place=place, optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): diff --git a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py index dfc7325acf..090c11ce1e 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py +++ b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py @@ -155,12 +155,15 @@ def train_program(): return [avg_cost, scale_infer] +def optimizer_func(): + return fluid.optimizer.SGD(learning_rate=0.2) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.SGD(learning_rate=0.2) trainer = fluid.Trainer( - train_func=train_program, place=place, optimizer=optimizer) + train_func=train_program, place=place, optimizer_func=optimizer_func) feed_order = [ 'user_id', 'gender_id', 'age_id', 'job_id', 'movie_id', 'category_id', diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py index 11e9fd1bec..9b61f7a00c 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py @@ -64,15 +64,18 @@ def train_program(word_dict): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adagrad(learning_rate=0.002) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) word_dict = paddle.dataset.imdb.word_dict() trainer = fluid.Trainer( train_func=partial(train_program, word_dict), place=place, - optimizer=optimizer) + optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py index 90757d54f9..aa7c567b4d 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py @@ -79,15 +79,18 @@ def train_program(word_dict): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adagrad(learning_rate=0.002) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) word_dict = paddle.dataset.imdb.word_dict() trainer = fluid.Trainer( train_func=partial(train_program, word_dict), place=place, - optimizer=optimizer) + optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py index 52b7d4a837..8c74be0f08 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py @@ -71,20 +71,25 @@ def train_program(word_dict): return [avg_cost, accuracy] +def optimizer_func(): + return fluid.optimizer.Adagrad(learning_rate=0.002) + + def train(use_cuda, train_program, params_dirname): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() - optimizer = fluid.optimizer.Adagrad(learning_rate=0.002) word_dict = paddle.dataset.imdb.word_dict() trainer = fluid.Trainer( train_func=partial(train_program, word_dict), place=place, - optimizer=optimizer) + optimizer_func=optimizer_func) def event_handler(event): if isinstance(event, fluid.EndEpochEvent): test_reader = paddle.batch( - paddle.dataset.imdb.test(word_dict), batch_size=BATCH_SIZE) + paddle.dataset.imdb.test(word_dict), + batch_size=BATCH_SIZE, + drop_last=False) avg_cost, acc = trainer.test( reader=test_reader, feed_order=['words', 'label']) @@ -110,7 +115,8 @@ def train(use_cuda, train_program, params_dirname): train_reader = paddle.batch( paddle.reader.shuffle( paddle.dataset.imdb.train(word_dict), buf_size=25000), - batch_size=BATCH_SIZE) + batch_size=BATCH_SIZE, + drop_last=False) trainer.train( num_epochs=1, diff --git a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py index eeb8e67087..ba44f72d9b 100644 --- a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py +++ b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py @@ -80,6 +80,10 @@ def train_program(is_sparse): return avg_cost +def optimizer_func(): + return fluid.optimizer.SGD(learning_rate=0.001) + + def train(use_cuda, train_program, params_dirname): train_reader = paddle.batch( paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE) @@ -104,9 +108,7 @@ def train(use_cuda, train_program, params_dirname): sys.exit("got NaN loss, training failed.") trainer = fluid.Trainer( - train_func=train_program, - optimizer=fluid.optimizer.SGD(learning_rate=0.001), - place=place) + train_func=train_program, optimizer_func=optimizer_func, place=place) trainer.train( reader=train_reader, diff --git a/python/paddle/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/fluid/tests/book/test_label_semantic_roles.py index bc8a1aafc8..99d51ae007 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -76,8 +76,7 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, emb_layers.append(mark_embedding) hidden_0_layers = [ - fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') - for emb in emb_layers + fluid.layers.fc(input=emb, size=hidden_dim) for emb in emb_layers ] hidden_0 = fluid.layers.sums(input=hidden_0_layers) @@ -94,8 +93,8 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, for i in range(1, depth): mix_hidden = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') + fluid.layers.fc(input=input_tmp[0], size=hidden_dim), + fluid.layers.fc(input=input_tmp[1], size=hidden_dim) ]) lstm = fluid.layers.dynamic_lstm( diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py index 8818cf96fa..be347cd531 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py @@ -56,7 +56,7 @@ BATCH_SIZE = 200 # fix the order of training data train_reader = paddle.batch( - paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE) + paddle.dataset.uci_housing.train(), batch_size=BATCH_SIZE, drop_last=False) # train_reader = paddle.batch( # paddle.reader.shuffle( diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py index a1ca6d981f..fa696acdfa 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py @@ -80,21 +80,6 @@ def encoder_decoder(): return rnn() -def to_lodtensor(data, place): - seq_lens = [len(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - res = core.LoDTensor() - res.set(flattened_data, place) - res.set_lod([lod]) - return res - - def main(): rnn_out = encoder_decoder() label = layers.data( @@ -122,18 +107,21 @@ def main(): exe.run(framework.default_startup_program()) + feed_order = [ + 'src_word_id', 'target_language_word', 'target_language_next_word' + ] + + feed_list = [ + fluid.default_main_program().global_block().var(var_name) + for var_name in feed_order + ] + feeder = fluid.DataFeeder(feed_list, place) + batch_id = 0 for pass_id in xrange(10): for data in train_data(): - word_data = to_lodtensor(map(lambda x: x[0], data), place) - trg_word = to_lodtensor(map(lambda x: x[1], data), place) - trg_word_next = to_lodtensor(map(lambda x: x[2], data), place) outs = exe.run(fluid.default_main_program(), - feed={ - 'src_word_id': word_data, - 'target_language_word': trg_word, - 'target_language_next_word': trg_word_next - }, + feed=feeder.feed(data), fetch_list=[avg_cost]) avg_cost_val = np.array(outs[0]) print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) + diff --git a/python/paddle/fluid/tests/test_concurrency.py b/python/paddle/fluid/tests/no_test_concurrency.py similarity index 100% rename from python/paddle/fluid/tests/test_concurrency.py rename to python/paddle/fluid/tests/no_test_concurrency.py diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index fead95ffda..21182393bd 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -43,10 +43,12 @@ list(REMOVE_ITEM TEST_OPS test_warpctc_op) list(REMOVE_ITEM TEST_OPS test_dist_train) list(REMOVE_ITEM TEST_OPS test_parallel_executor_crf) list(REMOVE_ITEM TEST_OPS test_parallel_executor_fetch_feed) +# TODO(wuyi): this test hungs on CI, will add it back later +list(REMOVE_ITEM TEST_OPS test_listen_and_serv_op) foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach(TEST_OP) py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=${WARPCTC_LIB_DIR} SERIAL) py_test_modules(test_dist_train MODULES test_dist_train SERIAL) -# tests that need to be done in fixed timeout -set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20) +py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SERIAL) +py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL) diff --git a/python/paddle/fluid/tests/unittests/benchmark.py b/python/paddle/fluid/tests/unittests/benchmark.py new file mode 100644 index 0000000000..e891ee932f --- /dev/null +++ b/python/paddle/fluid/tests/unittests/benchmark.py @@ -0,0 +1,113 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import unittest +import time +import itertools + +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid.op import Operator +from op_test import OpTest + + +class BenchmarkSuite(OpTest): + def timeit_function(self, callback, iters, *args, **kwargs): + assert iters != 0, "Iters should >= 1" + start = time.time() + for i in range(iters): + callback(*args, **kwargs) + elapse = time.time() - start + return elapse / iters + + def _assert_cpu_gpu_same(self, cpu_outs, gpu_outs, fetch_list, atol): + for item_cpu_out, item_gpu_out, variable in zip(cpu_outs, gpu_outs, + fetch_list): + # the cpu version is baseline, expect gpu version keep same with cpu version. + expect = item_cpu_out + expect_t = np.array(item_cpu_out) + actual = item_gpu_out + actual_t = np.array(item_gpu_out) + var_name = variable if isinstance(variable, + basestring) else variable.name + self.assertTrue( + np.allclose( + actual_t, expect_t, atol=atol), + "Output (" + var_name + ") has diff" + str(actual_t) + "\n" + + str(expect_t)) + self.assertListEqual(actual.lod(), + expect.lod(), + "Output (" + var_name + ") has different lod") + + def _get_input_names(self): + inputs = [] + for name, value in self.inputs.iteritems(): + if isinstance(value, list): + inputs.extend([sub_name for sub_name, _ in value]) + inputs.append(name) + return inputs + + def _get_output_names(self): + outputs = [] + for var_name, var in self.outputs.iteritems(): + if isinstance(var, list): + for sub_var_name, sub_var in var: + outputs.append(sub_var_name) + else: + outputs.append(var_name) + if len(outputs) == 0: + for out_name, out_dup in Operator.get_op_outputs(self.op_type): + outputs.append(str(out_name)) + return outputs + + def check_output_stability(self, atol=1e-8): + places = self._get_places() + if len(places) < 2: + return + cpu_outs, fetch_list = self._calc_output(places[0]) + gpu_outs, _ = self._calc_output(places[1]) + self._assert_cpu_gpu_same(cpu_outs, gpu_outs, fetch_list, atol) + + def timeit_output_with_place(self, place, iters): + return self.timeit_function(self.calc_output, iters, place) + + def timeit_output(self, iters=100): + places = self._get_places() + elapses = [] + for place in places: + elapses.append(self.timeit_output_with_place(place, iters)) + for place, elapse in zip(places, elapses): + print("One pass of ({2}_op) at {0} cost {1}".format( + str(place), elapse, self.op_type)) + + def timeit_grad_with_place(self, place, iters=100): + inputs_to_check = self._get_input_names() + output_names = self._get_output_names() + return self.timeit_function( + self._get_gradient, + iters, + inputs_to_check, + place, + output_names, + no_grad_set=None) + + def timeit_grad(self, iters=100): + places = self._get_places() + elapses = [] + for place in places: + elapses.append(self.timeit_grad_with_place(place, iters)) + for place, elapse in zip(places, elapses): + print("One pass of ({2}_grad_op) at {0} cost {1}".format( + str(place), elapse, self.op_type)) diff --git a/python/paddle/fluid/tests/unittests/benchmark_sum_op.py b/python/paddle/fluid/tests/unittests/benchmark_sum_op.py new file mode 100644 index 0000000000..91a5f1bca4 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/benchmark_sum_op.py @@ -0,0 +1,82 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +import paddle.fluid as fluid +from benchmark import BenchmarkSuite +from op_test import OpTest + +# This is a demo op test case for operator benchmarking and high resolution number stability alignment. + + +class TestSumOp(BenchmarkSuite): + def setUp(self): + self.op_type = "sum" + self.customize_testcase() + self.customize_fetch_list() + + def customize_fetch_list(self): + """ + customize fetch list, configure the wanted variables. + >>> self.fetch_list = ["Out"] + """ + self.fetch_list = ["Out"] + # pass + + def customize_testcase(self): + # a test case + x0 = np.random.random((300, 400)).astype('float32') + x1 = np.random.random((300, 400)).astype('float32') + x2 = np.random.random((300, 400)).astype('float32') + + # NOTE: if the output is empty, then it will autofilled by benchmarkSuite. + # only the output dtype is used, the shape, lod and data is computed from input. + self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} + self.outputs = {"Out": x0 + x1 + x2} + + def test_check_output(self): + """ + compare the output with customized output. In this case, + you should set the correct output by hands. + >>> self.outputs = {"Out": x0 + x1 + x2} + """ + self.check_output(atol=1e-8) + + def test_output_stability(self): + # compare the cpu gpu output in high resolution. + self.check_output_stability() + + def test_timeit_output(self): + """ + perf the op, time cost will be averged in iters. + output example + >>> One pass of (sum_op) at CPUPlace cost 0.000461330413818 + >>> One pass of (sum_op) at CUDAPlace(0) cost 0.000556070804596 + """ + self.timeit_output(iters=100) + + def test_timeit_grad(self): + """ + perf the op gradient, time cost will be averged in iters. + output example + >>> One pass of (sum_grad_op) at CPUPlace cost 0.00279935121536 + >>> One pass of (sum_grad_op) at CUDAPlace(0) cost 0.00500632047653 + """ + self.timeit_grad(iters=100) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index b611470fa1..307caae4b0 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -15,13 +15,17 @@ import unittest import numpy as np import random +import time import itertools -import paddle.fluid.core as core import collections + +import paddle.fluid as fluid +import paddle.fluid.core as core from paddle.fluid.backward import append_backward from paddle.fluid.op import Operator from paddle.fluid.executor import Executor -from paddle.fluid.framework import Program, OpProtoHolder +from paddle.fluid.framework import Program, OpProtoHolder, Variable +from testsuite import create_op, set_input, append_input_output, append_loss_ops def randomize_probability(batch_size, class_num, dtype='float32'): @@ -33,73 +37,6 @@ def randomize_probability(batch_size, class_num, dtype='float32'): return prob -def create_op(scope, op_type, inputs, outputs, attrs): - kwargs = dict() - - op_maker = core.op_proto_and_checker_maker - op_role_attr_name = op_maker.kOpRoleAttrName() - - if op_role_attr_name not in attrs: - attrs[op_role_attr_name] = int(op_maker.OpRole.Forward) - - def __create_var__(name, var_name): - scope.var(var_name).get_tensor() - kwargs[name].append(var_name) - - for in_name, in_dup in Operator.get_op_inputs(op_type): - if in_name in inputs: - kwargs[in_name] = [] - if in_dup: - sub_in = inputs[in_name] - for item in sub_in: - sub_in_name, _ = item[0], item[1] - __create_var__(in_name, sub_in_name) - else: - __create_var__(in_name, in_name) - - for out_name, out_dup in Operator.get_op_outputs(op_type): - if out_name in outputs: - kwargs[out_name] = [] - if out_dup: - sub_out = outputs[out_name] - for item in sub_out: - sub_out_name, _ = item[0], item[1] - __create_var__(out_name, sub_out_name) - else: - __create_var__(out_name, out_name) - - for attr_name in Operator.get_op_attr_names(op_type): - if attr_name in attrs: - kwargs[attr_name] = attrs[attr_name] - - return Operator(op_type, **kwargs) - - -def set_input(scope, op, inputs, place): - def __set_input__(var_name, var): - if isinstance(var, tuple) or isinstance(var, np.ndarray): - tensor = scope.find_var(var_name).get_tensor() - if isinstance(var, tuple): - tensor.set_lod(var[1]) - var = var[0] - tensor.set_dims(var.shape) - tensor.set(var, place) - elif isinstance(var, float): - scope.find_var(var_name).set_float(var) - elif isinstance(var, int): - scope.find_var(var_name).set_int(var) - - for in_name, in_dup in Operator.get_op_inputs(op.type()): - if in_name in inputs: - if in_dup: - sub_in = inputs[in_name] - for item in sub_in: - sub_in_name, sub_in_val = item[0], item[1] - __set_input__(sub_in_name, sub_in_val) - else: - __set_input__(in_name, inputs[in_name]) - - def get_numeric_gradient(place, scope, op, @@ -173,54 +110,15 @@ def get_numeric_gradient(place, return gradient_flat.reshape(tensor_to_check.get_dims()) -def append_input_output(block, op_proto, np_list, is_input): - '''Insert VarDesc and generate Python variable instance''' - proto_list = op_proto.inputs if is_input else op_proto.outputs - - def create_var(block, name, np_list, var_proto): - if name not in np_list: - assert var_proto.intermediate, "{} not found".format(name) - shape = None - lod_level = None - else: - np_value = np_list[name] - if isinstance(np_value, tuple): - shape = list(np_value[0].shape) - lod_level = len(np_value[1]) - else: - shape = list(np_value.shape) - lod_level = 0 - return block.create_var( - dtype="float32", shape=shape, lod_level=lod_level, name=name) - - var_dict = {} - for var_proto in proto_list: - var_name = str(var_proto.name) - if is_input: - if (var_name not in np_list) and var_proto.dispensable: - continue - assert (var_name in np_list) or (var_proto.dispensable), \ - "Missing {} as input".format(var_name) - if var_proto.duplicable: - assert isinstance(np_list[var_name], list), \ - "Duplicable {} should be set as list".format(var_name) - var_list = [] - for (name, np_value) in np_list[var_name]: - var_list.append( - create_var(block, name, {name: np_value}, var_proto)) - var_dict[var_name] = var_list - else: - var_dict[var_name] = create_var(block, var_name, np_list, var_proto) - - return var_dict - - class OpTest(unittest.TestCase): @classmethod def setUpClass(cls): '''Fix random seeds to remove randomness from tests''' cls._np_rand_state = np.random.get_state() cls._py_rand_state = random.getstate() + cls.call_once = False + cls.dtype = "float32" + cls.outputs = {} np.random.seed(123) random.seed(124) @@ -231,6 +129,31 @@ class OpTest(unittest.TestCase): np.random.set_state(cls._np_rand_state) random.setstate(cls._py_rand_state) + def try_call_once(self, data_type): + if not self.call_once: + self.call_once = True + self.dtype = data_type + + def infer_dtype_from_inputs_outputs(self, inputs, outputs): + def infer_dtype(numpy_dict): + assert isinstance( + numpy_dict, + dict), "self.inputs, self.outputs must be numpy_dict" + for var_name, var_value in numpy_dict.iteritems(): + if isinstance(var_value, (np.ndarray, np.generic)): + self.try_call_once(var_value.dtype) + elif isinstance(var_value, (list, tuple)): + # the case of self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} + if len(var_value) > 1 and isinstance(var_value[1], ( + np.ndarray, np.generic)): + instance = var_value[1] + self.try_call_once(instance[1].dtype) + else: + self.try_call_once("float32") + + infer_dtype(inputs) + infer_dtype(outputs) + def feed_var(self, input_vars, place): feed_map = {} for var_name in input_vars: @@ -254,18 +177,14 @@ class OpTest(unittest.TestCase): return feed_map - def calc_output(self, place): - outs, _ = self._calc_output(place) - return outs - - def _calc_output(self, place): + def _append_ops(self, block): op_proto = OpProtoHolder.instance().get_op_proto(self.op_type) - - program = Program() - block = program.global_block() - - inputs = append_input_output(block, op_proto, self.inputs, True) - outputs = append_input_output(block, op_proto, self.outputs, False) + "infer datatype from inputs and outputs for this test case" + self.infer_dtype_from_inputs_outputs(self.inputs, self.outputs) + inputs = append_input_output(block, op_proto, self.inputs, True, + self.dtype) + outputs = append_input_output(block, op_proto, self.outputs, False, + self.dtype) op = block.append_op( type=self.op_type, inputs=inputs, @@ -275,22 +194,68 @@ class OpTest(unittest.TestCase): op.desc.infer_var_type(block.desc) op.desc.infer_shape(block.desc) - fetch_list = [] - for var_name, var in outputs.iteritems(): - if var_name in self.outputs: + def _get_io_vars(self, block, numpy_inputs): + inputs = {} + for name, value in numpy_inputs.iteritems(): + if isinstance(value, list): + var_list = [ + block.var(sub_name) for sub_name, sub_value in value + ] + inputs[name] = var_list + else: + inputs[name] = block.var(name) + return inputs + + def _get_inputs(self, block): + return self._get_io_vars(block, self.inputs) + + def _get_outputs(self, block): + return self._get_io_vars(block, self.outputs) + + def calc_output(self, place): + outs, _ = self._calc_output(place) + return outs + + def _calc_output(self, place, parallel=False): + + program = Program() + block = program.global_block() + self._append_ops(block) + + inputs = self._get_inputs(block) + outputs = self._get_outputs(block) + feed_map = self.feed_var(inputs, place) + + if parallel: + use_cuda = False + if isinstance(place, fluid.CUDAPlace(0)): + use_cuda = True + executor = fluid.ParallelExecutor( + use_cuda=use_cuda, loss_name=loss.name, main_program=program) + else: + executor = Executor(place) + + fetch_list = getattr(self, "fetch_list", []) + # if the fetch_list is customized by user, we use it directly. + # if not, fill the fetch_list by the user configured outputs in test. + if len(fetch_list) == 0: + for var_name, var in outputs.iteritems(): if isinstance(var, list): for v in var: fetch_list.append(v) else: fetch_list.append(var) - - feed_map = self.feed_var(inputs, place) - - exe = Executor(place) - outs = exe.run(program, - feed=feed_map, - fetch_list=fetch_list, - return_numpy=False) + # if the fetch_list still empty, fill the fetch_list by the operator output. + if len(fetch_list) == 0: + for out_name, out_dup in Operator.get_op_outputs(self.op_type): + fetch_list.append(str(out_name)) + # fetch_list = map(block.var, fetch_list) + if not isinstance(fetch_list[0], Variable): + fetch_list = map(block.var, fetch_list) + outs = executor.run(program, + feed=feed_map, + fetch_list=fetch_list, + return_numpy=False) return outs, fetch_list def check_output_with_place(self, place, atol): @@ -346,17 +311,19 @@ class OpTest(unittest.TestCase): "Output (" + out_name + ") has different lod at " + str(place)) - def check_output(self, atol=1e-5): - places = [core.CPUPlace()] + def _get_places(self): + places = [fluid.CPUPlace()] if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type): places.append(core.CUDAPlace(0)) + return places + + def check_output(self, atol=1e-5): + places = self._get_places() for place in places: self.check_output_with_place(place, atol) def check_output_customized(self, checker): - places = [core.CPUPlace()] - if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type): - places.append(core.CUDAPlace(0)) + places = self._get_places() for place in places: outs = self.calc_output(place) outs = [np.array(out) for out in outs] @@ -389,9 +356,7 @@ class OpTest(unittest.TestCase): in_place=False, max_relative_error=0.005, user_defined_grads=None): - places = [core.CPUPlace()] - if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type): - places.append(core.CUDAPlace(0)) + places = self._get_places() for place in places: self.check_grad_with_place(place, inputs_to_check, output_names, no_grad_set, numeric_grad_delta, @@ -438,35 +403,6 @@ class OpTest(unittest.TestCase): max_relative_error, "Gradient Check On %s" % str(place)) - @staticmethod - def _create_var_descs_(block, var_dict): - # FIXME: Try unify with `append_input_output` - for param_name in var_dict: - var = var_dict[param_name] - if not isinstance(var, list) and not isinstance(var, tuple): - var = [(param_name, var, None)] - if not isinstance(var[0], list) and not isinstance(var[0], tuple): - var = [(param_name, var[0], var[1])] - - for i, item in enumerate(var): - if not isinstance(item[0], basestring): - item = [[param_name] + list(item)] - if len(item) == 2: - if isinstance(item[1], tuple): - var[i] = [item[0], item[1][0], item[1][1]] - else: - # only set var name and value, set lod to None - var[i] = list(item) + [None] - var_descs = [(block.create_var( - name=name, shape=each.shape, dtype=each.dtype), each, lod) - for name, each, lod in var] - - yield param_name, var_descs - - @staticmethod - def _merge_list(iterable): - return reduce(lambda a, b: list(a) + list(b), iterable, []) - @staticmethod def _numpy_to_lod_tensor(np_value, lod, place): tensor = core.LoDTensor() @@ -497,83 +433,31 @@ class OpTest(unittest.TestCase): input.dtype = np.uint16 return input - def _get_gradient(self, input_to_check, place, output_names, no_grad_set): + def _get_gradient(self, + input_to_check, + place, + output_names, + no_grad_set, + parallel=False): prog = Program() block = prog.global_block() - inputs_with_np = { - key: value - for (key, value) in OpTest._create_var_descs_( - block, getattr(self, 'inputs', {})) - } - outputs_with_np = { - key: val - for (key, val) in OpTest._create_var_descs_( - block, getattr(self, 'outputs', {})) - } - inputs = { - k: [item[0] for item in inputs_with_np[k]] - for k in inputs_with_np - } - outputs = { - k: [item[0] for item in outputs_with_np[k]] - for k in outputs_with_np - } - - op = block.append_op( - type=self.op_type, - inputs=inputs, - outputs=outputs, - attrs=getattr(self, 'attrs', {})) - - # infer variable type and infer shape in compile-time - op.desc.infer_var_type(block.desc) - op.desc.infer_shape(block.desc) - - mean_inputs = map(block.var, output_names) - - if len(mean_inputs) == 1: - loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1]) - op = block.append_op( - inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean') - op.desc.infer_var_type(block.desc) - op.desc.infer_shape(block.desc) - else: - avg_sum = [] - for cur_loss in mean_inputs: - cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1]) - op = block.append_op( - inputs={"X": [cur_loss]}, - outputs={"Out": [cur_avg_loss]}, - type="mean") - op.desc.infer_var_type(block.desc) - op.desc.infer_shape(block.desc) - avg_sum.append(cur_avg_loss) - - loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1]) - op_sum = block.append_op( - inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum') - op_sum.desc.infer_var_type(block.desc) - op_sum.desc.infer_shape(block.desc) - - loss = block.create_var(dtype=loss_sum.dtype, shape=[1]) - op_loss = block.append_op( - inputs={"X": loss_sum}, - outputs={"Out": loss}, - type='scale', - attrs={'scale': 1.0 / float(len(avg_sum))}) - op_loss.desc.infer_var_type(block.desc) - op_loss.desc.infer_shape(block.desc) - + self._append_ops(block) + loss = append_loss_ops(block, output_names) param_grad_list = append_backward( loss=loss, parameter_list=input_to_check, no_grad_set=no_grad_set) - feed_dict = { - item[0].name: OpTest._numpy_to_lod_tensor(item[1], item[2], place) - for p_name in inputs_with_np for item in inputs_with_np[p_name] - } + inputs = self._get_inputs(block) + feed_dict = self.feed_var(inputs, place) fetch_list = [g for p, g in param_grad_list] - executor = Executor(place) + if parallel: + use_cuda = False + if isinstance(place, fluid.CUDAPlace(0)): + use_cuda = True + executor = fluid.ParallelExecutor( + use_cuda=use_cuda, loss_name=loss.name, main_program=program) + else: + executor = Executor(place) return map(np.array, executor.run(prog, feed_dict, fetch_list, return_numpy=False)) diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py index c9c3c64871..829c5a1a5f 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +import multiprocessing +import os import unittest import paddle.fluid as fluid import time @@ -23,6 +25,7 @@ __all__ = ['TestParallelExecutorBase'] class TestParallelExecutorBase(unittest.TestCase): def check_network_convergence(self, method, + use_cuda=True, memory_opt=True, iter=50, batch_size=None, @@ -53,7 +56,7 @@ class TestParallelExecutorBase(unittest.TestCase): adam.minimize(loss) if memory_opt: fluid.memory_optimize(main) - place = fluid.CUDAPlace(0) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() startup_exe = fluid.Executor(place) startup_exe.run(startup) exec_strategy = fluid.ExecutionStrategy() @@ -64,7 +67,7 @@ class TestParallelExecutorBase(unittest.TestCase): if use_parallel_executor: exe = fluid.ParallelExecutor( - True, + use_cuda, loss_name=loss.name, exec_strategy=exec_strategy, build_strategy=build_strategy) @@ -72,7 +75,9 @@ class TestParallelExecutorBase(unittest.TestCase): exe = fluid.Executor(place=place) if batch_size is not None: - batch_size *= fluid.core.get_cuda_device_count() + batch_size *= fluid.core.get_cuda_device_count( + ) if use_cuda else int( + os.environ.get('CPU_NUM', multiprocessing.cpu_count())) begin = time.time() first_loss, = run_executor( exe=exe, feed=feed_dict, fetch_list=[loss.name]) diff --git a/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py b/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py new file mode 100644 index 0000000000..e04412f809 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py @@ -0,0 +1,82 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class BaseTestCase(OpTest): + def initTestCase(self): + self.op_type = 'arg_min' + self.dims = (3, 4, 5) + self.dtype = 'float32' + self.axis = 0 + + def setUp(self): + self.initTestCase() + self.x = (1000 * np.random.random(self.dims)).astype(self.dtype) + self.inputs = {'X': self.x} + self.attrs = {'axis': self.axis} + if self.op_type == "arg_min": + self.outputs = {'Out': np.argmin(self.x, axis=self.axis)} + else: + self.outputs = {'Out': np.argmax(self.x, axis=self.axis)} + + def test_check_output(self): + self.check_output() + + +class TestCase0(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_max' + self.dims = (3, 4, 5) + self.dtype = 'float32' + self.axis = 0 + + +class TestCase1(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_min' + self.dims = (3, 4) + self.dtype = 'float64' + self.axis = 1 + + +class TestCase2(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_max' + self.dims = (3, 4) + self.dtype = 'int64' + self.axis = 0 + + +class TestCase3(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_max' + self.dims = (3, ) + self.dtype = 'int64' + self.axis = 0 + + +class TestCase4(BaseTestCase): + def initTestCase(self): + self.op_type = 'arg_min' + self.dims = (1, ) + self.dtype = 'int32' + self.axis = 0 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py index bffb4f3b66..87c11e7880 100644 --- a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py +++ b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py @@ -17,7 +17,10 @@ import numpy as np from op_test import OpTest -def bilinear_interp_np(input, out_h, out_w): +def bilinear_interp_np(input, out_h, out_w, out_size): + if out_size is not None: + out_h = out_size[0] + out_w = out_size[1] batch_size, channel, in_h, in_w = input.shape if out_h > 1: ratio_h = (in_h - 1.0) / (out_h - 1.0) @@ -49,12 +52,15 @@ def bilinear_interp_np(input, out_h, out_w): class TestBilinearInterpOp(OpTest): def setUp(self): + self.out_size = None self.init_test_case() self.op_type = "bilinear_interp" input_np = np.random.random(self.input_shape).astype("float32") - output_np = bilinear_interp_np(input_np, self.out_h, self.out_w) - + output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, + self.out_size) self.inputs = {'X': input_np} + if self.out_size is not None: + self.inputs['OutSize'] = self.out_size self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} self.outputs = {'Out': output_np} @@ -68,6 +74,7 @@ class TestBilinearInterpOp(OpTest): self.input_shape = [2, 3, 4, 4] self.out_h = 2 self.out_w = 2 + self.out_size = np.array([3, 3]).astype("int32") class TestCase1(TestBilinearInterpOp): @@ -91,5 +98,29 @@ class TestCase3(TestBilinearInterpOp): self.out_w = 128 +class TestCase4(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + self.out_size = np.array([2, 2]).astype("int32") + + +class TestCase5(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + self.out_size = np.array([11, 11]).astype("int32") + + +class TestCase6(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + self.out_size = np.array([65, 129]).astype("int32") + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_box_coder_op.py b/python/paddle/fluid/tests/unittests/test_box_coder_op.py index 56f5af91d8..b4c48d85f2 100644 --- a/python/paddle/fluid/tests/unittests/test_box_coder_op.py +++ b/python/paddle/fluid/tests/unittests/test_box_coder_op.py @@ -19,7 +19,8 @@ import math from op_test import OpTest -def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): +def box_coder(target_box, prior_box, prior_box_var, output_box, code_type, + box_normalized): prior_box_x = ( (prior_box[:, 2] + prior_box[:, 0]) / 2).reshape(1, prior_box.shape[0]) prior_box_y = ( @@ -30,6 +31,9 @@ def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): (prior_box[:, 3] - prior_box[:, 1])).reshape(1, prior_box.shape[0]) prior_box_var = prior_box_var.reshape(1, prior_box_var.shape[0], prior_box_var.shape[1]) + if not box_normalized: + prior_box_height = prior_box_height + 1 + prior_box_width = prior_box_width + 1 if (code_type == "EncodeCenterSize"): target_box_x = ((target_box[:, 2] + target_box[:, 0]) / 2).reshape( @@ -40,6 +44,9 @@ def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): target_box.shape[0], 1) target_box_height = ((target_box[:, 3] - target_box[:, 1])).reshape( target_box.shape[0], 1) + if not box_normalized: + target_box_height = target_box_height + 1 + target_box_width = target_box_width + 1 output_box[:,:,0] = (target_box_x - prior_box_x) / prior_box_width / \ prior_box_var[:,:,0] @@ -64,9 +71,13 @@ def box_coder(target_box, prior_box, prior_box_var, output_box, code_type): output_box[:, :, 1] = target_box_y - target_box_height / 2 output_box[:, :, 2] = target_box_x + target_box_width / 2 output_box[:, :, 3] = target_box_y + target_box_height / 2 + if not box_normalized: + output_box[:, :, 2] = output_box[:, :, 2] - 1 + output_box[:, :, 3] = output_box[:, :, 3] - 1 -def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type): +def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type, + box_normalized): n = target_box.shape[0] m = prior_box.shape[0] output_box = np.zeros((n, m, 4), dtype=np.float32) @@ -74,11 +85,11 @@ def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type): if (code_type == "EncodeCenterSize"): box_coder(target_box[lod[i]:lod[i + 1], :], prior_box, prior_box_var, output_box[lod[i]:lod[i + 1], :, :], - code_type) + code_type, box_normalized) elif (code_type == "DecodeCenterSize"): box_coder(target_box[lod[i]:lod[i + 1], :, :], prior_box, prior_box_var, output_box[lod[i]:lod[i + 1], :, :], - code_type) + code_type, box_normalized) return output_box @@ -93,15 +104,45 @@ class TestBoxCoderOp(OpTest): prior_box_var = np.random.random((10, 4)).astype('float32') target_box = np.random.random((5, 10, 4)).astype('float32') code_type = "DecodeCenterSize" + box_normalized = False output_box = batch_box_coder(prior_box, prior_box_var, target_box, - lod[0], code_type) + lod[0], code_type, box_normalized) self.inputs = { 'PriorBox': prior_box, 'PriorBoxVar': prior_box_var, 'TargetBox': target_box, } - self.attrs = {'code_type': 'decode_center_size'} + self.attrs = { + 'code_type': 'decode_center_size', + 'box_normalized': False + } + self.outputs = {'OutputBox': output_box} + + +class TestBoxCoderOpWithoutBoxVar(OpTest): + def test_check_output(self): + self.check_output() + + def setUp(self): + self.op_type = "box_coder" + lod = [[0, 1, 2, 3, 4, 5]] + prior_box = np.random.random((10, 4)).astype('float32') + prior_box_var = np.ones((10, 4)).astype('float32') + target_box = np.random.random((5, 10, 4)).astype('float32') + code_type = "DecodeCenterSize" + box_normalized = False + output_box = batch_box_coder(prior_box, prior_box_var, target_box, + lod[0], code_type, box_normalized) + + self.inputs = { + 'PriorBox': prior_box, + 'TargetBox': target_box, + } + self.attrs = { + 'code_type': 'decode_center_size', + 'box_normalized': False + } self.outputs = {'OutputBox': output_box} @@ -116,15 +157,16 @@ class TestBoxCoderOpWithLoD(OpTest): prior_box_var = np.random.random((10, 4)).astype('float32') target_box = np.random.random((20, 4)).astype('float32') code_type = "EncodeCenterSize" + box_normalized = True output_box = batch_box_coder(prior_box, prior_box_var, target_box, - lod[0], code_type) + lod[0], code_type, box_normalized) self.inputs = { 'PriorBox': prior_box, 'PriorBoxVar': prior_box_var, 'TargetBox': (target_box, lod), } - self.attrs = {'code_type': 'encode_center_size'} + self.attrs = {'code_type': 'encode_center_size', 'box_normalized': True} self.outputs = {'OutputBox': output_box} diff --git a/python/paddle/fluid/tests/unittests/test_checkpoint.py b/python/paddle/fluid/tests/unittests/test_checkpoint.py new file mode 100644 index 0000000000..e22400a045 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_checkpoint.py @@ -0,0 +1,75 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddle.fluid as fluid +import unittest +import os +import tempfile + + +class TestCheckpoint(unittest.TestCase): + def setUp(self): + self.dirname = tempfile.mktemp() + self.max_num_checkpoints = 3 + self.epoch_interval = 1 + self.step_interval = 1 + self.trainer_id = 0 + self.chief = self.trainer_id == 0 + self.place = fluid.CPUPlace() + self.epoch_id = 100 + self.step_id = 20 + + def test_checkpoint(self): + self.save_checkpoint() + serial = fluid.io.get_latest_checkpoint_serial(self.dirname) + self.assertTrue(serial >= 0) + trainer_args = ["epoch_id", "step_id"] + epoch_id, step_id = fluid.io.load_trainer_args( + self.dirname, serial, self.trainer_id, trainer_args) + self.assertEqual(self.step_id, int(step_id)) + self.assertEqual(self.epoch_id, int(epoch_id)) + + program = fluid.Program() + with fluid.program_guard(program): + exe = fluid.Executor(self.place) + fluid.io.load_checkpoint(exe, self.dirname, serial, program) + + fluid.io.clean_checkpoint(self.dirname, delete_dir=True) + self.assertFalse(os.path.isdir(self.dirname)) + + def save_checkpoint(self): + config = fluid.CheckpointConfig(self.dirname, self.max_num_checkpoints, + self.epoch_interval, self.step_interval) + + trainer_args = {} + trainer_args["epoch_id"] = self.epoch_id + trainer_args["step_id"] = self.step_id + + program = fluid.Program() + with fluid.program_guard(program): + program.global_block().create_var( + name="scale_0", + psersistable=True, + dtype="float32", + shape=[32, 32]) + + exe = fluid.Executor(self.place) + for i in xrange(10): + fluid.io.save_checkpoint(exe, config.checkpoint_dir, + self.trainer_id, trainer_args, program, + config.max_num_checkpoints) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_crop_op.py b/python/paddle/fluid/tests/unittests/test_crop_op.py index 20cc3a643f..4016089c01 100644 --- a/python/paddle/fluid/tests/unittests/test_crop_op.py +++ b/python/paddle/fluid/tests/unittests/test_crop_op.py @@ -42,9 +42,9 @@ class TestCropOp(OpTest): def setUp(self): self.op_type = "crop" self.crop_by_input = False + self.offset_by_input = False self.attrs = {} self.initTestCase() - self.attrs['offsets'] = self.offsets if self.crop_by_input: self.inputs = { 'X': np.random.random(self.x_shape).astype("float32"), @@ -55,6 +55,10 @@ class TestCropOp(OpTest): self.inputs = { 'X': np.random.random(self.x_shape).astype("float32"), } + if self.offset_by_input: + self.inputs['Offsets'] = np.array(self.offsets).astype('int32') + else: + self.attrs['offsets'] = self.offsets self.outputs = { 'Out': crop(self.inputs['X'], self.offsets, self.crop_shape) } @@ -101,5 +105,22 @@ class TestCase4(TestCropOp): self.crop_by_input = True +class TestCase5(TestCropOp): + def initTestCase(self): + self.x_shape = (3, 4, 5) + self.crop_shape = [2, 2, 3] + self.offsets = [1, 0, 2] + self.offset_by_input = True + + +class TestCase6(TestCropOp): + def initTestCase(self): + self.x_shape = (10, 9, 14) + self.crop_shape = [3, 3, 5] + self.offsets = [3, 5, 4] + self.crop_by_input = True + self.offset_by_input = True + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index fa49bd41a5..b4379ad447 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -13,39 +13,16 @@ # limitations under the License. import unittest - import paddle.fluid as fluid -import paddle.fluid.core as core -import paddle.fluid.layers as layers from paddle.fluid.transpiler.distribute_transpiler import delete_ops -import numpy + +from transpiler_test import TranspilerTest -class TestDistTranspiler(unittest.TestCase): +class TestDistTranspiler(TranspilerTest): def setUp(self): - self.trainer_id = 0 - self.trainers = 2 - self.pservers = 2 - self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175" self.current_pserver_ep = "127.0.0.1:6174" - def net_conf(self): - x = fluid.layers.data(name='x', shape=[1000], dtype='float32') - - y_predict = fluid.layers.fc(input=x, - size=1000, - act=None, - param_attr=fluid.ParamAttr(name='fc_w')) - - y = fluid.layers.data(name='y', shape=[1], dtype='float32') - - cost = fluid.layers.square_error_cost(input=y_predict, label=y) - avg_cost = fluid.layers.mean(cost) - sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) - - optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) - return optimize_ops, params_grads - def test_transpiler(self): trainer = self.get_trainer() pserver, startup = self.get_pserver(self.current_pserver_ep) @@ -70,14 +47,6 @@ class TestDistTranspiler(unittest.TestCase): fc_w_var = startup.global_block().var("fc_w.block1") self.assertEqual(fc_w_var.shape, (500, 1000)) - def get_main_program(self): - main = fluid.Program() - - with fluid.program_guard(main): - self.net_conf() - - return main - def get_expect_trainer_ops(self): trainer = fluid.Program() @@ -86,31 +55,12 @@ class TestDistTranspiler(unittest.TestCase): delete_ops(trainer.global_block(), optimize_ops) ops = [op.type for op in trainer.global_block().ops] + [ - "split_byref", "send_vars", "send_barrier", "recv", "recv", + "split_byref", "send", "send_barrier", "recv", "recv", "fetch_barrier", "concat" ] - ops.insert(ops.index("elementwise_add_grad") + 1, "send_vars") + ops.insert(ops.index("elementwise_add_grad") + 1, "send") return ops - def get_trainer(self): - return self._transpiler_instance().get_trainer_program() - - def get_pserver(self, ep): - t = self._transpiler_instance() - pserver = t.get_pserver_program(ep) - startup = t.get_startup_program(ep, pserver) - return pserver, startup - - def _transpiler_instance(self): - main = self.get_main_program() - t = fluid.DistributeTranspiler() - t.transpile( - self.trainer_id, - program=main, - pservers=self.pserver_eps, - trainers=self.trainers) - return t - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py b/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py index 2232939075..95af51f1b2 100644 --- a/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py +++ b/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py @@ -30,9 +30,6 @@ class Memory(object): assert val.dtype == self.ex.dtype self.cur = val - def ex(self): - return self.ex - def next(self): self.ex = self.cur self.cur = None diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py index 1f52bd90d0..96d47906a0 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py @@ -252,5 +252,25 @@ class TestFP16ElementwiseAddOp_rowwise_add_1(TestFP16ElementwiseAddOp): self.axis = 1 +class TestElementwiseAddOp_channelwise_add(TestElementwiseAddOp): + def init_input_output(self): + self.x = np.random.rand(3, 20, 20).astype(self.dtype) + self.y = np.random.rand(3, 1, 1).astype(self.dtype) + self.out = self.x + self.y + + def init_axis(self): + self.axis = -1 + + +class TestFP16ElementwiseAddOp_channelwise_add(TestFP16ElementwiseAddOp): + def init_input_output(self): + self.x = np.random.rand(3, 10, 20).astype(self.dtype) + self.y = np.random.rand(3, 1, 1).astype(self.dtype) + self.out = self.x + self.y + + def init_axis(self): + self.axis = -1 + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_gather_op.py b/python/paddle/fluid/tests/unittests/test_gather_op.py index 6fd043c27e..4ae9086480 100644 --- a/python/paddle/fluid/tests/unittests/test_gather_op.py +++ b/python/paddle/fluid/tests/unittests/test_gather_op.py @@ -20,8 +20,9 @@ from op_test import OpTest class TestGatherOp(OpTest): def setUp(self): self.op_type = "gather" - xnp = np.random.random((10, 20)).astype("float32") - self.inputs = {'X': xnp, 'Index': np.array([1, 3, 5]).astype("int32")} + self.config() + xnp = np.random.random(self.x_shape).astype("float32") + self.inputs = {'X': xnp, 'Index': np.array(self.index).astype("int32")} self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]} def test_check_output(self): @@ -30,6 +31,16 @@ class TestGatherOp(OpTest): def test_check_grad(self): self.check_grad(['X'], 'Out') + def config(self): + self.x_shape = (10, 20) + self.index = [1, 3, 5] + + +class TestCase1(TestGatherOp): + def config(self): + self.x_shape = (10) + self.index = [1, 3, 5] + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 63ae51c4f4..f6e516bbe7 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -379,13 +379,35 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(output) print(str(program)) - def test_upsampling_bilinear2d(self): + def test_resize_bilinear(self): program = Program() with program_guard(program): x = layers.data(name='x', shape=[3, 9, 6], dtype="float32") - output = layers.upsampling_bilinear2d(x, out_shape=[12, 12]) + output = layers.resize_bilinear(x, out_shape=[12, 12]) self.assertIsNotNone(output) - output = layers.upsampling_bilinear2d(x, scale=3) + output = layers.resize_bilinear(x, scale=3) + self.assertIsNotNone(output) + print(str(program)) + + def test_polygon_box_transform(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[8, 4, 4], dtype="float32") + output = layers.polygon_box_transform(input=x) + self.assertIsNotNone(output) + print(str(program)) + + def test_l2_normalize(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[8, 7, 10], dtype="float32") + output = layers.l2_normalize(x, axis=1) + + def test_maxout(self): + program = Program() + with program_guard(program): + data = layers.data(name='x', shape=[8, 6, 6], dtype="float32") + output = layers.maxout(x=data, groups=2) self.assertIsNotNone(output) print(str(program)) diff --git a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py index cf89f9d0eb..d1d709551c 100644 --- a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py +++ b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py @@ -23,7 +23,7 @@ from multiprocessing import Process from op_test import OpTest -def run_pserver(use_cuda, sync_mode, ip, port, trainer_count, trainer_id): +def run_pserver(use_cuda, sync_mode, ip, port, trainers, trainer_id): x = fluid.layers.data(name='x', shape=[1], dtype='float32') y_predict = fluid.layers.fc(input=x, size=1, act=None) y = fluid.layers.data(name='y', shape=[1], dtype='float32') @@ -39,15 +39,8 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainer_count, trainer_id): place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) - port = os.getenv("PADDLE_INIT_PORT", port) - pserver_ips = os.getenv("PADDLE_INIT_PSERVERS", ip) # ip,ip... - eplist = [] - for ip in pserver_ips.split(","): - eplist.append(':'.join([ip, port])) - pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("TRAINERS", trainer_count)) - current_endpoint = os.getenv("POD_IP", ip) + ":" + port - trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID", trainer_id)) + pserver_endpoints = ip + ":" + port + current_endpoint = ip + ":" + port t = fluid.DistributeTranspiler() t.transpile( trainer_id, @@ -62,47 +55,52 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainer_count, trainer_id): class TestListenAndServOp(OpTest): def setUp(self): - self.sleep_time = 5 + self.ps_timeout = 5 self.ip = "127.0.0.1" self.port = "6173" - self.trainer_count = 1 + self.trainers = 1 self.trainer_id = 1 - def _raise_signal(self, parent_pid, raised_signal): - time.sleep(self.sleep_time) - ps_command = subprocess.Popen( - "ps -o pid --ppid %d --noheaders" % parent_pid, - shell=True, - stdout=subprocess.PIPE) - ps_output = ps_command.stdout.read() - retcode = ps_command.wait() - assert retcode == 0, "ps command returned %d" % retcode - - for pid_str in ps_output.split("\n")[:-1]: - try: - os.kill(int(pid_str), raised_signal) - except Exception: - continue - def _start_pserver(self, use_cuda, sync_mode): p = Process( target=run_pserver, - args=(use_cuda, sync_mode, self.ip, self.port, self.trainer_count, + args=(use_cuda, sync_mode, self.ip, self.port, self.trainers, self.trainer_id)) p.start() + return p.pid + + def _wait_ps_ready(self, pid): + start_left_time = self.ps_timeout + sleep_time = 0.5 + while True: + assert start_left_time >= 0, "wait ps ready failed" + time.sleep(sleep_time) + try: + # the listen_and_serv_op would touch a file which contains the listen port + # on the /tmp directory until it was ready to process all the RPC call. + os.stat("/tmp/paddle.%d.port" % pid) + return + except os.error: + start_left_time -= sleep_time + + def test_rpc_interfaces(self): + # TODO(Yancey1989): need to make sure the rpc interface correctly. + pass def test_handle_signal_in_serv_op(self): # run pserver on CPU in sync mode - self._start_pserver(False, True) + pid = self._start_pserver(False, True) + self._wait_ps_ready(pid) - # raise SIGINT to pserver - self._raise_signal(os.getpid(), signal.SIGINT) + # raise SIGTERM to pserver + os.kill(pid, signal.SIGTERM) # run pserver on CPU in async mode - self._start_pserver(False, False) + pid = self._start_pserver(False, False) + self._wait_ps_ready(pid) # raise SIGTERM to pserver - self._raise_signal(os.getpid(), signal.SIGTERM) + os.kill(pid, signal.SIGTERM) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_lstm_op.py b/python/paddle/fluid/tests/unittests/test_lstm_op.py index f8ff5a3361..e726f99d49 100644 --- a/python/paddle/fluid/tests/unittests/test_lstm_op.py +++ b/python/paddle/fluid/tests/unittests/test_lstm_op.py @@ -194,107 +194,104 @@ class TestLstmOp(OpTest): ['Input', 'Weight', 'Bias'], ['Hidden'], max_relative_error=5e-4) -class TestLstmOpHasInitial(TestLstmOp): - def set_argument(self): - self.lod = [[0, 2, 5, 7]] - self.D = 16 - - self.act_gate = 'sigmoid' - self.act_cell = 'tanh' - self.act_cand = 'tanh' - - self.has_initial_state = True - self.is_reverse = True - self.use_peepholes = True - - def test_check_grad(self): - # TODO(qingqing) remove folowing lines after the check_grad is refined. - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight', 'Bias', 'H0', 'C0'], ['Hidden'], - max_relative_error=5e-4) - - def test_check_grad_ingore_bias(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('Bias')) - - def test_check_grad_ingore_weight(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Bias'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('Weight')) - - def test_check_grad_ingore_input(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Weight', 'Bias'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('Input')) - - def test_check_grad_ingore_h0(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight', 'Bias', 'C0'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('H0')) - - def test_check_grad_ingore_c0(self): - N = len(self.lod[0]) - 1 - self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') - self.outputs['BatchCellPreAct'] = np.zeros( - (N, self.D)).astype('float64') - self.check_grad( - ['Input', 'Weight', 'Bias', 'H0'], ['Hidden'], - max_relative_error=5e-4, - no_grad_set=set('C0')) - - -class TestLstmOpRerverse(TestLstmOp): - def set_argument(self): - self.lod = [[0, 2, 5, 7]] - self.D = 16 - - self.act_gate = 'sigmoid' - self.act_cell = 'tanh' - self.act_cand = 'tanh' - - self.has_initial_state = False - self.is_reverse = True - self.use_peepholes = True - - -class TestLstmOpNotUsePeepholes(TestLstmOp): - def set_argument(self): - self.lod = [[0, 2, 5, 7]] - self.D = 16 - - self.act_gate = 'sigmoid' - self.act_cell = 'tanh' - self.act_cand = 'tanh' - - self.has_initial_state = False - self.is_reverse = True - self.use_peepholes = False - +# class TestLstmOpHasInitial(TestLstmOp): +# def set_argument(self): +# self.lod = [[0, 2, 5, 7]] +# self.D = 16 + +# self.act_gate = 'sigmoid' +# self.act_cell = 'tanh' +# self.act_cand = 'tanh' + +# self.has_initial_state = True +# self.is_reverse = True +# self.use_peepholes = True + +# def test_check_grad(self): +# # TODO(qingqing) remove folowing lines after the check_grad is refined. +# N = len(self.lod[0]) - 1 +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight', 'Bias', 'H0', 'C0'], ['Hidden'], +# max_relative_error=5e-4) + +# def test_check_grad_ingore_bias(self): +# N = len(self.lod[0]) - 1 +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('Bias')) + +# def test_check_grad_ingore_weight(self): +# N = len(self.lod[0]) - 1 +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Bias'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('Weight')) + +# def test_check_grad_ingore_input(self): +# N = len(self.lod[0]) - 1 +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Weight', 'Bias'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('Input')) + +# def test_check_grad_ingore_h0(self): +# N = len(self.lod[0]) - 1 +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight', 'Bias', 'C0'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('H0')) + +# def test_check_grad_ingore_c0(self): +# N = len(self.lod[0]) - 1 +# self.outputs['BatchGate'] = np.zeros((N, 4 * self.D)).astype('float64') +# self.outputs['BatchCellPreAct'] = np.zeros( +# (N, self.D)).astype('float64') +# self.check_grad( +# ['Input', 'Weight', 'Bias', 'H0'], ['Hidden'], +# max_relative_error=5e-4, +# no_grad_set=set('C0')) + +# class TestLstmOpRerverse(TestLstmOp): +# def set_argument(self): +# self.lod = [[0, 2, 5, 7]] +# self.D = 16 + +# self.act_gate = 'sigmoid' +# self.act_cell = 'tanh' +# self.act_cand = 'tanh' + +# self.has_initial_state = False +# self.is_reverse = True +# self.use_peepholes = True + +# class TestLstmOpNotUsePeepholes(TestLstmOp): +# def set_argument(self): +# self.lod = [[0, 2, 5, 7]] +# self.D = 16 + +# self.act_gate = 'sigmoid' +# self.act_cell = 'tanh' +# self.act_cand = 'tanh' + +# self.has_initial_state = False +# self.is_reverse = True +# self.use_peepholes = False if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_mean_iou.py b/python/paddle/fluid/tests/unittests/test_mean_iou.py new file mode 100644 index 0000000000..64d42b693b --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_mean_iou.py @@ -0,0 +1,114 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import division +import unittest +import numpy as np +from op_test import OpTest + + +def compute_mean_iou(predictions, labels, num_classes, in_wrongs, in_corrects, + in_mean_ious): + assert predictions.shape == labels.shape + predictions = predictions.flatten() + labels = labels.flatten() + + out_wrong = np.zeros([num_classes]).astype("int32") + for _, wrong in in_wrongs: + out_wrong += wrong + out_correct = np.zeros([num_classes]).astype("int32") + for _, correct in in_corrects: + out_correct += correct + + for pred, label in zip(predictions, labels): + if pred == label: + out_correct[pred] += 1 + else: + out_wrong[pred] += 1 + out_wrong[label] += 1 + + denominator = out_wrong + out_correct + valid_count = (denominator != 0).sum() + denominator = np.where(denominator > 0, denominator, + np.ones(denominator.shape)) + mean_iou = (out_correct / denominator).sum() / valid_count + + for _, in_mean_iou in in_mean_ious: + mean_iou += in_mean_iou + return mean_iou, out_wrong, out_correct + + +class TestMeanIOUOp(OpTest): + def setUp(self): + self.config() + self.op_type = "mean_iou" + predictions = np.random.randint(0, self.num_classes, + self.image_size).astype("int32") + labels = np.random.randint(0, self.num_classes, + self.image_size).astype("int32") + + in_wrongs = [] + for i in range(self.in_wrong_num): + in_wrongs.append(("in_wrong_%d" % i, np.random.randint( + 0, 10, [self.num_classes]).astype("int32"))) + + in_corrects = [] + for i in range(self.in_correct_num): + in_corrects.append(("in_correct_%d" % i, np.random.randint( + 0, 10, [self.num_classes]).astype("int32"))) + + in_mean_ious = [] + for i in range(self.in_mean_iou_num): + in_mean_ious.append(("in_mean_iou_%d" % i, np.random.uniform( + 0, 1, [1]).astype("float32"))) + + self.inputs = { + 'Predictions': predictions, + 'Labels': labels, + 'InWrongs': in_wrongs, + 'InCorrects': in_corrects, + 'InMeanIou': in_mean_ious + } + self.attrs = {'num_classes': long(self.num_classes)} + mean_iou, out_wrong, out_correct = compute_mean_iou( + predictions, labels, self.num_classes, in_wrongs, in_corrects, + in_mean_ious) + self.outputs = { + 'OutMeanIou': mean_iou, + 'OutWrong': out_wrong, + 'OutCorrect': out_correct + } + + def config(self): + self.num_classes = 10 + self.image_size = [128, 128] + self.in_wrong_num = 0 + self.in_correct_num = 0 + self.in_mean_iou_num = 0 + + def test_check_output(self): + self.check_output() + + +class TestCase1(TestMeanIOUOp): + def config(self): + self.num_classes = 5 + self.image_size = [100, 128] + self.in_wrong_num = 2 + self.in_correct_num = 2 + self.in_mean_iou_num = 2 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_merge_ids_op.py b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py new file mode 100644 index 0000000000..f209bdf30f --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py @@ -0,0 +1,38 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestMergeIdsOp(OpTest): + def setUp(self): + self.op_type = "merge_ids" + ids = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64') + x0 = np.array([[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]]).astype('float32') + x1 = np.array([]).astype('float32') + x2 = np.array([[0.4, 0.5], [0.4, 0.5], [0.5, 0.6], + [0.5, 0.6]]).astype('float32') + out = np.array([[0.1, 0.2], [0.4, 0.5], [0.4, 0.5], [0.2, 0.3], + [0.5, 0.6], [0.5, 0.6], [0.3, 0.4]]).astype('float32') + self.inputs = {'Ids': ids, "X": [('x0', x0), ('x1', x1), ('x2', x2)]} + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_mul_op.py b/python/paddle/fluid/tests/unittests/test_mul_op.py index 862b7f8cb9..bbc782c1bc 100644 --- a/python/paddle/fluid/tests/unittests/test_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_mul_op.py @@ -22,8 +22,8 @@ class TestMulOp(OpTest): def setUp(self): self.op_type = "mul" self.inputs = { - 'X': np.random.random((32, 84)).astype("float32"), - 'Y': np.random.random((84, 100)).astype("float32") + 'X': np.random.random((2, 5)).astype("float32"), + 'Y': np.random.random((5, 3)).astype("float32") } self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} @@ -46,13 +46,16 @@ class TestMulOp2(OpTest): def setUp(self): self.op_type = "mul" self.inputs = { - 'X': np.random.random((15, 4, 12, 10)).astype("float32"), - 'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32") + 'X': np.random.random((3, 4, 4, 3)).astype("float32"), + 'Y': np.random.random((2, 6, 1, 2, 3)).astype("float32") } - self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} - result = np.dot(self.inputs['X'].reshape(15 * 4, 12 * 10), - self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9)) - result = result.reshape(15, 4, 8, 2, 9) + self.attrs = { + 'x_num_col_dims': 2, + 'y_num_col_dims': 2, + } + result = np.dot(self.inputs['X'].reshape(3 * 4, 4 * 3), + self.inputs['Y'].reshape(2 * 6, 1 * 2 * 3)) + result = result.reshape(3, 4, 1, 2, 3) self.outputs = {'Out': result} def test_check_output(self): @@ -73,9 +76,9 @@ class TestMulOp2(OpTest): class TestFP16MulOp1(OpTest): def setUp(self): self.op_type = "mul" - x = np.random.random((32, 84)).astype("float16") - y = np.random.random((84, 100)).astype("float16") - self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} + x = np.random.random((3, 5)).astype("float16") + y = np.random.random((5, 4)).astype("float16") + self.inputs = {'X': x.view(np.float16), 'Y': y.view(np.float16)} self.outputs = {'Out': np.dot(x, y)} def test_check_output(self): @@ -88,13 +91,15 @@ class TestFP16MulOp1(OpTest): class TestFP16MulOp2(OpTest): def setUp(self): self.op_type = "mul" - x = np.random.random((15, 4, 12, 10)).astype("float16") - y = np.random.random((4, 30, 8, 2, 9)).astype("float16") - self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} - self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} - result = np.dot( - x.reshape(15 * 4, 12 * 10), y.reshape(4 * 30, 8 * 2 * 9)) - result = result.reshape(15, 4, 8, 2, 9) + x = np.random.random((3, 4, 4, 3)).astype("float16") + y = np.random.random((2, 6, 1, 2, 3)).astype("float16") + self.inputs = {'X': x.view(np.float16), 'Y': y.view(np.float16)} + self.attrs = { + 'x_num_col_dims': 2, + 'y_num_col_dims': 2, + } + result = np.dot(x.reshape(3 * 4, 4 * 3), y.reshape(2 * 6, 1 * 2 * 3)) + result = result.reshape(3, 4, 1, 2, 3) self.outputs = {'Out': result} def test_check_output(self): diff --git a/python/paddle/fluid/tests/unittests/test_norm_op.py b/python/paddle/fluid/tests/unittests/test_norm_op.py index 6feda175fb..108a665f37 100644 --- a/python/paddle/fluid/tests/unittests/test_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_norm_op.py @@ -17,44 +17,23 @@ import numpy as np from op_test import OpTest -def norm(input, scale, epsilon): - s0, s1, s2, s3 = input.shape - x_square = input * input - for i in xrange(s0): - input_batch = input[i:i + 1, :, :, :] - input_batch = input_batch.reshape(s1, s2 * s3) - x_square_batch = x_square[i:i + 1, :, :, :] - x_square_batch = x_square_batch.reshape(s1, s2 * s3) - square_colsum = x_square_batch.sum(axis=0) + epsilon - tmp = pow(square_colsum, 0.5) - tmp = np.reciprocal(tmp) - tmp_tile = np.tile(tmp, s1) - tmp_tile = tmp_tile.reshape(s1, s2 * s3) - scale_tile = np.tile(scale, (1, s2 * s3)) - scale_tile = scale_tile.reshape(s1, s2 * s3) - out_batch = input_batch * tmp_tile * scale_tile - out_batch = out_batch.reshape(1, s1, s2, s3) - if i == 0: - out = out_batch - else: - out = np.concatenate((out, out_batch), 0) - out.reshape(s0, s1, s2, s3) - return out +def l2_norm(x, axis, epsilon): + x2 = x**2 + s = np.sum(x2, axis=axis, keepdims=True) + r = np.sqrt(s + epsilon) + y = x / np.broadcast_to(r, x.shape) + return y, r class TestNormOp(OpTest): def setUp(self): self.op_type = "norm" self.init_test_case() - input = np.random.random(self.shape).astype("float32") - scale = np.array([10, 10, 10]) - self.inputs = { - 'X': input.astype('float32'), - 'Scale': scale.astype('float32') - } - self.attrs = {'epsilon': self.epsilon} - output = norm(input, scale, self.epsilon) - self.outputs = {'Out': output.astype('float32')} + x = np.random.random(self.shape).astype("float64") + y, norm = l2_norm(x, self.axis, self.epsilon) + self.inputs = {'X': x} + self.attrs = {'epsilon': self.epsilon, 'axis': self.axis} + self.outputs = {'Out': y, 'Norm': norm} def test_check_output(self): self.check_output() @@ -63,8 +42,23 @@ class TestNormOp(OpTest): self.check_grad(['X'], 'Out') def init_test_case(self): - self.shape = [2, 3, 2, 2] - self.epsilon = 1e-6 + self.shape = [2, 3, 4, 4] + self.axis = 1 + self.epsilon = 1e-8 + + +class TestNormOp2(TestNormOp): + def init_test_case(self): + self.shape = [5, 3, 9, 7] + self.axis = 0 + self.epsilon = 1e-8 + + +class TestNormOp3(TestNormOp): + def init_test_case(self): + self.shape = [5, 3, 2, 7] + self.axis = -1 + self.epsilon = 1e-8 if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py b/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py index ef34893943..198c68866d 100644 --- a/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py +++ b/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py @@ -70,8 +70,9 @@ class TestNormalization(unittest.TestCase): def l2_normalize(self, data, axis, epsilon): """ Compute the groundtruth. """ - output = data * np.reciprocal( - np.sum(np.square(data), axis=axis, keepdims=True)) + output = data / np.broadcast_to( + np.sqrt(np.sum(np.square(data), axis=axis, keepdims=True)), + data.shape) return output def test_l2_normalize(self): diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py index 66e138b03f..163975555e 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py @@ -17,6 +17,7 @@ import paddle.fluid as fluid import unittest import paddle import numpy as np +import os word_dict, verb_dict, label_dict = conll05.get_dict() word_dict_len = len(word_dict) @@ -101,7 +102,11 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, class TestCRFModel(unittest.TestCase): - def check_network_convergence(self, is_sparse, build_strategy=None): + def check_network_convergence(self, + is_sparse, + build_strategy=None, + use_cuda=True): + os.environ['CPU_NUM'] = str(4) main = fluid.Program() startup = fluid.Program() with fluid.program_guard(main, startup): @@ -145,12 +150,12 @@ class TestCRFModel(unittest.TestCase): paddle.dataset.conll05.test(), buf_size=8192), batch_size=16) - place = fluid.CUDAPlace(0) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup) pe = fluid.ParallelExecutor( - use_cuda=True, + use_cuda=use_cuda, loss_name=avg_cost.name, build_strategy=build_strategy) @@ -172,25 +177,33 @@ class TestCRFModel(unittest.TestCase): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce self.check_network_convergence( - is_sparse=True, build_strategy=build_strategy) + is_sparse=True, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy, use_cuda=False) def test_update_dense_parameter_all_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce self.check_network_convergence( - is_sparse=False, build_strategy=build_strategy) + is_sparse=False, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy, use_cuda=False) def test_update_sparse_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce self.check_network_convergence( - is_sparse=True, build_strategy=build_strategy) + is_sparse=True, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy, use_cuda=False) def test_update_dense_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce self.check_network_convergence( - is_sparse=False, build_strategy=build_strategy) + is_sparse=False, build_strategy=build_strategy, use_cuda=True) + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy, use_cuda=False) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py index 24f8d28c03..79702475cc 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py @@ -18,6 +18,7 @@ import paddle.fluid as fluid import unittest import numpy as np import paddle +import os def Lenet(data, class_dim): @@ -35,7 +36,7 @@ def Lenet(data, class_dim): class TestFetchOp(unittest.TestCase): - def parallel_exe(self, train_inputs, seed): + def parallel_exe(self, train_inputs, seed, use_cuda): main = fluid.Program() startup = fluid.Program() startup.random_seed = seed @@ -59,13 +60,13 @@ class TestFetchOp(unittest.TestCase): # conv2d_1.b_0@GRAD. Those variables should not be pruned. # fluid.memory_optimize(main) - place = fluid.CUDAPlace(0) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup) feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) pe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name, main_program=main) + use_cuda=use_cuda, loss_name=loss.name, main_program=main) fetch_list = [] all_vars = main.global_block().vars @@ -88,14 +89,16 @@ class TestFetchOp(unittest.TestCase): for i in range(iters): train_inputs.append(tst_reader_iter.next()) - self.parallel_exe(train_inputs, seed=1) + os.environ['CPU_NUM'] = str(4) + self.parallel_exe(train_inputs, seed=1, use_cuda=True) + self.parallel_exe(train_inputs, seed=1, use_cuda=False) class TestFeedParallel(unittest.TestCase): - def test_main(self): + def parallel_exe(self, use_cuda, seed): main = fluid.Program() startup = fluid.Program() - startup.random_seed = 1 + startup.random_seed = seed with fluid.scope_guard(fluid.core.Scope()): with fluid.program_guard(main, startup): data = fluid.layers.data( @@ -111,15 +114,18 @@ class TestFeedParallel(unittest.TestCase): regularization=fluid.regularizer.L2Decay(1e-4)) opt.minimize(loss) - place = fluid.CUDAPlace(0) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() feeder = fluid.DataFeeder(place=place, feed_list=[data, label]) reader = feeder.decorate_reader( paddle.batch( flowers.train(), batch_size=16), multi_devices=True) + exe = fluid.Executor(place) exe.run(startup) + pe = fluid.ParallelExecutor( - use_cuda=True, loss_name=loss.name, main_program=main) + use_cuda=use_cuda, loss_name=loss.name, main_program=main) for batch_id, data in enumerate(reader()): loss_np = np.array(pe.run(feed=data, fetch_list=[loss.name])[0]) @@ -127,6 +133,11 @@ class TestFeedParallel(unittest.TestCase): if batch_id == 2: break + def test_feed_op(self): + os.environ['CPU_NUM'] = str(4) + self.parallel_exe(use_cuda=True, seed=1) + self.parallel_exe(use_cuda=False, seed=1) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py index 015703c3e2..a801d99aa1 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py @@ -18,6 +18,7 @@ import numpy as np import paddle import paddle.dataset.mnist as mnist import unittest +import os MNIST_RECORDIO_FILE = "./mnist_test_pe.recordio" @@ -85,6 +86,7 @@ def fc_with_batchnorm(use_feed): class TestMNIST(TestParallelExecutorBase): @classmethod def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) # Convert mnist to recordio file with fluid.program_guard(fluid.Program(), fluid.Program()): reader = paddle.batch(mnist.train(), batch_size=4) @@ -99,9 +101,12 @@ class TestMNIST(TestParallelExecutorBase): fluid.recordio_writer.convert_reader_to_recordio_file( MNIST_RECORDIO_FILE, reader, feeder) - def check_simple_fc_convergence(self, balance_parameter_opt_between_cards): - self.check_network_convergence(simple_fc_net) - self.check_network_convergence(simple_fc_net, allow_op_delay=True) + def check_simple_fc_convergence(self, + balance_parameter_opt_between_cards, + use_cuda=True): + self.check_network_convergence(simple_fc_net, use_cuda=use_cuda) + self.check_network_convergence( + simple_fc_net, use_cuda=use_cuda, allow_op_delay=True) img = np.zeros(shape=[32, 784], dtype='float32') label = np.ones(shape=[32, 1], dtype='int64') @@ -109,17 +114,21 @@ class TestMNIST(TestParallelExecutorBase): simple_fc_net, feed_dict={"image": img, "label": label}, + use_cuda=use_cuda, balance_parameter_opt_between_cards=balance_parameter_opt_between_cards ) def test_simple_fc(self): - self.check_simple_fc_convergence(False) + self.check_simple_fc_convergence(False, use_cuda=True) + self.check_simple_fc_convergence(False, use_cuda=False) def test_simple_fc_with_new_strategy(self): - self.check_simple_fc_convergence(True) + self.check_simple_fc_convergence(True, use_cuda=True) + self.check_simple_fc_convergence(True, use_cuda=False) def check_simple_fc_parallel_accuracy(self, - balance_parameter_opt_between_cards): + balance_parameter_opt_between_cards, + use_cuda=True): img = np.zeros(shape=[32, 784], dtype='float32') label = np.ones(shape=[32, 1], dtype='int64') single_first_loss, single_last_loss = self.check_network_convergence( @@ -127,12 +136,14 @@ class TestMNIST(TestParallelExecutorBase): seed=1000, feed_dict={"image": img, "label": label}, + use_cuda=use_cuda, use_parallel_executor=False) parallel_first_loss, parallel_last_loss = self.check_network_convergence( method=simple_fc_net, seed=1000, feed_dict={"image": img, "label": label}, + use_cuda=use_cuda, use_parallel_executor=True, balance_parameter_opt_between_cards=balance_parameter_opt_between_cards ) @@ -143,28 +154,33 @@ class TestMNIST(TestParallelExecutorBase): self.assertAlmostEquals(p_l, single_last_loss[0], delta=1e-6) def test_simple_fc_parallel_accuracy(self): - self.check_simple_fc_parallel_accuracy(False) + self.check_simple_fc_parallel_accuracy(False, use_cuda=True) + self.check_simple_fc_parallel_accuracy(False, use_cuda=False) def test_simple_fc_parallel_accuracy_with_new_strategy(self): - self.check_simple_fc_parallel_accuracy(True) + self.check_simple_fc_parallel_accuracy(True, use_cuda=True) + self.check_simple_fc_parallel_accuracy(True, use_cuda=False) - def check_batchnorm_fc_convergence(self, - balance_parameter_opt_between_cards): - self.check_network_convergence(fc_with_batchnorm) + def check_batchnorm_fc_convergence( + self, balance_parameter_opt_between_cards, use_cuda): + self.check_network_convergence(fc_with_batchnorm, use_cuda=use_cuda) img = np.zeros(shape=[32, 784], dtype='float32') label = np.ones(shape=[32, 1], dtype='int64') self.check_network_convergence( fc_with_batchnorm, feed_dict={"image": img, "label": label}, + use_cuda=use_cuda, balance_parameter_opt_between_cards=balance_parameter_opt_between_cards ) def test_batchnorm_fc(self): - self.check_batchnorm_fc_convergence(False) + self.check_batchnorm_fc_convergence(False, use_cuda=True) + self.check_batchnorm_fc_convergence(False, use_cuda=False) def test_batchnorm_fc_with_new_strategy(self): - self.check_batchnorm_fc_convergence(True) + self.check_batchnorm_fc_convergence(True, use_cuda=True) + self.check_batchnorm_fc_convergence(True, use_cuda=False) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py index a3fa140cbb..066299e6c6 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py @@ -15,6 +15,7 @@ import paddle.fluid as fluid from parallel_executor_test_base import TestParallelExecutorBase import unittest +import os def squeeze_excitation(input, num_channels, reduction_ratio): @@ -130,22 +131,30 @@ def SE_ResNeXt50Small(batch_size=2, use_feed=False): class TestResnet(TestParallelExecutorBase): - def check_resnet_convergence(self, balance_parameter_opt_between_cards): + def check_resnet_convergence(self, + balance_parameter_opt_between_cards, + use_cuda=True, + iter=20): + os.environ['CPU_NUM'] = str(4) + import functools batch_size = 2 self.check_network_convergence( functools.partial( SE_ResNeXt50Small, batch_size=batch_size), - iter=20, + iter=iter, batch_size=batch_size, + use_cuda=use_cuda, balance_parameter_opt_between_cards=balance_parameter_opt_between_cards ) def test_resnet(self): - self.check_resnet_convergence(False) + self.check_resnet_convergence(False, use_cuda=True) + self.check_resnet_convergence(False, use_cuda=False, iter=5) def test_resnet_with_new_strategy(self): - self.check_resnet_convergence(True) + self.check_resnet_convergence(True, use_cuda=True) + self.check_resnet_convergence(True, use_cuda=False, iter=5) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py index 93a5f76786..31ba8c1d60 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py @@ -15,6 +15,7 @@ import paddle.fluid as fluid import numpy as np import unittest +import os def simple_fc_net(): @@ -35,7 +36,8 @@ def simple_fc_net(): class ParallelExecutorTestingDuringTraining(unittest.TestCase): - def check_network_convergence(self, build_strategy=None): + def check_network_convergence(self, use_cuda, build_strategy=None): + os.environ['CPU_NUM'] = str(4) main = fluid.Program() startup = fluid.Program() with fluid.program_guard(main, startup): @@ -49,19 +51,19 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): image = np.random.normal(size=(batch_size, 784)).astype('float32') label = np.random.randint(0, 10, (batch_size, 1), dtype="int64") - place = fluid.CUDAPlace(0) + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() exe = fluid.Executor(place) exe.run(startup) feed_dict = {'image': image, 'label': label} train_exe = fluid.ParallelExecutor( - use_cuda=True, + use_cuda=use_cuda, loss_name=loss.name, main_program=main, build_strategy=build_strategy) test_exe = fluid.ParallelExecutor( - use_cuda=True, + use_cuda=use_cuda, main_program=test_program, share_vars_from=train_exe, build_strategy=build_strategy) @@ -81,12 +83,18 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase): def test_parallel_testing(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce - self.check_network_convergence(build_strategy) + self.check_network_convergence( + use_cuda=True, build_strategy=build_strategy) + self.check_network_convergence( + use_cuda=False, build_strategy=build_strategy) def test_parallel_testing_with_new_strategy(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce - self.check_network_convergence(build_strategy) + self.check_network_convergence( + use_cuda=True, build_strategy=build_strategy) + self.check_network_convergence( + use_cuda=False, build_strategy=build_strategy) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py index c81df66d98..b6215fddb1 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py @@ -19,6 +19,7 @@ from parallel_executor_test_base import TestParallelExecutorBase import unittest import paddle import paddle.dataset.wmt16 as wmt16 +import os WMT16_RECORDIO_FILE = "./wmt16_test_pe.recordio" @@ -149,6 +150,7 @@ def transformer(use_feed): class TestTransformer(TestParallelExecutorBase): @classmethod def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) reader = paddle.batch( wmt16.train(ModelHyperParams.src_vocab_size, ModelHyperParams.trg_vocab_size), @@ -167,7 +169,8 @@ class TestTransformer(TestParallelExecutorBase): @unittest.skip("transformer is buggy in multi gpu") def test_main(self): - self.check_network_convergence(transformer) + self.check_network_convergence(transformer, use_cuda=True) + self.check_network_convergence(transformer, use_cuda=False) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_reverse_op.py b/python/paddle/fluid/tests/unittests/test_reverse_op.py new file mode 100644 index 0000000000..f845575a02 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_reverse_op.py @@ -0,0 +1,67 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestReverseOp(OpTest): + def initTestCase(self): + self.x = np.random.random((3, 4)).astype('float32') + self.axis = [0] + + def setUp(self): + self.initTestCase() + self.op_type = "reverse" + self.inputs = {"X": self.x} + self.attrs = {'axis': self.axis} + out = self.x + for a in self.axis: + out = np.flip(out, axis=a) + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + +class TestCase0(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4)).astype('float32') + self.axis = [1] + + +class TestCase1(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4)).astype('float32') + self.axis = [0, 1] + + +class TestCase2(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4, 5)).astype('float32') + self.axis = [0, 2] + + +class TestCase3(TestReverseOp): + def initTestCase(self): + self.x = np.random.random((3, 4, 5)).astype('float32') + self.axis = [1, 2] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_shape_op.py b/python/paddle/fluid/tests/unittests/test_shape_op.py new file mode 100644 index 0000000000..a62ee05007 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_shape_op.py @@ -0,0 +1,47 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestShapeOp(OpTest): + def setUp(self): + self.op_type = "shape" + self.config() + self.shape = [2, 3] + input = np.zeros(self.shape) + self.inputs = {'Input': input} + self.outputs = {'Out': np.array(self.shape)} + + def config(self): + self.shape = [2, 3] + + def test_check_output(self): + self.check_output() + + +class case1(TestShapeOp): + def config(self): + self.shape = [2] + + +class case2(TestShapeOp): + def config(self): + self.shape = [1, 2, 3] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_simple_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_simple_dist_transpiler.py new file mode 100644 index 0000000000..f4aa7426bc --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_simple_dist_transpiler.py @@ -0,0 +1,80 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np + +import paddle.fluid as fluid +from paddle.fluid.transpiler.distribute_transpiler import delete_ops + +from transpiler_test import TranspilerTest + + +class TestSimpleDistTranspiler(TranspilerTest): + def setUp(self): + self.current_pserver_ep = "127.0.0.1:6175" + + def test_simple_transpiler(self): + np.random.seed(1) + + trainer = self.get_trainer() + pserver, startup = self.get_pserver(self.current_pserver_ep) + self.assertEqual([op.type for op in trainer.global_block().ops], + self.get_expect_trainer_ops()) + + self.assertEqual(len(pserver.blocks), 2) + # block0: listen_and_serv + self.assertEqual([op.type for op in pserver.blocks[0].ops], + ["listen_and_serv"]) + # block1: optimize pass + self.assertEqual([op.type for op in pserver.blocks[1].ops], + ["sum", "scale", "sgd"]) + + # confirm startup program + self.assertEqual([op.type for op in startup.global_block().ops], + ["fill_constant", "uniform_random", "uniform_random"]) + + # the variable #fc_w will NOT be splited + fc_w_var = startup.global_block().var("fc_w@GRAD") + self.assertEqual(fc_w_var.shape, (1000, 1000)) + + fc_w_var = startup.global_block().var("fc_w@GRAD.trainer_0") + self.assertEqual(fc_w_var.shape, (1000, 1000)) + + def get_expect_trainer_ops(self): + trainer = fluid.Program() + + with fluid.program_guard(trainer): + optimize_ops, params_grads = self.net_conf() + + delete_ops(trainer.global_block(), optimize_ops) + ops = [op.type for op in trainer.global_block().ops] + [ + "send", "send_barrier", "recv", "recv", "fetch_barrier" + ] + ops.insert(ops.index("elementwise_add_grad") + 1, "send") + return ops + + def _transpiler_instance(self): + main = self.get_main_program() + t = fluid.DistributeTranspiler() + t.transpile( + self.trainer_id, + program=main, + pservers=self.pserver_eps, + trainers=self.trainers, + slice_var_up=False) + return t + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_slice_op.py b/python/paddle/fluid/tests/unittests/test_slice_op.py new file mode 100644 index 0000000000..1a48bce3bb --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_slice_op.py @@ -0,0 +1,62 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestSliceOp(OpTest): + def setUp(self): + self.op_type = "slice" + self.config() + self.inputs = {'Input': self.input} + self.outputs = {'Out': self.out} + self.attrs = { + 'axes': self.axes, + 'starts': self.starts, + 'ends': self.ends + } + + def config(self): + self.input = np.random.random([3, 4, 5, 6]).astype("float32") + self.starts = [1, 0, 2] + self.ends = [3, 3, 4] + self.axes = [0, 1, 2] + self.out = self.input[1:3, 0:3, 2:4, :] + + def test_check_output(self): + self.check_output() + + +class TestCase1(TestSliceOp): + def config(self): + self.input = np.random.random([3, 4, 5, 6]).astype("float32") + self.starts = [-3, 0, 2] + self.ends = [3, 100, -1] + self.axes = [0, 1, 2] + self.out = self.input[-3:3, 0:100, 2:-1, :] + + +class TestCase2(TestSliceOp): + def config(self): + self.input = np.random.random([3, 4, 5, 6]).astype("float32") + self.starts = [-3, 0, 2] + self.ends = [3, 100, -1] + self.axes = [0, 1, 3] + self.out = self.input[-3:3, 0:100, :, 2:-1] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_split_var.py b/python/paddle/fluid/tests/unittests/test_slice_var.py similarity index 84% rename from python/paddle/fluid/tests/unittests/test_split_var.py rename to python/paddle/fluid/tests/unittests/test_slice_var.py index 0c5e8901b9..82305b23a1 100644 --- a/python/paddle/fluid/tests/unittests/test_split_var.py +++ b/python/paddle/fluid/tests/unittests/test_slice_var.py @@ -14,14 +14,14 @@ import math import unittest -from paddle.fluid.transpiler.distribute_transpiler import split_dense_variable +from paddle.fluid.transpiler.distribute_transpiler import slice_variable import paddle.fluid as fluid import paddle.fluid.core as core import random -class TestSplitVar(unittest.TestCase): - def check_split_output(self, shapes, expected_sizes, min_size): +class TestSliceVar(unittest.TestCase): + def check_slice_output(self, shapes, expected_sizes, min_size): var_list = [] program = fluid.Program() for shape in shapes: @@ -31,7 +31,7 @@ class TestSplitVar(unittest.TestCase): # dtype=core.VarDesc.VarType.LOD_TENSOR, shape=shape) var_list.append(var) - blocks = split_dense_variable(var_list, 10, min_size) + blocks = slice_variable(var_list, 10, min_size) all_sizes = [] for s in expected_sizes: for s2 in s: @@ -49,7 +49,7 @@ class TestSplitVar(unittest.TestCase): [1150, 1150, 1150, 1150, 1150, 1150, 1100] ] - self.check_split_output(shapes, expected_sizes, 1024) + self.check_slice_output(shapes, expected_sizes, 1024) def test_check_output_8k(self): shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10], @@ -57,7 +57,7 @@ class TestSplitVar(unittest.TestCase): expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000], [35937, 35937, 35937, 35937, 35937, 35937]] - self.check_split_output(shapes, expected_sizes, 8192) + self.check_slice_output(shapes, expected_sizes, 8192) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/testsuite.py b/python/paddle/fluid/tests/unittests/testsuite.py new file mode 100644 index 0000000000..1dc94a80c9 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/testsuite.py @@ -0,0 +1,182 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np + +import paddle.fluid.core as core +from paddle.fluid.op import Operator + + +def as_lodtensor(np_array, lod, place): + tensor = core.LoDTensor() + tensor.set(np_value, place) + if lod is not None: + tensor.set_lod(lod) + return tensor + + +def create_op(scope, op_type, inputs, outputs, attrs): + kwargs = dict() + + op_maker = core.op_proto_and_checker_maker + op_role_attr_name = op_maker.kOpRoleAttrName() + + if op_role_attr_name not in attrs: + attrs[op_role_attr_name] = int(op_maker.OpRole.Forward) + + def __create_var__(name, var_name): + scope.var(var_name).get_tensor() + kwargs[name].append(var_name) + + for in_name, in_dup in Operator.get_op_inputs(op_type): + if in_name in inputs: + kwargs[in_name] = [] + if in_dup: + sub_in = inputs[in_name] + for item in sub_in: + sub_in_name, _ = item[0], item[1] + __create_var__(in_name, sub_in_name) + else: + __create_var__(in_name, in_name) + + for out_name, out_dup in Operator.get_op_outputs(op_type): + if out_name in outputs: + kwargs[out_name] = [] + if out_dup: + sub_out = outputs[out_name] + for item in sub_out: + sub_out_name, _ = item[0], item[1] + __create_var__(out_name, sub_out_name) + else: + __create_var__(out_name, out_name) + + for attr_name in Operator.get_op_attr_names(op_type): + if attr_name in attrs: + kwargs[attr_name] = attrs[attr_name] + + return Operator(op_type, **kwargs) + + +def set_input(scope, op, inputs, place): + def __set_input__(var_name, var): + if isinstance(var, tuple) or isinstance(var, np.ndarray): + tensor = scope.find_var(var_name).get_tensor() + if isinstance(var, tuple): + tensor.set_lod(var[1]) + var = var[0] + tensor.set_dims(var.shape) + tensor.set(var, place) + elif isinstance(var, float): + scope.find_var(var_name).set_float(var) + elif isinstance(var, int): + scope.find_var(var_name).set_int(var) + + for in_name, in_dup in Operator.get_op_inputs(op.type()): + if in_name in inputs: + if in_dup: + sub_in = inputs[in_name] + for item in sub_in: + sub_in_name, sub_in_val = item[0], item[1] + __set_input__(sub_in_name, sub_in_val) + else: + __set_input__(in_name, inputs[in_name]) + + +def append_input_output(block, op_proto, np_list, is_input, dtype): + '''Insert VarDesc and generate Python variable instance''' + proto_list = op_proto.inputs if is_input else op_proto.outputs + + def create_var(block, name, np_list, var_proto): + dtype = None + shape = None + lod_level = None + if name not in np_list: + assert var_proto.intermediate, "{} not found".format(name) + else: + np_value = np_list[name] + if isinstance(np_value, tuple): + dtype = np_value[0].dtype + # output shape, lod should be infered from input. + if is_input: + shape = list(np_value[0].shape) + lod_level = len(np_value[1]) + else: + dtype = np_value.dtype + if is_input: + shape = list(np_value.shape) + lod_level = 0 + return block.create_var( + dtype=dtype, shape=shape, lod_level=lod_level, name=name) + + var_dict = {} + for var_proto in proto_list: + var_name = str(var_proto.name) + if is_input: + if (var_name not in np_list) and var_proto.dispensable: + continue + assert (var_name in np_list) or (var_proto.dispensable), \ + "Missing {} as input".format(var_name) + if var_proto.duplicable: + assert isinstance(np_list[var_name], list), \ + "Duplicable {} should be set as list".format(var_name) + var_list = [] + for (name, np_value) in np_list[var_name]: + var_list.append( + create_var(block, name, {name: np_value}, var_proto)) + var_dict[var_name] = var_list + else: + var_dict[var_name] = create_var(block, var_name, np_list, var_proto) + + return var_dict + + +def append_loss_ops(block, output_names): + mean_inputs = map(block.var, output_names) + # for item in mean_inputs: + # print(item) + # print("Item", item.dtype) + + if len(mean_inputs) == 1: + loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1]) + op = block.append_op( + inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean') + op.desc.infer_var_type(block.desc) + op.desc.infer_shape(block.desc) + else: + avg_sum = [] + for cur_loss in mean_inputs: + cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1]) + op = block.append_op( + inputs={"X": [cur_loss]}, + outputs={"Out": [cur_avg_loss]}, + type="mean") + op.desc.infer_var_type(block.desc) + op.desc.infer_shape(block.desc) + avg_sum.append(cur_avg_loss) + + loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1]) + op_sum = block.append_op( + inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum') + op_sum.desc.infer_var_type(block.desc) + op_sum.desc.infer_shape(block.desc) + + loss = block.create_var(dtype=loss_sum.dtype, shape=[1]) + op_loss = block.append_op( + inputs={"X": loss_sum}, + outputs={"Out": loss}, + type='scale', + attrs={'scale': 1.0 / float(len(avg_sum))}) + op_loss.desc.infer_var_type(block.desc) + op_loss.desc.infer_shape(block.desc) + return loss diff --git a/python/paddle/fluid/tests/unittests/transpiler_test.py b/python/paddle/fluid/tests/unittests/transpiler_test.py new file mode 100644 index 0000000000..d84c5d9c41 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/transpiler_test.py @@ -0,0 +1,73 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np + +import paddle.fluid as fluid +import paddle.fluid.core as core +import paddle.fluid.layers as layers + + +class TranspilerTest(unittest.TestCase): + @classmethod + def setUpClass(self): + self.trainer_id = 0 + self.trainers = 2 + self.pservers = 2 + self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175" + + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w')) + + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) + + optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) + return optimize_ops, params_grads + + def get_main_program(self): + main = fluid.Program() + + with fluid.program_guard(main): + self.net_conf() + + return main + + def get_trainer(self): + return self._transpiler_instance().get_trainer_program() + + def get_pserver(self, ep): + t = self._transpiler_instance() + pserver = t.get_pserver_program(ep) + startup = t.get_startup_program(ep, pserver) + return pserver, startup + + def _transpiler_instance(self): + main = self.get_main_program() + t = fluid.DistributeTranspiler() + t.transpile( + self.trainer_id, + program=main, + pservers=self.pserver_eps, + trainers=self.trainers) + return t diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index 7da123dd92..efc28d8993 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -27,11 +27,8 @@ import parallel_executor from transpiler import distribute_transpiler __all__ = [ - 'Trainer', - 'BeginEpochEvent', - 'EndEpochEvent', - 'BeginStepEvent', - 'EndStepEvent', + 'Trainer', 'BeginEpochEvent', 'EndEpochEvent', 'BeginStepEvent', + 'EndStepEvent', 'CheckpointConfig' ] @@ -59,6 +56,35 @@ class EndStepEvent(object): self.metrics = metrics +class CheckpointConfig(object): + def __init__(self, + checkpoint_dir=None, + max_num_checkpoints=3, + epoch_interval=1, + step_interval=10): + if checkpoint_dir is None: + self.checkpoint_dir = os.getcwd() + else: + self.checkpoint_dir = checkpoint_dir + + self.max_num_checkpoints = max_num_checkpoints + + if epoch_interval < 1: + self.epoch_interval = 1 + else: + self.epoch_interval = epoch_interval + + if step_interval < 1: + self.step_interval = 10 + else: + self.step_interval = step_interval + + self.epoch_id = 0 + self.step_id = 0 + self.load_serial = None + self.is_pserver = False + + def check_and_get_place(place): """ Check the type of place or get the default place @@ -90,23 +116,32 @@ class Trainer(object): Args: train_func(callable): A function which will return loss. The loss must be a scalar. - optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer + optimizer_func(callable): A function that returns an Optimizer object. place: The device place of this trainer. """ def __init__(self, train_func, - optimizer, + optimizer_func, param_path=None, place=None, - parallel=False): + parallel=False, + checkpoint_config=None): self.__stop = False self.parallel = parallel # 1. we need to generate a framework.Program by calling # program_func. Reference: fluid.program_guard in # test_word2vec.py - if not isinstance(optimizer, opt_module.Optimizer): - raise TypeError("The optimizer should be an instance of Optimizer") + + # config for checkpoint + # only chief worker will save variables + self.trainer_id = 0 + self.checkpoint_cfg = checkpoint_config + if self.checkpoint_cfg: + assert isinstance(self.checkpoint_cfg, CheckpointConfig) + serial = io.get_latest_checkpoint_serial( + self.checkpoint_cfg.checkpoint_dir) + self.checkpoint_cfg.load_serial = serial if serial >= 0 else None self.scope = core.Scope() @@ -117,12 +152,15 @@ class Trainer(object): program_func_outs = train_func() self.train_func_outputs = program_func_outs if isinstance( program_func_outs, list) else [program_func_outs] - self.test_program = self.train_program.clone() + self.test_program = self.train_program.clone(for_test=True) + + # The first element of program_func_outs is loss. + loss = self.train_func_outputs[0] + + optimizer = optimizer_func() if not isinstance(optimizer, opt_module.Optimizer): raise TypeError( "The optimizer should be an instance of Optimizer") - # The fisrt element of program_func_outs is loss. - loss = self.train_func_outputs[0] optimize_ops, params_grads = optimizer.minimize(loss) self.place = check_and_get_place(place) @@ -136,9 +174,25 @@ class Trainer(object): exe = executor.Executor(place) exe.run(self.startup_program) - if param_path: + if self.checkpoint_cfg and self.checkpoint_cfg.load_serial: + with self._prog_and_scope_guard(): + exe = executor.Executor(place) + io.load_checkpoint(exe, self.checkpoint_cfg.checkpoint_dir, + self.checkpoint_cfg.load_serial, + self.startup_program) + + if not self.checkpoint_cfg.is_pserver: + epoch_id, step_id = io.load_trainer_args( + self.checkpoint_cfg.checkpoint_dir, + self.checkpoint_cfg.load_serial, self.trainer_id, + self._get_checkpoint_load_args()) + self.checkpoint_cfg.epoch_id = int(epoch_id) + self.checkpoint_cfg.step_id = int(step_id) + + if param_path and os.path.isdir(param_path): # load params from param_path into scope - io.load_persistables(exe, dirname=param_path) + io.load_persist_vars_without_grad( + exe, dirname=param_path, program=self.startup_program) def _transpile_nccl2_dist(self): # PADDLE_TRAINER_IPS @@ -193,14 +247,18 @@ class Trainer(object): current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port # the unique trainer id, starting from 0, needed by trainer # only - trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) + self.trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0")) + # the role, should be either PSERVER or TRAINER training_role = os.getenv("PADDLE_TRAINING_ROLE") with self._prog_and_scope_guard(): t = distribute_transpiler.DistributeTranspiler() t.transpile( - trainer_id, pservers=pserver_endpoints, trainers=trainers) + self.trainer_id, pservers=pserver_endpoints, trainers=trainers) if training_role == "PSERVER": + if self.checkpoint_cfg: + self.is_pserver = True + self.train_program = t.get_pserver_program(current_endpoint) self.startup_program = t.get_startup_program(current_endpoint, self.train_program) @@ -293,11 +351,26 @@ class Trainer(object): self._train_by_any_executor(event_handler, exe, num_epochs, reader) def _train_by_any_executor(self, event_handler, exe, num_epochs, reader): - for epoch_id in range(num_epochs): + if self.checkpoint_cfg: + epochs = [ + epoch_id for epoch_id in range(num_epochs) + if epoch_id >= self.checkpoint_cfg.epoch_id + ] + else: + epochs = [epoch_id for epoch_id in range(num_epochs)] + + for epoch_id in epochs: event_handler(BeginEpochEvent(epoch_id)) for step_id, data in enumerate(reader()): if self.__stop: + if self.checkpoint_cfg: + self._clean_checkpoint() return + + if self.checkpoint_cfg and self.checkpoint_cfg.load_serial \ + and self.checkpoint_cfg.step_id >= step_id and self.checkpoint_cfg.epoch_id == epoch_id: + continue + begin_event = BeginStepEvent(epoch_id, step_id) event_handler(begin_event) if begin_event.fetch_metrics: @@ -308,8 +381,13 @@ class Trainer(object): ]) else: metrics = exe.run(feed=data, fetch_list=[]) + + if self.checkpoint_cfg: + self._save_checkpoint(epoch_id, step_id) event_handler(EndStepEvent(epoch_id, step_id, metrics)) event_handler(EndEpochEvent(epoch_id)) + if self.checkpoint_cfg: + self._clean_checkpoint() def _test_by_executor(self, reader, feed_order, fetch_list): with executor.scope_guard(self.scope): @@ -348,6 +426,38 @@ class Trainer(object): loss_name=self.train_func_outputs[0].name) return self._get_parallel_executor() + def _clean_checkpoint(self): + assert self.checkpoint_cfg + io.clean_checkpoint(checkpoint_dir=self.checkpoint_cfg.checkpoint_dir) + + def _get_checkpoint_load_args(self): + """ + epoch_id and step_id are runtime arguments, they are not variables, will load them independently. + """ + return ["epoch_id", "step_id"] + + def _get_checkpoint_save_args(self, epoch_id, step_id): + """ + epoch_id and step_id are runtime arguments, they are not variables, will save them independently. + """ + trainer_args = {} + trainer_args["epoch_id"] = epoch_id + trainer_args["step_id"] = step_id + return trainer_args + + def _save_checkpoint(self, epoch_id, step_id): + assert self.checkpoint_cfg + + if epoch_id % self.checkpoint_cfg.epoch_interval == 0 and step_id % self.checkpoint_cfg.step_interval == 0: + exe = executor.Executor(self.place) + io.save_checkpoint( + executor=exe, + checkpoint_dir=self.checkpoint_cfg.checkpoint_dir, + trainer_id=self.trainer_id, + trainer_args=self._get_checkpoint_save_args(epoch_id, step_id), + main_program=self.train_program, + max_num_checkpoints=self.checkpoint_cfg.max_num_checkpoints) + def build_feed_var_list(program, feed_order): if not isinstance(program, framework.Program): diff --git a/python/paddle/fluid/transpiler/__init__.py b/python/paddle/fluid/transpiler/__init__.py index 045ca537b2..cf18090f71 100644 --- a/python/paddle/fluid/transpiler/__init__.py +++ b/python/paddle/fluid/transpiler/__init__.py @@ -15,10 +15,9 @@ from distribute_transpiler import DistributeTranspiler from inference_transpiler import InferenceTranspiler from memory_optimization_transpiler import memory_optimize, release_memory -from distribute_transpiler_simple import SimpleDistributeTranspiler from ps_dispatcher import HashName, RoundRobin __all__ = [ - "DistributeTranspiler", "InferenceTranspiler", "SimpleDistributeTranspiler", - "memory_optimize", "release_memory", "HashName", "RoundRobin" + "DistributeTranspiler", "InferenceTranspiler", "memory_optimize", + "release_memory", "HashName", "RoundRobin" ] diff --git a/python/paddle/fluid/transpiler/details/__init__.py b/python/paddle/fluid/transpiler/details/__init__.py new file mode 100644 index 0000000000..dc597c3384 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/__init__.py @@ -0,0 +1,16 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from program_utils import * +from ufind import * diff --git a/python/paddle/fluid/transpiler/details/program_utils.py b/python/paddle/fluid/transpiler/details/program_utils.py new file mode 100644 index 0000000000..f10b496306 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/program_utils.py @@ -0,0 +1,37 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +def delete_ops(block, ops): + try: + start = list(block.ops).index(ops[0]) + end = list(block.ops).index(ops[-1]) + [block.remove_op(start) for _ in xrange(end - start + 1)] + except Exception, e: + raise e + block.program.sync_with_cpp() + + +def find_op_by_input_arg(block, arg_name): + for index, op in enumerate(block.ops): + if arg_name in op.input_arg_names: + return index + return -1 + + +def find_op_by_output_arg(block, arg_name): + for index, op in enumerate(block.ops): + if arg_name in op.output_arg_names: + return index + return -1 diff --git a/python/paddle/fluid/transpiler/details/ufind.py b/python/paddle/fluid/transpiler/details/ufind.py new file mode 100644 index 0000000000..0e30d0e3f9 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/ufind.py @@ -0,0 +1,64 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +class UnionFind(object): + """ Union-find data structure. + + Union-find is a data structure that keeps track of a set of elements partitioned + into a number of disjoint (non-overlapping) subsets. + + Reference: + https://en.wikipedia.org/wiki/Disjoint-set_data_structure + + Args: + elements(list): The initialize element list. + """ + + def __init__(self, elementes=None): + self._parents = [] # index -> parent index + self._index = {} # element -> index + self._curr_idx = 0 + if not elementes: + elementes = [] + for ele in elementes: + self._parents.append(self._curr_idx) + self._index.update({ele: self._curr_idx}) + self._curr_idx += 1 + + def find(self, x): + # Find the root index of given element x, + # execute the path compress while findind the root index + if not x in self._index: + return -1 + idx = self._index[x] + while idx != self._parents[idx]: + t = self._parents[idx] + self._parents[idx] = self._parents[t] + idx = t + return idx + + def union(self, x, y): + # Union two given element + x_root = self.find(x) + y_root = self.find(y) + + if x_root == y_root: + return + self._parents[x_root] = y_root + + def is_connected(self, x, y): + # If two given elements have the same root index, + # then they are connected. + return self.find(x) == self.find(y) diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index e9b7d9e9d2..9c604170b8 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -11,19 +11,46 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. +""" +Transpile the program to distributed data-parallelism programs. +The main_program will be transformed to use a remote parameter server +to do parameter optimization. And the optimization graph will be put +into a parameter server program. + +Use different methods to split trainable variables to different +parameter servers. + +Steps to transpile trainer: +1. split variable to multiple blocks, aligned by product(dim[1:]) (width). +2. rename splited grad variables to add trainer_id suffix ".trainer_%d". +3. modify trainer program add split_op to each grad variable. +4. append send_op to send splited variables to server and +5. add recv_op to fetch params(splited blocks or origin param) from server. +6. append concat_op to merge splited blocks to update local weights. + +Steps to transpile pserver: +1. create new program for parameter server. +2. create params and grad variables that assigned to current server instance. +3. create a sub-block in the server side program +4. append ops that should run on current server instance. +5. add listen_and_serv op +""" from __future__ import print_function import math +import numpy as np from ps_dispatcher import RoundRobin, HashName, PSDispatcher from .. import core, framework from ..framework import Program, default_main_program, \ default_startup_program, \ Variable, Parameter, grad_var_name +from details import * LOOKUP_TABLE_TYPE = "lookup_table" LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" +OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName() RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName( ) RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC @@ -40,62 +67,11 @@ class VarBlock: return "%s:%d:%d" % (self.varname, self.offset, self.size) -class UnionFind(object): - """ Union-find data structure. - - Union-find is a data structure that keeps track of a set of elements partitioned - into a number of disjoint (non-overlapping) subsets. - - Reference: - https://en.wikipedia.org/wiki/Disjoint-set_data_structure - - Args: - elements(list): The initialize element list. - """ - - def __init__(self, elementes=None): - self._parents = [] # index -> parent index - self._index = {} # element -> index - self._curr_idx = 0 - if not elementes: - elementes = [] - for ele in elementes: - self._parents.append(self._curr_idx) - self._index.update({ele: self._curr_idx}) - self._curr_idx += 1 - - def find(self, x): - # Find the root index of given element x, - # execute the path compress while findind the root index - if not x in self._index: - return -1 - idx = self._index[x] - while idx != self._parents[idx]: - t = self._parents[idx] - self._parents[idx] = self._parents[t] - idx = t - return idx - - def union(self, x, y): - # Union two given element - x_root = self.find(x) - y_root = self.find(y) - - if x_root == y_root: - return - self._parents[x_root] = y_root - - def is_connected(self, x, y): - # If two given elements have the same root index, - # then they are connected. - return self.find(x) == self.find(y) - - def same_or_split_var(p_name, var_name): return p_name == var_name or p_name.startswith(var_name + ".block") -def split_dense_variable(var_list, service_count, min_block_size=8192): +def slice_variable(var_list, slice_count, min_block_size=8192): """ We may need to split dense tensor to one or more blocks and put them equally onto parameter server. One block is a sub-tensor @@ -103,25 +79,25 @@ def split_dense_variable(var_list, service_count, min_block_size=8192): We need to have a minimal block size so that the calculations in the parameter server side can gain better performance. By default - minimum block size 8K elements (maybe 16bit or 32bit or 64bit). + minimum block size 8K elements (maybe 16bit or 32bit or 64bit). Args: var_list (list): List of variables. - service_count (int): Numel of pserver services. A pserver may have two - or more listening ports. + slice_count (int): Numel of count that variables will be sliced, which + could be the pserver services' count. min_block_size (int): Minimum splitted block size. Returns: - blocks (list[(varname, block_id, current_block_size)]): A list + blocks (list[(varname, block_id, current_block_size)]): A list of VarBlocks. Each VarBlock specifies a shard of the var. """ blocks = [] for var in var_list: - split_count = service_count + split_count = slice_count var_numel = reduce(lambda x, y: x * y, var.shape) max_pserver_count = int(math.floor(var_numel / float(min_block_size))) if max_pserver_count == 0: max_pserver_count = 1 - if max_pserver_count < service_count: + if max_pserver_count < slice_count: split_count = max_pserver_count block_size = int(math.ceil(var_numel / float(split_count))) @@ -141,99 +117,15 @@ def split_dense_variable(var_list, service_count, min_block_size=8192): return blocks -def delete_ops(block, ops): - try: - start = list(block.ops).index(ops[0]) - end = list(block.ops).index(ops[-1]) - [block.remove_op(start) for _ in xrange(end - start + 1)] - except Exception, e: - raise e - block.program.sync_with_cpp() - - -def find_op_by_input_arg(block, arg_name): - for index, op in enumerate(block.ops): - if arg_name in op.input_arg_names: - return index - return -1 - - -def find_op_by_output_arg(block, arg_name): - for index, op in enumerate(block.ops): - if arg_name in op.output_arg_names: - return index - return -1 - - class DistributeTranspiler: - def transpile(self, - trainer_id, - program=None, - pservers="127.0.0.1:6174", - trainers=1, - split_method=RoundRobin, - sync_mode=True): - """ - Transpile the program to distributed data-parallelism programs. - The main_program will be transformed to use a remote parameter server - to do parameter optimization. And the optimization graph will be put - into a parameter server program. - - Use different methods to split trainable variables to different - parameter servers. - - Steps to transpile trainer: - 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). - 2. rename splited grad variables to add trainer_id suffix ".trainer_%d". - 3. modify trainer program add split_op to each grad variable. - 4. append send_op to send splited variables to server and fetch - params(splited blocks or origin param) from server. - 5. append concat_op to merge splited blocks to update local weights. - - Steps to transpile pserver: - 1. create new program for parameter server. - 2. create params and grad variables that assigned to current server instance. - 3. create a sub-block in the server side program - 4. append ops that should run on current server instance. - 5. add listen_and_serv op - - :param trainer_id: one unique id for each trainer in a job. - :type trainer_id: int - :param program: program to transpile, default is default_main_program - :type program: Program - :param pservers: parameter server endpoints like "m1:6174,m2:6174" - :type pservers: string - :param trainers: total number of workers/trainers in the job - :type trainers: int - :param split_method: A function to determin how to split variables - to different servers equally. - :type split_method: function - :type sync_mode: boolean default True - :param sync_mode: if sync_mode is set True, it means that dist transpiler - will transpile the program into sync_mode pserver and trainer program. - """ - assert (split_method.__bases__[0] == PSDispatcher) - if program is None: - program = default_main_program() - self.origin_program = program - self.trainer_num = trainers - self.sync_mode = sync_mode - # TODO(typhoonzero): currently trainer_id is fetched from cluster system - # like Kubernetes, we should port this to use etcd later when developing - # fluid distributed training with fault-tolerance. - self.trainer_id = trainer_id - pserver_endpoints = pservers.split(",") - self.pserver_endpoints = pserver_endpoints - self.optimize_ops, params_grads = self._get_optimize_pass() - ps_dispatcher = split_method(pserver_endpoints) - + def _has_distributed_lookup_table(self): # process lookup_table_op # 1. check all lookup_table_op is distributed # 2. check all lookup_table_op share the same table. distributed_lookup_table_ops = [] # support only one distributed_lookup_table now self.table_name = None - for op in program.global_block().ops: + for op in self.origin_program.global_block().ops: if op.type == LOOKUP_TABLE_TYPE: if op.attrs['is_distributed'] is True: if self.table_name is None: @@ -246,20 +138,13 @@ class DistributeTranspiler: if self.table_name is not None: assert op.input("W")[0] != self.table_name - self.has_distributed_lookup_table = len( - distributed_lookup_table_ops) > 0 - - # step1: For large parameters and gradients, split them into smaller - # blocks. - param_list = [] - grad_list = [] - for p, g in params_grads: - # skip parameter marked not trainable - if type(p) == Parameter and p.trainable == False: - continue - param_list.append(p) - grad_list.append(g) + return len(distributed_lookup_table_ops) > 0 + def _update_dist_lookup_table_vars(self, param_list, grad_list, + params_grads): + # TODO(wuyi): put find a way to put dist lookup table stuff all together. + # update self.table_param_grad and self.trainer_side_table_grad_list + program = self.origin_program if self.has_distributed_lookup_table: param_list = [ param for param in param_list if param.name != self.table_name @@ -277,7 +162,7 @@ class DistributeTranspiler: self.trainer_side_table_grad_list = [ program.global_block().create_var( name="%s.trainer_%d.pserver_%d" % - (table_grad_var.name, trainer_id, index), + (table_grad_var.name, self.trainer_id, index), type=table_grad_var.type, shape=table_grad_var.shape, dtype=table_grad_var.dtype) @@ -292,24 +177,58 @@ class DistributeTranspiler: dtype=table_grad_var.dtype) for index in range(len(self.pserver_endpoints)) ] + return param_list, grad_list - grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints)) - param_blocks = split_dense_variable(param_list, len(pserver_endpoints)) + def _init_splited_vars(self, slice_var_up): + # update these mappings for further transpile: + # 1. param_var_mapping: param var name -> [splited params vars] + # 2. grad_var_mapping: grad var name -> [splited grads vars] + # 3. grad_param_mapping: grad.blockx -> param.blockx + # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []} + + param_list = [] + grad_list = [] + param_grad_set = set() + for p, g in self.params_grads: + # skip parameter marked not trainable + if type(p) == Parameter and p.trainable == False: + continue + if p.name not in param_grad_set: + param_list.append(p) + param_grad_set.add(p.name) + if g.name not in param_grad_set: + grad_list.append(g) + param_grad_set.add(g.name) + + param_list, grad_list = self._update_dist_lookup_table_vars( + param_list, grad_list, self.params_grads) + + if slice_var_up: + # when we slice var up into blocks, we will slice the var according to + # pserver services' count. A pserver may have two or more listening ports. + grad_blocks = slice_variable(grad_list, len(self.pserver_endpoints)) + param_blocks = slice_variable(param_list, + len(self.pserver_endpoints)) + else: + # when we do NOT slice var up into blocks, we will always slice params + # grads into one block. + grad_blocks = slice_variable(grad_list, 1) + param_blocks = slice_variable(param_list, 1) assert (len(grad_blocks) == len(param_blocks)) - # step2: Create new vars for the parameters and gradients blocks and - # add ops to do the split. - param_var_mapping = self._create_vars_from_blocklist(program, - param_blocks) - grad_var_mapping = self._create_vars_from_blocklist( - program, grad_blocks, add_trainer_suffix=self.trainer_num > 1) - grad_param_mapping = dict() + + # origin_varname -> [splited_var] + self.param_var_mapping = self._create_vars_from_blocklist( + self.origin_program, param_blocks) + self.grad_var_mapping = self._create_vars_from_blocklist( + self.origin_program, + grad_blocks, + add_trainer_suffix=self.trainer_num > 1) + self.grad_param_mapping = dict() for g, p in zip(grad_blocks, param_blocks): g_name, g_bid, _ = g.split(":") p_name, p_bid, _ = p.split(":") - grad_param_mapping[grad_var_mapping[g_name][int(g_bid)]] = \ - param_var_mapping[p_name][int(p_bid)] - - # step 3: transpile trainer side program, insert recv op and send op. + self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ + self.param_var_mapping[p_name][int(p_bid)] # create mapping of endpoint -> split var to create pserver side program self.param_grad_ep_mapping = dict() @@ -322,11 +241,66 @@ class DistributeTranspiler: }) for ep in self.pserver_endpoints ] + def transpile(self, + trainer_id, + program=None, + pservers="127.0.0.1:6174", + trainers=1, + slice_var_up=True, + split_method=RoundRobin, + sync_mode=True): + """ + :param trainer_id: one unique id for each trainer in a job. + :type trainer_id: int + :param program: program to transpile, default is default_main_program + :type program: Program + :param pservers: parameter server endpoints like "m1:6174,m2:6174" + :type pservers: string + :param trainers: total number of workers/trainers in the job + :type trainers: int + :param split_method: A function to determin how to split variables + to different servers equally. + :type split_method: function + :type sync_mode: boolean default True + :param sync_mode: if sync_mode is set True, it means that dist transpiler + will transpile the program into sync_mode pserver and trainer program. + """ + assert (split_method.__bases__[0] == PSDispatcher) + if program is None: + program = default_main_program() + self.origin_program = program + self.trainer_num = trainers + self.sync_mode = sync_mode + self.trainer_id = trainer_id + pserver_endpoints = pservers.split(",") + self.pserver_endpoints = pserver_endpoints + self.optimize_ops, self.params_grads = self._get_optimize_pass() + + ps_dispatcher = split_method(self.pserver_endpoints) + self.has_distributed_lookup_table = self._has_distributed_lookup_table() + + # split and create vars, then put splited vars in dicts for later use. + self._init_splited_vars(slice_var_up) + # step 3.1: insert send op to send gradient vars to parameter servers ps_dispatcher.reset() send_vars = [] - for orig_varname, splited_vars in grad_var_mapping.items(): + + # in general cases, the number of pservers is times of 2, and this + # will lead to uneven distribution among weights and bias: + # fc_w@GRAD_trainer_0, fc_w@GRAD_trainer_1 --> pserver1 + # fc_b@GRAD_trainer_0, fc_b@GRAD_trainer_1 --> pserver2 + # shuffle the map will avoid the uneven distribution above + grad_var_mapping_items = self.grad_var_mapping.items() + if not slice_var_up: + np.random.shuffle(grad_var_mapping_items) + + for orig_varname, splited_vars in grad_var_mapping_items: eplist = ps_dispatcher.dispatch(splited_vars) + + if not slice_var_up: + assert (len(splited_vars) == 1) + if len(splited_vars) == 1: orig_varname = splited_vars[0].name index = find_op_by_output_arg(program.global_block(), @@ -343,7 +317,7 @@ class DistributeTranspiler: program.global_block().insert_op( index=index + 1, - type="send_vars", + type="send", inputs={"X": splited_vars}, outputs={}, attrs={ @@ -367,15 +341,16 @@ class DistributeTranspiler: # step 3.2: insert recv op to receive parameters from parameter server recv_vars = [] for _, var in enumerate(send_vars): - recv_vars.append(grad_param_mapping[var]) + recv_vars.append(self.grad_param_mapping[var]) ps_dispatcher.reset() eplist = ps_dispatcher.dispatch(recv_vars) for i, ep in enumerate(eplist): self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i]) self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i]) + # step4: Concat the parameters splits together after recv. - for varname, splited_var in param_var_mapping.iteritems(): + for varname, splited_var in self.param_var_mapping.iteritems(): eps = [] for var in splited_var: index = [v.name for v in recv_vars].index(var.name) @@ -399,7 +374,7 @@ class DistributeTranspiler: RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) - for varname, splited_var in param_var_mapping.iteritems(): + for varname, splited_var in self.param_var_mapping.iteritems(): if len(splited_var) <= 1: continue orig_param = program.global_block().vars[varname] @@ -440,7 +415,6 @@ class DistributeTranspiler: # we don't need to create them when grad arrives. # change client side var name to origin name by # removing ".trainer_%d" suffix - suff_idx = v.name.find(".trainer_") if suff_idx >= 0: orig_var_name = v.name[:suff_idx] @@ -477,24 +451,14 @@ class DistributeTranspiler: # located on current pserver opt_op_on_pserver = [] for _, op in enumerate(self.optimize_ops): - if self._is_opt_op(op) and self._is_opt_op_on_pserver(endpoint, op): + if self._is_optimizer_op(op) and self._is_opt_op_on_pserver( + endpoint, op): opt_op_on_pserver.append(op) # step 3.3 # Iterate through the ops, and if an op and the optimize ops # which located on current pserver are in one set, then # append it into the sub program. - # We try to put optimization program run parallelly, assume - # optimization program always looks like: - # - # prevop -> prevop -> opt op -> following op -> following op; -> - # prevop -> prevop -> opt op -> following op -> following op; -> - # global op -> global op - # - # we put operators that can run parallelly to many program blocks. - # in above example, we seperate ops by the ";". Global ops must run - # after all the optimize ops finished. - global_ops = [] # HACK: optimization global ops only used to scale beta1 and beta2 # replace it with dependency engine. @@ -502,12 +466,18 @@ class DistributeTranspiler: if self._is_adam_connected_op(op): global_ops.append(op) - def __append_optimize_op__(op, block, grad_to_block_id): - if self._is_opt_op(op): + def __append_optimize_op__(op, block, grad_to_block_id, merged_var): + if self._is_optimizer_op(op): self._append_pserver_ops(block, op, endpoint, grad_to_block_id, - self.origin_program) + self.origin_program, merged_var) else: - self._append_pserver_non_opt_ops(block, op) + self._append_pserver_non_opt_ops(block, op, endpoint) + + def __op_have_grad_input__(op): + for varname in op.input_arg_names: + if varname.find("@GRAD") >= 0: + return varname + return "" # append lr decay ops to the child block if exists lr_ops = self._get_lr_ops() @@ -515,17 +485,26 @@ class DistributeTranspiler: lr_decay_block = pserver_program.create_block( pserver_program.num_blocks - 1) for _, op in enumerate(lr_ops): - self._append_pserver_non_opt_ops(lr_decay_block, op) + self._append_pserver_non_opt_ops(lr_decay_block, op, endpoint) # append op to the current block grad_to_block_id = [] pre_block_idx = pserver_program.num_blocks - 1 for idx, opt_op in enumerate(opt_op_on_pserver): per_opt_block = pserver_program.create_block(pre_block_idx) + # append grad merging ops before clip and weight decay + for _, op in enumerate(self.optimize_ops): + # find the origin @GRAD var before clipping + grad_varname_for_block = __op_have_grad_input__(op) + if ufind.is_connected(op, opt_op) and grad_varname_for_block: + merged_var = self._append_pserver_grad_merge_ops( + per_opt_block, grad_varname_for_block, endpoint, + grad_to_block_id, self.origin_program) for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if ufind.is_connected(op, opt_op) and op not in global_ops: - __append_optimize_op__(op, per_opt_block, grad_to_block_id) + __append_optimize_op__(op, per_opt_block, grad_to_block_id, + merged_var) # append global ops if global_ops: @@ -533,46 +512,41 @@ class DistributeTranspiler: pserver_program.num_blocks - 1) for glb_op in global_ops: __append_optimize_op__(glb_op, opt_state_block, - grad_to_block_id) - - # NOT USED: single block version: - # - # for _, op in enumerate(self.optimize_ops): - # for _, opt_op in enumerate(opt_op_on_pserver): - # if ufind.is_connected(op, opt_op): - # __append_optimize_op__(glb_op, optimize_block) - # break + grad_to_block_id, None) # process distributed lookup_table - prefetch_block = None + prefetch_var_name_to_block_id = [] if self.has_distributed_lookup_table: pserver_index = self.pserver_endpoints.index(endpoint) table_opt_block = self._create_table_optimize_block( pserver_index, pserver_program, pre_block_idx, grad_to_block_id) - prefetch_block = self._create_prefetch_block( + prefetch_var_name_to_block_id = self._create_prefetch_block( pserver_index, pserver_program, table_opt_block) # NOTE: if has_distributed_lookup_table is False, then prefetch_block will # not be executed, so it's safe to use optimize_block to hold the place if self.has_distributed_lookup_table: - assert prefetch_block is not None + assert len(prefetch_var_name_to_block_id) > 0 else: - assert prefetch_block is None - prefetch_block = pserver_program.global_block() + assert len(prefetch_var_name_to_block_id) == 0 + + attrs = { + "OptimizeBlock": pserver_program.block(1), + "endpoint": endpoint, + "Fanin": self.trainer_num, + "sync_mode": self.sync_mode, + "grad_to_block_id": grad_to_block_id + } + if len(prefetch_var_name_to_block_id) > 0: + attrs['prefetch_var_name_to_block_id'] \ + = prefetch_var_name_to_block_id # step5 append the listen_and_serv op pserver_program.global_block().append_op( type="listen_and_serv", inputs={'X': recv_inputs}, outputs={}, - attrs={ - "OptimizeBlock": pserver_program.block(1), - "endpoint": endpoint, - "Fanin": self.trainer_num, - "PrefetchBlock": prefetch_block, - "sync_mode": self.sync_mode, - "grad_to_block_id": grad_to_block_id - }) + attrs=attrs) pserver_program.sync_with_cpp() return pserver_program @@ -631,12 +605,21 @@ class DistributeTranspiler: attrs=op.attrs) return s_prog + # ====================== private transpiler functions ===================== + # transpiler function for dis lookup_table def _replace_lookup_table_op_with_prefetch(self, program, pserver_endpoints): # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op - self.prefetch_input_vars = None - self.prefetch_output_vars = None + # self.all_prefetch_input_vars = + # [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1] + # [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]] + self.all_prefetch_input_vars = [] + + # self.all_prefetch_input_vars = + # [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1] + # [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]] + self.all_prefetch_output_vars = [] continue_search_lookup_table_op = True while continue_search_lookup_table_op: @@ -646,26 +629,27 @@ class DistributeTranspiler: if op.type == LOOKUP_TABLE_TYPE: continue_search_lookup_table_op = True - op_index = list(all_ops).index(op) + lookup_table_op_index = list(all_ops).index(op) ids_name = op.input("Ids") out_name = op.output("Out") - if self.prefetch_input_vars is None: - ids_var = program.global_block().vars[ids_name[0]] - self.prefetch_input_vars = self.create_splited_vars( - source_var=ids_var, - block=program.global_block(), - tag="_prefetch_in_") - if self.prefetch_output_vars is None: - out_var = program.global_block().vars[out_name[0]] - self.prefetch_output_vars = self.create_splited_vars( - source_var=out_var, - block=program.global_block(), - tag="_prefetch_out_") + ids_var = program.global_block().vars[ids_name[0]] + prefetch_input_vars = self.create_splited_vars( + source_var=ids_var, + block=program.global_block(), + tag="_prefetch_in_") + self.all_prefetch_input_vars.append(prefetch_input_vars) + + out_var = program.global_block().vars[out_name[0]] + prefetch_output_vars = self.create_splited_vars( + source_var=out_var, + block=program.global_block(), + tag="_prefetch_out_") + self.all_prefetch_output_vars.append(prefetch_output_vars) # insert split_ids_op program.global_block().insert_op( - index=op_index, + index=lookup_table_op_index, type="split_ids", inputs={ 'Ids': [ @@ -673,14 +657,14 @@ class DistributeTranspiler: for varname in ids_name ] }, - outputs={"Out": self.prefetch_input_vars}) + outputs={"Out": prefetch_input_vars}) # insert prefetch_op program.global_block().insert_op( - index=op_index + 1, + index=lookup_table_op_index + 1, type="prefetch", - inputs={'X': self.prefetch_input_vars}, - outputs={"Out": self.prefetch_output_vars}, + inputs={'X': prefetch_input_vars}, + outputs={"Out": prefetch_output_vars}, attrs={ "epmap": pserver_endpoints, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE @@ -688,16 +672,21 @@ class DistributeTranspiler: # insert concat_op program.global_block().insert_op( - index=op_index + 2, - type="concat", - inputs={'X': self.prefetch_output_vars}, + index=lookup_table_op_index + 2, + type="merge_ids", + inputs={ + 'Ids': [ + program.global_block().vars[varname] + for varname in ids_name + ], + 'X': prefetch_output_vars + }, outputs={ "Out": [ program.global_block().vars[varname] for varname in out_name ] - }, - attrs={"axis": 0}) + }) # delete lookup_table_op delete_ops(program.global_block(), [op]) @@ -705,7 +694,7 @@ class DistributeTranspiler: break def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints): - # 2. add split_ids_op and send_vars_op to send gradient to pservers + # 2. add split_ids_op and send_op to send gradient to pservers # there should only be one table_name all_ops = program.global_block().ops table_grad_name = grad_var_name(self.table_name) @@ -722,11 +711,11 @@ class DistributeTranspiler: outputs={"Out": self.trainer_side_table_grad_list}) program.global_block().insert_op( index=op_index + 2, - type="send_vars", + type="send", inputs={'X': self.trainer_side_table_grad_list}, outputs={}, attrs={ - "sync_send": True, + "sync_mode": True, "epmap": pserver_endpoints, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) @@ -736,30 +725,34 @@ class DistributeTranspiler: optimize_block): # STEP: create prefetch block table_var = pserver_program.global_block().vars[self.table_name] - prefetch_block = pserver_program.create_block(optimize_block.idx) - trainer_ids = self.prefetch_input_vars[pserver_index] - pserver_ids = pserver_program.global_block().create_var( - name=trainer_ids.name, - type=trainer_ids.type, - shape=trainer_ids.shape, - dtype=trainer_ids.dtype) - trainer_out = self.prefetch_output_vars[pserver_index] - pserver_out = pserver_program.global_block().create_var( - name=trainer_out.name, - type=trainer_out.type, - shape=trainer_out.shape, - dtype=trainer_out.dtype) - prefetch_block.append_op( - type="lookup_sparse_table", - inputs={'Ids': pserver_ids, - "W": table_var}, - outputs={"Out": pserver_out}, - attrs={ - "is_sparse": True, # has no effect on lookup_table op - "is_distributed": True, - "padding_idx": -1 - }) - return prefetch_block + prefetch_var_name_to_block_id = [] + for index in range(len(self.all_prefetch_input_vars)): + prefetch_block = pserver_program.create_block(optimize_block.idx) + trainer_ids = self.all_prefetch_input_vars[index][pserver_index] + pserver_ids = pserver_program.global_block().create_var( + name=trainer_ids.name, + type=trainer_ids.type, + shape=trainer_ids.shape, + dtype=trainer_ids.dtype) + trainer_out = self.all_prefetch_output_vars[index][pserver_index] + pserver_out = pserver_program.global_block().create_var( + name=trainer_out.name, + type=trainer_out.type, + shape=trainer_out.shape, + dtype=trainer_out.dtype) + prefetch_block.append_op( + type="lookup_sparse_table", + inputs={'Ids': pserver_ids, + "W": table_var}, + outputs={"Out": pserver_out}, + attrs={ + "is_sparse": True, # has no effect on lookup_table op + "is_distributed": True, + "padding_idx": -1 + }) + prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str( + prefetch_block.idx)) + return prefetch_var_name_to_block_id def _create_table_optimize_block(self, pserver_index, pserver_program, pre_block_idx, grad_to_block_id): @@ -836,7 +829,6 @@ class DistributeTranspiler: return table_opt_block - # ====================== private transpiler functions ===================== def _create_vars_from_blocklist(self, program, block_list, @@ -849,8 +841,8 @@ class DistributeTranspiler: program (ProgramDesc): ProgramDesc which gradients blong. block_list (list[(varname, block_id, block_size)]): List of gradient blocks. add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True. - Returns: - var_mapping (dict(varname->[new_varname_variable])):A dict mapping + Returns: + var_mapping (dict(varname->[new_varname_variable])):A dict mapping from original var name to each var split. """ @@ -863,6 +855,9 @@ class DistributeTranspiler: if not block_map.has_key(varname): block_map[varname] = [] block_map[varname].append((long(offset), long(size))) + # Do not remove this important debug message: + print("block map: %s" % block_map) + for varname, splited in block_map.iteritems(): orig_var = program.global_block().var(varname) if len(splited) == 1: @@ -979,17 +974,74 @@ class DistributeTranspiler: pass return orig_shape - def _orig_varname(self, varname): - suff_idx = varname.find(".trainer_") + def _get_varname_parts(self, varname): + # returns origin, blockid, trainerid orig_var_name = "" - if suff_idx >= 0: - orig_var_name = varname[:suff_idx] + trainer_part = "" + block_part = "" + trainer_idx = varname.find(".trainer_") + if trainer_idx >= 0: + trainer_part = varname[trainer_idx + 1:] + else: + trainer_idx = len(varname) + block_index = varname.find(".block") + if block_index >= 0: + block_part = varname[block_index + 1:trainer_idx] + else: + block_index = len(varname) + orig_var_name = varname[0:min(block_index, trainer_idx)] + return orig_var_name, block_part, trainer_part + + def _orig_varname(self, varname): + orig, _, _ = self._get_varname_parts(varname) + return orig + + def _append_pserver_grad_merge_ops(self, optimize_block, + grad_varname_for_block, endpoint, + grad_to_block_id, origin_program): + program = optimize_block.program + pserver_block = program.global_block() + grad_block = None + for g in self.param_grad_ep_mapping[endpoint]["grads"]: + if self._orig_varname(g.name) == \ + self._orig_varname(grad_varname_for_block): + grad_block = g + break + if not grad_block: + # do not append this op if current endpoint + # is not dealing with this grad block + return + orig_varname, block_name, trainer_name = self._get_varname_parts( + grad_block.name) + if block_name: + merged_var_name = '.'.join([orig_varname, block_name]) else: - orig_var_name = varname - return orig_var_name + merged_var_name = orig_varname + merged_var = \ + pserver_block.vars[merged_var_name] + grad_to_block_id.append(merged_var.name + ":" + str(optimize_block.idx)) + if self.sync_mode and self.trainer_num > 1: + vars2merge = [] + for i in xrange(self.trainer_num): + per_trainer_name = "%s.trainer_%d" % \ + (merged_var_name, i) + vars2merge.append(pserver_block.vars[per_trainer_name]) + + optimize_block.append_op( + type="sum", + inputs={"X": vars2merge}, + outputs={"Out": merged_var}) + # TODO(panyx0718): What if it's SELECTED_ROWS. + if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: + optimize_block.append_op( + type="scale", + inputs={"X": merged_var}, + outputs={"Out": merged_var}, + attrs={"scale": 1.0 / float(self.trainer_num)}) + return merged_var def _append_pserver_ops(self, optimize_block, opt_op, endpoint, - grad_to_block_id, origin_program): + grad_to_block_id, origin_program, merged_var): program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() @@ -997,40 +1049,6 @@ class DistributeTranspiler: # moment can use the updated shape for key in opt_op.input_names: if key == "Grad": - grad_block = None - for g in self.param_grad_ep_mapping[endpoint]["grads"]: - if same_or_split_var( - self._orig_varname(g.name), - self._orig_varname(opt_op.input(key)[0])): - grad_block = g - break - if not grad_block: - # do not append this op if current endpoint - # is not dealing with this grad block - return - merged_var = \ - pserver_block.vars[self._orig_varname(grad_block.name)] - grad_to_block_id.append(merged_var.name + ":" + str( - optimize_block.idx)) - if self.sync_mode and self.trainer_num > 1: - vars2merge = [] - for i in xrange(self.trainer_num): - per_trainer_name = "%s.trainer_%d" % \ - (self._orig_varname(grad_block.name), i) - vars2merge.append(pserver_block.vars[per_trainer_name]) - - optimize_block.append_op( - type="sum", - inputs={"X": vars2merge}, - outputs={"Out": merged_var}) - # TODO(panyx0718): What if it's SELECTED_ROWS. - if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: - optimize_block.append_op( - type="scale", - inputs={"X": merged_var}, - outputs={"Out": merged_var}, - attrs={"scale": 1.0 / float(self.trainer_num)}) - new_inputs[key] = merged_var elif key == "Param": # param is already created on global program @@ -1089,17 +1107,31 @@ class DistributeTranspiler: outputs=outputs, attrs=opt_op.attrs) - def _append_pserver_non_opt_ops(self, optimize_block, opt_op): + def _is_splited_grad_var(self, var, var_dict): + grad_block = None + for _, g in var_dict.iteritems(): + if self._orig_varname(g.name) == self._orig_varname(var.name): + if g.name.find(".trainer_") == -1: + grad_block = g + break + return grad_block + + def _append_pserver_non_opt_ops(self, optimize_block, opt_op, endpoint): program = optimize_block.program # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op( self.origin_program.global_block().vars, opt_op) - for varlist in inputs.itervalues(): + for key, varlist in inputs.iteritems(): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - if not program.global_block().vars.has_key(var.name): + # for ops like clipping and weight decay, get the splited var + # for inputs/outputs + grad_block = self._is_splited_grad_var( + var, program.global_block().vars) + if grad_block: + inputs[key] = grad_block + elif not program.global_block().vars.has_key(var.name): program.global_block().create_var( name=var.name, persistable=var.persistable, @@ -1108,13 +1140,16 @@ class DistributeTranspiler: outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) - - for varlist in outputs.itervalues(): + for key, varlist in outputs.iteritems(): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - program.global_block().clone_variable(var) + grad_block = self._is_splited_grad_var( + var, program.global_block().vars) + if grad_block: + outputs[key] = grad_block + elif not program.global_block().vars.has_key(var.name): + program.global_block().clone_variable(var) optimize_block.append_op( type=opt_op.type, @@ -1160,9 +1195,17 @@ class DistributeTranspiler: ufind.union(op1, op2) return ufind - def _is_opt_op(self, op): - # NOTE: It's a HACK implement. - # optimize op: SGDOptimize, MomentumOptimizer, AdamOptimizer and etc... + def _is_opt_role_op(self, op): + # NOTE: depend on oprole to find out whether this op is for + # optimize + op_maker = core.op_proto_and_checker_maker + optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize + if op_maker.kOpRoleAttrName() in op.attrs and \ + int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role): + return True + return False + + def _is_optimizer_op(self, op): if "Param" in op.input_names and \ "LearningRate" in op.input_names: return True @@ -1212,7 +1255,7 @@ class DistributeTranspiler: # find learning rate variables by optimize op lr_vars = set() for op in self.optimize_ops: - if self._is_opt_op(op): + if self._is_optimizer_op(op): lr_vars.add(op.input("LearningRate")[0]) find_ops = [] @@ -1229,7 +1272,7 @@ class DistributeTranspiler: # NOTE: we need to skip all optimize ops, since it is connected # with forward/backward ops and lr ops, we only need the lr ops. if op1 != op2 and self._is_op_connected(op1, op2) and \ - not self._is_opt_op(op1) and not self._is_opt_op(op2): + not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2): ufind.union(op1, op2) # find all ops which is related with lr var for op1 in block.ops: @@ -1250,13 +1293,21 @@ class DistributeTranspiler: block = self.origin_program.global_block() opt_ops = [] params_grads = [] + origin_var_dict = self.origin_program.global_block().vars for op in block.ops: - if self._is_opt_op(op): + if self._is_opt_role_op(op): opt_ops.append(op) - params_grads.append((self.origin_program.global_block().var( - op.input("Param")[0]), - self.origin_program.global_block().var( - op.input("Grad")[0]))) + # HACK(wuyi): if we find grad vars from input of optimize + # ops, we may get the output of clip op. Use syntax "@GRAD" + # and op_role_var to get the pair. + for input_name in op.input_arg_names: + if input_name.find("@GRAD") != -1 and \ + op.attrs[RPC_OP_ROLE_ATTR_NAME]: + param_name = op.attrs[OP_ROLE_VAR_ATTR_NAME][0] + params_grads.append([ + origin_var_dict[param_name], + origin_var_dict[input_name] + ]) elif self._is_adam_connected_op(op): opt_ops.append(op) else: diff --git a/python/paddle/fluid/transpiler/distribute_transpiler_simple.py b/python/paddle/fluid/transpiler/distribute_transpiler_simple.py deleted file mode 100644 index ea8c27cdca..0000000000 --- a/python/paddle/fluid/transpiler/distribute_transpiler_simple.py +++ /dev/null @@ -1,254 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -from ..framework import Program, default_main_program, Parameter, Variable -from ..layer_helper import LayerHelper - - -def hash_name_to_server(params_grads, pserver_endpoints): - """ - :param param_grads: - :return: a map of pserver endpoint -> - params -> [param list] - grads -> [grad list] - """ - - def _hash_param(param_name, total): - return hash(param_name) % total - - param_grad_map = dict() - for param, grad in params_grads: - if param.trainable is True and grad is not None: - server_id = _hash_param(param.name, len(pserver_endpoints)) - server_for_param = pserver_endpoints[server_id] - if not param_grad_map.has_key(server_for_param): - param_grad_map[server_for_param] = {"params": [], "grads": []} - param_grad_map[server_for_param]["params"].append(param) - param_grad_map[server_for_param]["grads"].append(grad) - - return param_grad_map - - -def round_robin(params_grads, pserver_endpoints): - assert (len(params_grads) > len(pserver_endpoints)) - - param_grad_map = dict() - pserver_idx = 0 - for param, grad in params_grads: - if param.trainable is True: - server_for_param = pserver_endpoints[pserver_idx] - if not param_grad_map.has_key(server_for_param): - param_grad_map[server_for_param] = {"params": [], "grads": []} - - param_grad_map[server_for_param]["params"].append(param) - param_grad_map[server_for_param]["grads"].append(grad) - - pserver_idx += 1 - if pserver_idx >= len(pserver_endpoints): - pserver_idx = 0 - return param_grad_map - - -class SimpleDistributeTranspiler: - def transpile(self, - optimize_ops, - params_grads, - program=None, - pservers="127.0.0.1:6174", - trainers=1, - split_method=round_robin): - """ - Transpile the program to a distributed data-parallelism programs. - - The main_program will be transform to use a remote parameter server - to do parameter optimization. And the optimization graph will be put - in to a parameter server program. - - Use different methods to split trainable varialbles to different - parameter servers. - - Example to run: - - exe = fluid.Executor(place) - t = fluid.DistributeTranspiler() - t.transpile(optimize_ops, params_grads, pservers="127.0.0.1:6174", trainers=1) - - pserver_endpoint = os.getenv("PSERVER") - if pserver_endpoint: - pserver_prog = t.get_pserver_program(pserver_endpoint, optimize_ops) - exe.run(fluid.default_startup_program()) - exe.run(pserver_prog) - else: - feeder = fluid.DataFeeder(feed_list=[images, label], place=place) - exe.run(fluid.default_startup_program()) - - for pass_id in range(PASS_NUM): - ... - - :param optimize_ops: op list of optimization, should be the - return value of Optimizer.minimize - :type optimize_ops: list - :param program: program to optimize, default default_main_program - :param pservers: parameter server endpoints like "m1:6174,m2:6174" - :type pservers: string - - :return: return a list of programs - """ - if program is None: - program = default_main_program() - self.program = program - self.trainers = trainers - self.optimize_ops = optimize_ops - self._optimize_distributed( - optimize_ops, - program, - params_grads, - pservers=pservers, - trainers=trainers, - split_method=split_method) - - def _clone_param(self, block, v): - assert isinstance(v, Parameter) - new_p = Parameter( - block=block, - shape=v.shape, - dtype=v.dtype, - type=v.type, - lod_level=v.lod_level, - stop_gradient=v.stop_gradient, - trainable=v.trainable, - optimize_attr=v.optimize_attr, - regularizer=v.regularizer, - name=v.name) - block.vars[new_p.name] = new_p - - def _clone_var(self, block, var): - assert isinstance(var, Variable) - return block.create_var( - name=var.name, - shape=var.shape, - dtype=var.dtype, - type=var.type, - lod_level=var.lod_level, - persistable=var.persistable) - - def _optimize_distributed(self, optimize_ops, program, params_and_grads, - **kwargs): - if kwargs.has_key("split_method"): - split_method = kwargs["split_method"] - else: - split_method = round_robin - - assert (callable(split_method)) - pserver_endpoints = kwargs["pservers"].split(",") - self.param_grad_map = split_method(params_and_grads, pserver_endpoints) - - send_op_ordered_inputs = [] - send_op_ordered_outputs = [] - epmap = [] - for ep, v in self.param_grad_map.iteritems(): - send_op_ordered_inputs.extend(v["grads"]) - send_op_ordered_outputs.extend(v["params"]) - for i in v["grads"]: - epmap.append(ep) - send_op = program.global_block().append_op( - type="send", - inputs={"X": send_op_ordered_inputs - }, # inputs is a list of tensors to be send - outputs={"Out": send_op_ordered_outputs}, - attrs={"endpoints": pserver_endpoints, - "epmap": epmap}) - - def get_trainer_program(self): - # remove optimize ops and add a send op to main_program - self.program.global_block().delete_ops(self.optimize_ops) - return self.program - - def _create_var_for_trainers(self, block, var, trainers): - var_list = [] - for i in xrange(trainers): - var_each = block.create_var( - name="%s.trainer_%d" % (var.name, i), - psersistable=var.persistable, - dtype=var.dtype, - shape=var.shape) - var_list.append(var_each) - return var_list - - def get_pserver_program(self, endpoint, optimize_ops): - pserver_program = Program() - for v in self.param_grad_map[endpoint]["params"]: - self._clone_param(pserver_program.global_block(), v) - - optimize_sub_program = Program() - grad_var_names = [ - var.name for var in self.param_grad_map[endpoint]["grads"] - ] - for opt_op in optimize_ops: - for _, var in opt_op.inputs.iteritems(): - # NOTE: append operators to merge gradients from multiple - # trainers. If trainers == 1, this is not needed. - if self.trainers > 1 and var.name in grad_var_names: - vars2merge = self._create_var_for_trainers( - optimize_sub_program.global_block(), var, self.trainers) - merged_var = optimize_sub_program.global_block().create_var( - name=var.name, - persistable=var.persistable, - dtype=var.dtype, - shape=var.shape) - optimize_sub_program.global_block().append_op( - type="sum", - inputs={"X": vars2merge}, - outputs={"Out": merged_var}) - optimize_sub_program.global_block().append_op( - type="scale", - inputs={"X": merged_var}, - outputs={"Out": merged_var}, - attrs={"scale": 1.0 / float(self.trainers)}) - else: - optimize_sub_program.global_block().create_var( - name=var.name, - persistable=var.persistable, - dtype=var.dtype, - shape=var.shape) - - if opt_op.inputs.has_key("Grad"): - if opt_op.inputs["Grad"].name in grad_var_names: - optimize_sub_program.global_block().append_op( - type=opt_op.type, - inputs=opt_op.inputs, - outputs=opt_op.outputs, - attrs=opt_op.attrs) - else: - optimize_sub_program.global_block().append_op( - type=opt_op.type, - inputs=opt_op.inputs, - outputs=opt_op.outputs, - attrs=opt_op.attrs) - pserver_program.global_block().append_op( - type="recv", - inputs={"RX": - self.param_grad_map[endpoint]["grads"]}, # grads to recv - outputs={}, - attrs={ - "OptimizeBlock": optimize_sub_program.global_block(), - "endpoint": endpoint, - "ParamList": - [p.name for p in self.param_grad_map[endpoint]["params"]], - "GradList": - [p.name for p in self.param_grad_map[endpoint]["grads"]], - "Trainers": self.trainers - }) - pserver_program.sync_with_cpp() - return pserver_program diff --git a/python/paddle/trainer_config_helpers/attrs.py b/python/paddle/trainer_config_helpers/attrs.py index e6f87ce61b..4e3beaf639 100644 --- a/python/paddle/trainer_config_helpers/attrs.py +++ b/python/paddle/trainer_config_helpers/attrs.py @@ -240,14 +240,15 @@ class ExtraLayerAttribute(object): :type error_clipping_threshold: float :param drop_rate: Dropout rate. Dropout will create a mask on layer output. The dropout rate is the zero rate of this mask. The - details of what dropout is please refer to `here - `_. + details of what dropout is please refer to `JMLRdropout + `_. :type drop_rate: float :param device: device ID of layer. device=-1, use CPU. device>=0, use GPU. - The details allocation in parallel_nn please refer to `here - `_. + The details allocation in parallel_nn please refer to `use_case + `_. :type device: int """ diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index ebc31b23e0..e6a03759ef 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -2556,7 +2556,7 @@ def img_conv_layer(input, the output will be obtained by concatenating the two results. The details of grouped convolution, please refer to: - `ImageNet Classification with Deep Convolutional Neural Networks + `ImageNet Classification With Deep Convolutional Neural Networks `_ The example usage is: @@ -5678,8 +5678,8 @@ def warp_ctc_layer(input, `_ library, which is used in `Deep Speech 2: End-toEnd Speech Recognition in English and Mandarin `_, to compute Connectionist Temporal - Classification (CTC) loss. Besides, another `warp-ctc - `_ repository, which is forked from + Classification (CTC) loss. Besides, another `warp-ctc repository + `_ , which is forked from the official one, is maintained to enable more compiling options. During the building process, PaddlePaddle will clone the source codes, build and install it to :code:`third_party/install/warpctc` directory. diff --git a/python/paddle/v2/dataset/flowers.py b/python/paddle/v2/dataset/flowers.py index 7bdddeaabe..357a4e9b00 100644 --- a/python/paddle/v2/dataset/flowers.py +++ b/python/paddle/v2/dataset/flowers.py @@ -119,7 +119,8 @@ def reader_creator(data_file, yield sample, int(label) - 1 if use_xmap: - return xmap_readers(mapper, reader, cpu_count(), buffered_size) + cpu_num = int(os.environ.get('CPU_NUM', cpu_count())) + return xmap_readers(mapper, reader, cpu_num, buffered_size) else: return map_readers(mapper, reader) diff --git a/python/paddle/v2/minibatch.py b/python/paddle/v2/minibatch.py index 317cf037c6..3c6a53db3c 100644 --- a/python/paddle/v2/minibatch.py +++ b/python/paddle/v2/minibatch.py @@ -15,7 +15,7 @@ __all__ = ['batch'] -def batch(reader, batch_size): +def batch(reader, batch_size, drop_last=True): """ Create a batched reader. @@ -23,6 +23,8 @@ def batch(reader, batch_size): :type reader: callable :param batch_size: size of each mini-batch :type batch_size: int + :param drop_last: drop the last batch, if the size of last batch is not equal to batch_size. + :type drop_last: bool :return: the batched reader. :rtype: callable """ @@ -35,7 +37,7 @@ def batch(reader, batch_size): if len(b) == batch_size: yield b b = [] - if b: + if drop_last == False and len(b) != 0: yield b return batch_reader diff --git a/python/setup.py.in b/python/setup.py.in index c42601d335..8257f1d5e2 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -69,7 +69,8 @@ packages=['paddle', 'paddle.fluid.proto', 'paddle.fluid.proto.profiler', 'paddle.fluid.layers', - 'paddle.fluid.transpiler'] + 'paddle.fluid.transpiler', + 'paddle.fluid.transpiler.details'] if '${WITH_FLUID_ONLY}'== 'OFF': packages+=['paddle.proto', diff --git a/tools/codestyle/.gitignore b/tools/codestyle/.gitignore new file mode 100644 index 0000000000..0d20b6487c --- /dev/null +++ b/tools/codestyle/.gitignore @@ -0,0 +1 @@ +*.pyc diff --git a/tools/codestyle/docstring_checker.py b/tools/codestyle/docstring_checker.py index 48100e5bf9..54a6904626 100644 --- a/tools/codestyle/docstring_checker.py +++ b/tools/codestyle/docstring_checker.py @@ -126,9 +126,10 @@ class DocstringChecker(BaseChecker): 'W9002': ('Doc string does not end with "." period', symbol + "-end-with", 'Used when a doc string does not end with a period'), - 'W9003': ('All args with their types must be mentioned in doc string', - symbol + "-with-all-args", - 'Used when not all arguments are in the doc string '), + 'W9003': + ('All args with their types must be mentioned in doc string %s', + symbol + "-with-all-args", + 'Used when not all arguments are in the doc string '), 'W9005': ('Missing docstring or docstring is too short', symbol + "-missing", 'Add docstring longer >=10'), 'W9006': ('Docstring indent error, use 4 space for indent', @@ -178,6 +179,8 @@ class DocstringChecker(BaseChecker): self.indent_style(node) def missing_doc_string(self, node): + if node.name.startswith("__") or node.name.startswith("_"): + return True if node.tolineno - node.fromlineno <= 10: return True @@ -199,12 +202,16 @@ class DocstringChecker(BaseChecker): doc = node.doc lines = doc.splitlines() + line_num = 0 for l in lines: + if line_num == 0: + continue cur_indent = len(l) - len(l.lstrip()) if cur_indent % indent != 0: self.add_message('W9006', node=node, line=node.fromlineno) return False + line_num += 1 return True @@ -320,15 +327,19 @@ class DocstringChecker(BaseChecker): return True parsed_args = doc.args + args_not_documented = set(args) - set(parsed_args) if len(args) > 0 and len(parsed_args) <= 0: - print "debug:parsed args: ", parsed_args - self.add_message('W9003', node=node, line=node.fromlineno) + self.add_message( + 'W9003', + node=node, + line=node.fromlineno, + args=list(args_not_documented)) return False for t in args: if t not in parsed_args: - print t, " with (type) not in ", parsed_args - self.add_message('W9003', node=node, line=node.fromlineno) + self.add_message( + 'W9003', node=node, line=node.fromlineno, args=[t, ]) return False return True diff --git a/tools/codestyle/docstring_checker.pyc b/tools/codestyle/docstring_checker.pyc deleted file mode 100644 index 1ce612ca23..0000000000 Binary files a/tools/codestyle/docstring_checker.pyc and /dev/null differ diff --git a/tools/codestyle/pylint_pre_commit.hook b/tools/codestyle/pylint_pre_commit.hook index e7c92ba671..150a3f5666 100755 --- a/tools/codestyle/pylint_pre_commit.hook +++ b/tools/codestyle/pylint_pre_commit.hook @@ -7,13 +7,13 @@ DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" export PYTHONPATH=$DIR:$PYTHONPATH # The trick to remove deleted files: https://stackoverflow.com/a/2413151 -for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do +for file in $(git diff --name-status | awk '$1 != "D" {print $2}'); do pylint --disable=all --load-plugins=docstring_checker \ --enable=doc-string-one-line,doc-string-end-with,doc-string-with-all-args,doc-string-triple-quotes,doc-string-missing,doc-string-indent-error,doc-string-with-returns,doc-string-with-raises $file; TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?); done -#exit $TOTAL_ERRORS +exit $TOTAL_ERRORS #For now, just warning: -exit 0 +#exit 0