m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/simplify_data_structures

Browse Source
fea/docker_cudnn7
Yu Yang 7 years ago
parent f051c768e5 31464f3425
commit ad40faaafb
80 changed files with 1439 additions and 1011 deletions
Show Stats Download Patch File Download Diff File

66
benchmark/fluid/machine_translation.py
Unescape View File

@ -48,6 +48,13 @@ parser.add_argument(
type=int,
default=16,
help="The sequence number of a mini-batch data. (default: %(default)d)")
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(
"--dict_size",
type=int,
@ -72,16 +79,21 @@ parser.add_argument(
default=3,
help="The width for beam searching. (default: %(default)d)")
parser.add_argument(
"--use_gpu",
type=distutils.util.strtobool,
default=True,
help="Whether to use gpu. (default: %(default)d)")
'--device',
type=str,
default='GPU',
choices=['CPU', 'GPU'],
help="The device type.")
parser.add_argument(
"--max_length",
type=int,
default=250,
help="The maximum length of sequence when doing generation. "
"(default: %(default)d)")
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
def lstm_step(x_t, hidden_t_prev, cell_t_prev, size):
@ -281,7 +293,7 @@ def train():
paddle.dataset.wmt14.test(args.dict_size), buf_size=1000),
batch_size=args.batch_size)
place = core.CUDAPlace(0) if args.use_gpu else core.CPUPlace()
place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
exe = Executor(place)
exe.run(framework.default_startup_program())
@ -307,14 +319,20 @@ def train():
return total_loss / count
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in xrange(args.pass_num):
pass_start_time = time.time()
words_seen = 0
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_batch_generator()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
src_seq, word_num = to_lodtensor(map(lambda x: x[0], data), place)
words_seen += word_num
num_samples += word_num
trg_seq, word_num = to_lodtensor(map(lambda x: x[1], data), place)
words_seen += word_num
num_samples += word_num
lbl_seq, _ = to_lodtensor(map(lambda x: x[2], data), place)
fetch_outs = exe.run(framework.default_main_program(),
@ -325,24 +343,36 @@ def train():
},
fetch_list=[avg_cost])
avg_cost_val = np.array(fetch_outs[0])
print('pass_id=%d, batch_id=%d, train_loss: %f' %
(pass_id, batch_id, avg_cost_val))
iters += 1
loss = np.array(fetch_outs[0])
print(
"Pass = %d, Iter = %d, Loss = %f" % (pass_id, iters, loss)
) # The accuracy is the accumulation of batches, but not the current batch.
pass_end_time = time.time()
test_loss = do_validation()
time_consumed = pass_end_time - pass_start_time
words_per_sec = words_seen / time_consumed
print("pass_id=%d, test_loss: %f, words/s: %f, sec/pass: %f" %
(pass_id, test_loss, words_per_sec, time_consumed))
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))
# evaluation
if args.with_test:
test_loss = do_validation()
exit(0)
def infer():
pass
def print_arguments(args):
print('----------- seq2seq Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
args = parser.parse_args()
print_arguments(args)
if args.infer_only:
infer()
else:

64
benchmark/fluid/mnist.py
Unescape View File

@ -35,6 +35,12 @@ def parse_args():
parser = argparse.ArgumentParser("mnist model benchmark.")
parser.add_argument(
'--batch_size', type=int, default=128, help='The minibatch size.')
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=35, help='The number of minibatches.')
parser.add_argument(
@ -53,19 +59,14 @@ def parse_args():
'--use_nvprof',
action='store_true',
help='If set, use nvprof for CUDA.')
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
args = parser.parse_args()
return args
def print_arguments(args):
vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
vars(args)['device'] == 'GPU')
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
def cnn_model(data):
conv_pool_1 = fluid.nets.simple_img_conv_pool(
input=data,
@ -161,16 +162,22 @@ def run_benchmark(model, args):
paddle.dataset.mnist.train(), batch_size=args.batch_size)
accuracy = fluid.average.WeightedAverage()
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
accuracy.reset()
pass_start = time.time()
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
img_data = np.array(
map(lambda x: x[0].reshape([1, 28, 28]), data)).astype(DTYPE)
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape([len(y_data), 1])
start = time.time()
outs = exe.run(
fluid.default_main_program(),
feed={"pixel": img_data,
@ -178,21 +185,36 @@ def run_benchmark(model, args):
fetch_list=[avg_cost, batch_acc, batch_size_tensor]
) # The accuracy is the accumulation of batches, but not the current batch.
accuracy.add(value=outs[1], weight=outs[2])
end = time.time()
iters += 1
num_samples += len(y_data)
loss = np.array(outs[0])
acc = np.array(outs[1])
print("pass=%d, batch=%d, loss=%f, error=%f, elapse=%f" %
(pass_id, batch_id, loss, 1 - acc, (end - start) / 1000))
train_losses.append(loss)
train_accs.append(acc)
print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %
(pass_id, iters, loss, acc))
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
train_elapsed = time.time() - start_time
examples_per_sec = num_samples / train_elapsed
pass_end = time.time()
print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
(num_samples, train_elapsed, examples_per_sec))
# evaluation
if args.with_test:
test_avg_acc = eval_test(exe, batch_acc, batch_size_tensor,
inference_program)
exit(0)
train_avg_acc = accuracy.eval()
test_avg_acc = eval_test(exe, batch_acc, batch_size_tensor,
inference_program)
print("pass=%d, train_avg_acc=%f, test_avg_acc=%f, elapse=%f" %
(pass_id, train_avg_acc, test_avg_acc,
(pass_end - pass_start) / 1000))
def print_arguments(args):
vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
vars(args)['device'] == 'GPU')
print('----------- mnist Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':

38
benchmark/fluid/resnet.py
Unescape View File

@ -87,15 +87,6 @@ def parse_args():
return args
def print_arguments(args):
vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
vars(args)['device'] == 'GPU')
print('----------- Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
def conv_bn_layer(input, ch_out, filter_size, stride, padding, act='relu'):
conv1 = fluid.layers.conv2d(
input=input,
@ -279,32 +270,31 @@ def run_benchmark(model, args):
'label': label},
fetch_list=[avg_cost, batch_acc, batch_size_tensor])
iters += 1
num_samples += label[0]
num_samples += len(label)
accuracy.add(value=acc, weight=weight)
train_losses.append(loss)
train_accs.append(acc)
print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %
(pass_id, iters, loss, acc))
pass_train_acc = accuracy.eval()
# evaluation
if args.with_test:
pass_test_acc = test(exe)
train_elapsed = time.time() - start_time
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
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))
# evaluation
if args.with_test:
pass_test_acc = test(exe)
exit(0)
if args.use_cprof:
pr.disable()
s = StringIO.StringIO()
sortby = 'cumulative'
ps = pstats.Stats(pr, stream=s).sort_stats(sortby)
ps.print_stats()
print(s.getvalue())
def print_arguments(args):
vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
vars(args)['device'] == 'GPU')
print('----------- resnet Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':

70
benchmark/fluid/run.sh
Unescape View File

@ -1,7 +1,9 @@
#!/bin/bash
# This script benchmarking the PaddlePaddle Fluid on
# single thread single GPU.
export CUDNN_PATH=/paddle/cudnn_v5/cuda/lib
#export FLAGS_fraction_of_gpu_memory_to_use=0.0
export CUDNN_PATH=/paddle/cudnn_v5
# disable openmp and mkl parallel
#https://github.com/PaddlePaddle/Paddle/issues/7199
@ -25,25 +27,79 @@ export CUDA_VISIBLE_DEVICES=0
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=$CUDNN_PATH:$LD_LIBRARY_PATH
# only query the gpu used
nohup stdbuf -oL nvidia-smi \
--id=${CUDA_VISIBLE_DEVICES} \
--query-gpu=timestamp \
--query-compute-apps=pid,process_name,used_memory \
--format=csv \
--filename=mem.log \
-l 1 &
# mnist
# mnist gpu mnist 128
FLAGS_benchmark=true stdbuf -oL python fluid/mnist.py \
--device=GPU \
--batch_size=128 \
--skip_batch_num=5 \
--iterations=500 \
2>&1 | tee -a mnist_gpu_128.log
# vgg16
# cifar10 gpu cifar10 128
FLAGS_benchmark=true python fluid/vgg.py \
# gpu cifar10 128
FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \
--device=GPU \
--batch_size=128 \
--skip_batch_num=5 \
--iterations=30 \
2>&1 > vgg16_gpu_128.log
--iterations=30 \
2>&1 | tee -a vgg16_gpu_128.log
# flowers gpu 128
FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \
--device=GPU \
--batch_size=32 \
--data_set=flowers \
--skip_batch_num=5 \
--iterations=30 \
2>&1 | tee -a vgg16_gpu_flowers_32.log
# resnet50
# resnet50 gpu cifar10 128
FLAGS_benchmark=true python fluid/resnet.py \
FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \
--device=GPU \
--batch_size=128 \
--data_set=cifar10 \
--model=resnet_cifar10 \
--skip_batch_num=5 \
--iterations=30 \
2>&1 > resnet50_gpu_128.log
2>&1 | tee -a resnet50_gpu_128.log
# resnet50 gpu flowers 64
FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \
--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
# lstm
# lstm gpu imdb 32 # tensorflow only support batch=32
FLAGS_benchmark=true stdbuf -oL python fluid/stacked_dynamic_lstm.py \
--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
# seq2seq
# seq2seq gpu wmb 128
FLAGS_benchmark=true stdbuf -oL python fluid/machine_translation.py \
--device=GPU \
--batch_size=128 \
--skip_batch_num=5 \
--iterations=30 \
2>&1 | tee -a lstm_gpu_128.log

89
benchmark/fluid/stacked_dynamic_lstm.py
Unescape View File

@ -37,6 +37,14 @@ def parse_args():
type=int,
default=32,
help='The sequence number of a batch data. (default: %(default)d)')
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(
'--emb_dim',
type=int,
@ -64,6 +72,10 @@ def parse_args():
default=int(os.environ.get('CROP_SIZE', '1500')),
help='The max sentence length of input. Since this model use plain RNN,'
' Gradient could be explored if sentence is too long')
parser.add_argument(
'--with_test',
action='store_true',
help='If set, test the testset during training.')
args = parser.parse_args()
return args
@ -157,37 +169,43 @@ def main():
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
def train_loop(pass_num, crop_size):
with profiler.profiler(args.device, 'total') as prof:
for pass_id in range(pass_num):
train_reader = batch(
paddle.reader.shuffle(
crop_sentence(imdb.train(word_dict), crop_size),
buf_size=25000),
batch_size=args.batch_size)
word_nums = 0
pass_start_time = time.time()
for batch_id, data in enumerate(train_reader()):
tensor_words = to_lodtensor([x[0] for x in data], place)
for x in data:
word_nums += len(x[0])
label = numpy.array([x[1] for x in data]).astype("int64")
label = label.reshape((-1, 1))
loss_np, acc, weight = exe.run(
fluid.default_main_program(),
feed={"words": tensor_words,
"label": label},
fetch_list=[loss, batch_acc, batch_size_tensor])
print("pass_id=%d, batch_id=%d, loss=%f, acc=%f" %
(pass_id, batch_id, loss_np, acc))
pass_end_time = time.time()
time_consumed = pass_end_time - pass_start_time
words_per_sec = word_nums / time_consumed
print("pass_id=%d, sec/pass: %f, words/s: %f" %
(pass_id, time_consumed, words_per_sec))
train_loop(args.pass_num, args.crop_size)
train_reader = batch(
paddle.reader.shuffle(
crop_sentence(imdb.train(word_dict), args.crop_size),
buf_size=25000),
batch_size=args.batch_size)
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
train_accs = []
train_losses = []
for batch_id, data in enumerate(train_reader()):
if iters == args.skip_batch_num:
start_time = time.time()
num_samples = 0
if iters == args.iterations:
break
tensor_words = to_lodtensor([x[0] for x in data], place)
label = numpy.array([x[1] for x in data]).astype("int64")
label = label.reshape((-1, 1))
loss_np, acc, weight = exe.run(
fluid.default_main_program(),
feed={"words": tensor_words,
"label": label},
fetch_list=[loss, batch_acc, batch_size_tensor])
iters += 1
for x in data:
num_samples += len(x[0])
print(
"Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
(pass_id, iters, loss_np, acc)
) # The accuracy is the accumulation of batches, but not the current batch.
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))
exit(0)
def to_lodtensor(data, place):
@ -205,5 +223,14 @@ def to_lodtensor(data, place):
return res
def print_arguments(args):
print('----------- lstm Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')
if __name__ == '__main__':
args = parse_args()
print_arguments(args)
main()

16
benchmark/fluid/vgg.py
Unescape View File

@ -191,25 +191,29 @@ def main():
fetch_list=[avg_cost, batch_acc, batch_size_tensor])
accuracy.add(value=acc, weight=weight)
iters += 1
num_samples += len(data)
num_samples += len(y_data)
print(
"Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
(pass_id, iters, loss, acc)
) # The accuracy is the accumulation of batches, but not the current batch.
pass_train_acc = accuracy.eval()
# pass_train_acc = accuracy.eval()
train_losses.append(loss)
train_accs.append(acc)
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
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))
# evaluation
if args.with_test:
pass_test_acc = test(exe)
train_elapsed = time.time() - start_time
print("Pass: %d, Loss: %f, Train Accuray: %f\n" %
(pass_id, np.mean(train_losses), np.mean(train_accs)))
exit(0)
def print_arguments():
print('----------- Configuration Arguments -----------')
print('----------- vgg Configuration Arguments -----------')
for arg, value in sorted(vars(args).iteritems()):
print('%s: %s' % (arg, value))
print('------------------------------------------------')

12
cmake/generic.cmake
Unescape View File

@ -244,11 +244,11 @@ function(cc_test TARGET_NAME)
cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_executable(${TARGET_NAME} ${cc_test_SRCS})
# Support linking flags: --whole-archive (Linux) / -force_load (MacOS)
target_circle_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
target_circle_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
if("${cc_test_DEPS}" MATCHES "ARCHIVE_START")
list(REMOVE_ITEM cc_test_DEPS ARCHIVE_START ARCHIVE_END)
endif()
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
add_test(NAME ${TARGET_NAME}
COMMAND ${TARGET_NAME} ${cc_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
@ -311,8 +311,8 @@ function(nv_test TARGET_NAME)
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog)
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
add_test(${TARGET_NAME} ${TARGET_NAME})
endif()
endfunction(nv_test)
@ -387,8 +387,8 @@ function(hip_test TARGET_NAME)
endif()
add_executable(${TARGET_NAME} ${_cmake_options} ${_generated_files} ${_sources})
set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags)
target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
add_test(${TARGET_NAME} ${TARGET_NAME})
endif()
endfunction(hip_test)

87
doc/design/file_manager/README.md
Unescape View File

@ -1,87 +0,0 @@
# FileManager设计文档
## 目标
在本文档中，我们设计说明了名为FileManager系统，方便用户上传自己的训练数据以进行分布式训练
主要功能包括：
- 提供常用的命令行管理命令管理文件和目录
- 支持大文件的断点上传、下载
## 名词解释
- PFS：是`Paddlepaddle cloud File System`的缩写，是对用户文件存储空间的抽象，与之相对的是local filesystem。目前我们用CephFS来搭建。
- [CephFS](http://docs.ceph.com/docs/master/cephfs/)：一个POSIX兼容的文件系统。
- Chunk：逻辑划上文件分块的单位。
## 模块
### 架构图
<image src=./src/filemanager.png width=900>
### PFSClient
- 功能： 详细设计[link](./pfs/pfsclient.md)
- 提供用户管理文件的命令
- 需要可以跨平台执行
- 双向验证
PFSClient需要和Ingress之间做双向验证<sup>[tls](#tls)</sup>，所以用户需要首先在`cloud.paddlepaddle.org`上注册一下，申请用户空间，并且把系统生成的CA(certificate authority)、Key、CRT(CA signed certificate)下载到本地，然后才能使用PFSClient。
### [Ingress](https://kubernetes.io/docs/concepts/services-networking/ingress/)
- 功能：
提供七层协议的反向代理、基于粘性会话的负载均衡功能。
- 透传用户身份的办法
Ingress需要把PFSClient的身份信息传给PFSServer，配置的方法参考[link](http://www.integralist.co.uk/posts/clientcertauth.html#3)
### PFSServer
PFSServer提供RESTful API接口，接收处理PFSClient端的文件管理请求，并且把结果返回PFSClient端。
RESTful API
- /api/v1/files
- `GET /api/v1/files`: Get metadata of files or directories.
- `POST /api/v1/files`: Create files or directories.
- `PATCH /api/v1/files`: Update files or directories.
- `DELETE /api/v1/files`: Delete files or directories.
- /api/v1/file/chunks
- `GET /api/v1/storage/file/chunks`: Get chunks's metadata of a file.
- /api/v1/storage/files
- `GET /api/v1/storage/files`: Download files or directories.
- `POST /api/v1/storage/files`: Upload files or directories.
- /api/v1/storage/file/chunks
- `GET /api/v1/storage/file/chunks`: Download chunks's data.
- `POST /api/v1/storage/file/chunks`: Upload chunks's data.
## 文件传输优化
### 分块文件传输
用户文件可能是比较大的，上传到Cloud或者下载到本地的时间可能比较长，而且在传输的过程中也可能出现网络不稳定的情况。为了应对以上的问题，我们提出了Chunk的概念，一个Chunk由所在的文件偏移、数据、数据长度及校验值组成。文件的上传和下载都是通过对Chunk的操作来实现的。由于Chunk比较小（默认256K），完成一个传输动作完成的时间也比较短，不容易出错。PFSClient需要在传输完毕最后一个Chunk的时候检查destination文件的MD5值是否和source文件一致。
一个典型的Chunk如下所示：
```
type Chunk struct {
fileOffset int64
checksum uint32
len uint32
data []byte
}
```
### 生成sparse文件
当destination文件不存在或者大小和source文件不一致时，可以用[Fallocate](https://Go.org/pkg/syscall/#Fallocate)生成sparse文件，然后就可以并发写入多个Chunk。
### 覆盖不一致的部分
文件传输的的关键在于需要PFSClient端对比source和destination的文件Chunks的checksum是否保持一致，不一致的由PFSClient下载或者传输Chunk完成。这样已经传输成功的部分就不用重新传输了。
## 用户使用流程
参考[link](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/cluster_train/data_dispatch.md)
## 框架生成
用[swagger](https://github.com/swagger-api/swagger-codegen)生成PFSClient和PFSServer的框架部分，以便我们可以把更多的精力放到逻辑本身上。
## 参考文档
- <a name=tls></a>[TLS complete guide](https://github.com/k8sp/tls/blob/master/tls.md)
- [aws.s3](http://docs.aws.amazon.com/cli/latest/reference/s3/)
- [linux man document](https://linux.die.net/man/)

129
doc/design/file_manager/pfs/pfsclient.md
Unescape View File

@ -1,129 +0,0 @@
# PFSClient
## Description
The `pfs` command is a Command Line Interface to manage your files on PaddlePaddle Cloud
## Synopsis
```
paddle [options] pfs <subcommand> [parameters]
```
## Options
```
--profile (string)
Use a specific profile from your credential file.
--help (string)
Display more information about command
--version
Output version information and exit
--debug
Show detailed debugging log
--only-show-errors (boolean)
Only errors and warnings are displayed. All other output is suppressed.
```
## Path Arguments
When using a command, we need to specify path arguments. There are two path argument type: `localpath` and `pfspath`.
A `pfspath` begin with `/pfs`, eg: `/pfs/$DATACENTER/home/$USER/folder`.
[Here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/cluster_train/data_dispatch.md#上传训练文件) is how to config datacenters.
## order of Path Arguments
Commonly, if there are two path arguments, the first is the source, and the second is the destination.
## Subcommonds
- rm - remove files or directories
```
Synopsis:
rm [-r] [-v] <PFSPath> ...
Options:
-r
Remove directories and their contents recursively
-v
Cause rm to be verbose, showing files after they are removed.
Examples:
paddle pfs rm /pfs/$DATACENTER/home/$USER/file
paddle pfs rm -r /pfs/$DATACENTER/home/$USER/folder
```
- mv - move (rename) files
```
Synopsis:
mv [-f | -n] [-v] <LocalPath> <PFSPath>
mv [-f | -n] [-v] <LocalPath> ... <PFSPath>
mv [-f | -n] [-v] <PFSPath> <LocalPath>
mv [-f | -n] [-v] <PFSPath> ... <LocalPath>
mv [-f | -n] [-v] <PFSPath> <PFSPath>
mv [-f | -n] [-v] <PFSPath> ... <PFSPath>
Options:
-f
Do not prompt for confirmation before overwriting the destination path. (The -f option overrides previous -n options.)
-n
Do not overwrite an existing file. (The -n option overrides previous -f options.)
-v
Cause mv to be verbose, showing files after they are moved.
Examples:
paddle pfs mv ./text1.txt /pfs/$DATACENTER/home/$USER/text1.txt
```
- cp - copy files or directories
```
Synopsis:
cp [-r] [-f | -n] [-v] [--preserve--links] <LocalPath> <PFSPath>
cp [-r] [-f | -n] [-v] [--preserve--links] <LocalPath> ... <PFSPath>
cp [-r] [-f | -n] [-v] [--preserve--links] <PFSPath> <LocalPath>
cp [-r] [-f | -n] [-v] [--preserve--links] <PFSPath> ... <LocalPath>
cp [-r] [-f | -n] [-v] [--preserve--links] <PFSPath> <PFSPath>
cp [-r] [-f | -n] [-v] [--preserve--links] <PFSPath> ... <PFSPath>
Options:
-r
Copy directories recursively
-f
Do not prompt for confirmation before overwriting the destination path. (The -f option overrides previous -n options.)
-n
Do not overwrite an existing file. (The -n option overrides previous -f options.)
-v
Cause cp to be verbose, showing files after they are copied.
--preserve--links
Reserve links when copy links
Examples:
paddle pfs cp ./file /pfs/$DATACENTER/home/$USER/file
paddle pfs cp /pfs/$DATACENTER/home/$USER/file ./file
```
- ls- list files
```
Synopsis:
ls [-r] <PFSPath> ...
Options:
-R
List directory(ies) recursively
Examples:
paddle pfs ls /pfs/$DATACENTER/home/$USER/file
paddle pfs ls /pfs/$DATACENTER/home/$USER/folder
```
- mkdir - mkdir directory(ies)
Create intermediate directory(ies) as required.
```
Synopsis:
mkdir <PFSPath> ...
Examples:
paddle pfs mkdir /pfs/$DATACENTER/home/$USER/folder
```

BIN
doc/design/file_manager/src/filemanager.graffle
View File

Binary file not shown.

BIN
doc/design/file_manager/src/filemanager.png
View File

Binary file not shown.
Side by Side

Before

Width:  |  Height:  |  Size: 142 KiB

2
doc/fluid/dev/index_cn.rst
Unescape View File

@ -9,5 +9,5 @@
use_eigen_cn.md
name_convention.md
support_new_device.md
releasing_process.md
releasing_process_cn.md
op_markdown_format.md

2
doc/fluid/dev/index_en.rst
Unescape View File

@ -9,5 +9,5 @@ Development
use_eigen_en.md
name_convention.md
support_new_device.md
releasing_process.md
releasing_process_en.md
op_markdown_format.md

54
doc/fluid/dev/releasing_process.md → doc/fluid/dev/releasing_process_cn.md
Unescape View File

@ -10,19 +10,10 @@ PaddlePaddle每次发新的版本，遵循以下流程:
* 使用Regression Test List作为检查列表，测试本次release的正确性。
* 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，到第二步
* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
* 编译这个版本的python wheel包，并发布到pypi。
* 由于pypi.python.org目前遵循[严格的命名规范PEP 513](https://www.python.org/dev/peps/pep-0513)，在使用twine上传之前，需要重命名wheel包中platform相关的后缀，比如将`linux_x86_64`修改成`manylinux1_x86_64`。
* pypi上的package名称为paddlepaddle和paddlepaddle_gpu，如果要上传GPU版本的包，需要修改build/python/setup.py中，name: "paddlepaddle_gpu"并重新打包wheel包：`python setup.py bdist_wheel`。
* 上传方法：
```
cd build/python
pip install twine
twine upload dist/[package to upload]
```
* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
1. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
1. 协同完成Release Note的书写
* 将这个版本的python wheel包发布到pypi。
* 更新Docker镜像（参考后面的操作细节）。
1. 第三步完成后，将`release/版本号`分支合入master分支，将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。
1. 协同完成Release Note的书写。
需要注意的是:
@ -31,13 +22,18 @@ PaddlePaddle每次发新的版本，遵循以下流程:
## 发布wheel包到pypi
使用[PaddlePaddle CI](https://paddleci.ngrok.io/project.html?projectId=Manylinux1&tab=projectOverview)
1. 使用[PaddlePaddle CI](https://paddleci.ngrok.io/project.html?projectId=Manylinux1&tab=projectOverview)
完成自动化二进制编译，参考下图，选择需要发布的版本（通常包含一个CPU版本和一个GPU版本），点击"run"右侧的"..."按钮，可以
弹出下面的选择框，在第二个tab (Changes)里选择需要发布的分支，这里选择0.11.0，然后点击"Run Build"按钮。等待编译完成后
可以在此页面的"Artifacts"下拉框中找到生成的3个二进制文件，分别对应CAPI，`cp27m`和`cp27mu`的版本。然后按照上述的方法
使用`twine`工具上传即可。
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/ci_build_whl.png">
弹出下面的选择框，在第二个tab (Changes)里选择需要发布的分支，这里选择0.11.0，然后点击"Run Build"按钮。
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/ci_build_whl.png">
1. 等待编译完成后可以在此页面的"Artifacts"下拉框中找到生成的3个二进制文件，分别对应CAPI，`cp27m`和`cp27mu`的版本。
1. 由于pypi.python.org目前遵循[严格的命名规范PEP 513](https://www.python.org/dev/peps/pep-0513)，在使用twine上传之前，需要重命名wheel包中platform相关的后缀，比如将`linux_x86_64`修改成`manylinux1_x86_64`。
1. 上传：
```
cd build/python
pip install twine
twine upload dist/[package to upload]
```
* 注：CI环境使用 https://github.com/PaddlePaddle/buildtools 这里的DockerImage作为编译环境以支持更多的Linux
发型版，如果需要手动编译，也可以使用这些镜像。这些镜像也可以从 https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/ 下载得到。
@ -48,10 +44,20 @@ PaddlePaddle每次发新的版本，遵循以下流程:
上述PaddlePaddle CI编译wheel完成后会自动将Docker镜像push到DockerHub，所以，发布Docker镜像只需要对自动push的镜像打上
版本号对应的tag即可：
1. 进入 https://hub.docker.com/r/paddlepaddle/paddle/tags/ 查看latest tag的更新时间是否在上述编译wheel包完成后是否最新。
1. 执行 `docker pull paddlepaddle/paddle:[latest tag]`，latest tag可以是latest或latest-gpu等。
1. 执行 `docker tag paddlepaddle/paddle:[latest tag] paddlepaddle/paddle:[version]`
1. 执行 `docker push paddlepaddle/paddle:[version]`
```
docker pull [镜像]:latest
docker tag [镜像]:latest [镜像]:[version]
docker push [镜像]:[version]
```
需要更新的镜像tag包括：
* `[version]`: CPU版本
* `[version]-openblas`: openblas版本
* `[version]-gpu`: GPU版本（CUDA 8.0 cudnn 5）
* `[version]-gpu-[cudaver]-[cudnnver]`: 不同cuda, cudnn版本的镜像
之后可进入 https://hub.docker.com/r/paddlepaddle/paddle/tags/ 查看是否发布成功。
## PaddlePaddle 分支规范
@ -76,7 +82,7 @@ PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-
### PaddlePaddle Book中所有章节
PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。
PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练（V2和Fluid）模型正确性。
<table>
<thead>

210
doc/fluid/dev/releasing_process_en.md
Unescape View File

@ -0,0 +1,210 @@
# PaddlePaddle Releasing Process
PaddlePaddle manages its branches using "git-flow branching model", and [Semantic Versioning](http://semver.org/) as it's version number semantics.
Each time we release a new PaddlePaddle version, we should follow the below steps:
1. Fork a new branch from `develop` named `release/[version]`, e.g. `release/0.10.0`.
1. Push a new tag on the release branch, the tag name should be like `[version]rc.patch`. The
first tag should be `0.10.0rc1`, and the second should be `0.10.0.rc2` and so on.
1. After that, we should do:
* Run all regression test on the Regression Test List (see PaddlePaddle TeamCity CI), to confirm
that this release has no major bugs.
* If regression test fails, we must fix those bugs and create a new `release/[version]`
branch from previous release branch.
* Modify `python/setup.py.in`, change the version number and change `ISTAGED` to `True`.
* Publish PaddlePaddle release wheel packages to pypi (see below instructions for detail).
* Update the Docker images (see below instructions for detail).
1. After above step, merge `release/[version]` branch to master and push a tag on the master commit,
then merge `master` to `develop`.
1. Update the Release Note.
***NOTE:***
* Do ***NOT*** merge commits from develop branch to release branches to keep the release branch contain
features only for current release, so that we can test on that version.
* If we want to fix bugs on release branches, we must merge the fix to master, develop and release branch.
## Publish Wheel Packages to pypi
1. Use our [CI tool](https://paddleci.ngrok.io/project.html?projectId=Manylinux1&tab=projectOverview)
to build all wheel packages needed to publish. As shown in the following picture, choose a build
version, click "..." button on the right side of "Run" button, and switch to the second tab in the
pop-up box, choose the current release branch and click "Run Build" button. You may repeat this
step to start different versions of builds.
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/ci_build_whl.png">
1. After the build succeeds, download the outputs under "Artifacts" including capi, `cp27m` and `cp27mu`.
1. Since pypi.python.org follows [PEP 513](https://www.python.org/dev/peps/pep-0513), before we
upload the package using `twine`, we need to rename the package from `linux_x86_64` to
`manylinux1_x86_64`.
1. Start the upload:
```
cd build/python
pip install twine
twine upload dist/[package to upload]
```
* NOTE: We use a special Docker image to build our releases to support more Linux distributions, you can
download it from https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/, or build it using
scripts under `tools/manylinux1`.
* pypi does not allow overwrite the already uploaded version of wheel package, even if you delete the
old version. you must change the version number before upload a new one.
## Publish Docker Images
Our CI tool will push latest images to DockerHub, so we only need to push a version tag like:
```
docker pull [image]:latest
docker tag [image]:latest [image]:[version]
docker push [image]:[version]
```
Tags that need to be updated are:
* `[version]`: CPU only version image
* `[version]-openblas`: openblas version image
* `[version]-gpu`: GPU version（using CUDA 8.0 cudnn 5）
* `[version]-gpu-[cudaver]-[cudnnver]`: tag for different cuda, cudnn versions
You can then checkout the latest pushed tags at https://hub.docker.com/r/paddlepaddle/paddle/tags/.
## Branching Model
We use [git-flow](http://nvie.com/posts/a-successful-git-branching-model/) as our branching model,
with some modifications:
* `master` branch is the stable branch. Each version on the master branch is tested and guaranteed.
* `develop` branch is for development. Each commit on develop branch has passed CI unit test, but no
regression tests are run.
* `release/[version]` branch is used to publish each release. Latest release version branches have
bugfix only for that version, but no feature updates.
* Developer forks are not required to follow
[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)
branching model, all forks is like a feature branch.
* Advise: developer fork's develop branch is used to sync up with main repo's develop branch.
* Advise: developer use it's fork's develop branch to for new branch to start developing.
* Use that branch on developer's fork to create pull requests and start reviews.
* developer can push new commits to that branch when the pull request is open.
* Bug fixes are also started from developers forked repo. And, bug fixes branch can merge to
`master`, `develop` and `releases`.
## PaddlePaddle Regression Test List
### All Chapters of PaddlePaddle Book
We need to guarantee that all the chapters of PaddlePaddle Book can run correctly. Including
V1 (`paddle_trainer` training) and V2 training and Fluid training.
<table>
<thead>
<tr>
<th></th>
<th>Linear Regression</th>
<th>Recognize Digits</th>
<th>Image Classification</th>
<th>Word2Vec</th>
<th>Personalized Recommendation</th>
<th>Sentiment Analysis</th>
<th>Semantic Role Labeling</th>
<th>Machine Translation</th>
</tr>
</thead>
<tbody>
<tr>
<td>API.V2 + Docker + GPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> API.V2 + Docker + CPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>`paddle_trainer` + Docker + GPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>`paddle_trainer` + Docker + CPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> API.V2 + Ubuntu + GPU</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>API.V2 + Ubuntu + CPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> `paddle_trainer` + Ubuntu + GPU</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> `paddle_trainer` + Ubuntu + CPU</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
</tbody>
</table>

1
paddle/.gitignore vendored
Unescape View File

@ -1,3 +1,4 @@
.timestamp
*.o
*.a
.svn

8
paddle/fluid/framework/CMakeLists.txt
Unescape View File

@ -7,9 +7,9 @@ cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
nv_test(dim_test SRCS dim_test.cu DEPS ddim)
if(WITH_GPU)
nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place paddle_memory device_context framework_proto)
nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place memory device_context framework_proto)
else()
cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place paddle_memory device_context framework_proto)
cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place memory device_context framework_proto)
endif()
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@ -21,9 +21,9 @@ endif()
cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place paddle_memory device_context init)
nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context init)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor paddle_memory)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor init)
cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)

72
paddle/fluid/framework/executor.cc
Unescape View File

@ -93,6 +93,43 @@ static void CheckTensorNANOrInf(const std::string& name,
"Tensor %s contains NAN", name);
}
void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
int block_id) {
auto& global_block = pdesc.Block(block_id);
const Scope* ancestor_scope = scope;
while (ancestor_scope->parent()) {
ancestor_scope = ancestor_scope->parent();
}
if (ancestor_scope != scope) {
for (auto& var : global_block.AllVars()) {
if (var->Name() == framework::kEmptyVarName) {
continue;
}
if (var->Persistable()) {
auto* ptr = const_cast<Scope*>(ancestor_scope)->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " global, which pointer is " << ptr;
} else {
auto* ptr = scope->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " locally, which pointer is " << ptr;
}
}
} else {
for (auto& var : global_block.AllVars()) {
auto* ptr = scope->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
<< ptr;
}
}
}
void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
bool create_local_scope, bool create_vars) {
platform::RecordBlock b(block_id);
@ -184,8 +221,8 @@ static bool has_fetch_operators(
void Executor::Run(const ProgramDesc& program, Scope* scope,
std::map<std::string, const LoDTensor*>& feed_targets,
std::map<std::string, LoDTensor*>& fetch_targets,
const std::string& feed_holder_name,
const std::string& fetch_holder_name, bool create_vars) {
bool create_vars, const std::string& feed_holder_name,
const std::string& fetch_holder_name) {
platform::RecordBlock b(kProgramId);
bool has_feed_ops =
has_feed_operators(program.Block(0), feed_targets, feed_holder_name);
@ -296,38 +333,13 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
bool create_local_scope, bool create_vars) {
auto& block = ctx->prog_.Block(ctx->block_id_);
Scope* local_scope = scope;
if (create_vars) {
if (create_local_scope) {
local_scope = &scope->NewScope();
for (auto& var : block.AllVars()) {
if (var->Name() == framework::kEmptyVarName) {
continue;
}
if (var->Persistable()) {
auto* ptr = scope->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " global, which pointer is " << ptr;
} else {
auto* ptr = local_scope->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create Variable " << var->Name()
<< " locally, which pointer is " << ptr;
}
}
} else {
for (auto& var : block.AllVars()) {
auto* ptr = local_scope->Var(var->Name());
InitializeVariable(ptr, var->GetType());
VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
<< ptr;
}
} // if (create_local_scope)
} // if (create_vars)
}
CreateVariables(ctx->prog_, local_scope, ctx->block_id_);
}
for (auto& op : ctx->ops_) {
VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);

6
paddle/fluid/framework/executor.h
Unescape View File

@ -54,9 +54,9 @@ class Executor {
void Run(const ProgramDesc& program, Scope* scope,
std::map<std::string, const LoDTensor*>& feed_targets,
std::map<std::string, LoDTensor*>& fetch_targets,
bool create_vars = true,
const std::string& feed_holder_name = "feed",
const std::string& fetch_holder_name = "fetch",
bool create_vars = true);
const std::string& fetch_holder_name = "fetch");
static std::unique_ptr<ExecutorPrepareContext> Prepare(
const ProgramDesc& program, int block_id);
@ -64,6 +64,8 @@ class Executor {
static std::vector<std::shared_ptr<ExecutorPrepareContext>> Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids);
void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id);
void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
bool create_local_scope = true,
bool create_vars = true);

103
paddle/fluid/framework/parallel_executor.cc
Unescape View File

@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/parallel_executor.h"
#include "paddle/fluid/platform/profiler.h"
#include <string>
#include <vector>
@ -24,6 +23,7 @@ limitations under the License. */
#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
#include "paddle/fluid/platform/profiler.h"
namespace paddle {
namespace framework {
@ -43,30 +43,40 @@ class ParallelExecutorPrivate {
#endif
};
std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
return member_->local_scopes_;
}
ParallelExecutor::ParallelExecutor(
size_t num_threads, bool use_event,
const std::vector<platform::Place> &places,
const std::unordered_set<std::string> &params,
const ProgramDesc &startup_program, const ProgramDesc &main_program,
const std::string &loss_var_name, Scope *scope, bool allow_op_delay)
const std::unordered_set<std::string> &bcast_vars,
const ProgramDesc &main_program, const std::string &loss_var_name,
Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay)
: member_(new ParallelExecutorPrivate(places)) {
member_->global_scope_ = scope;
// Step 1. RunStartupProgram and Bcast the params to devs.
Executor exe(places[0]);
exe.Run(startup_program, scope, 0);
// Step 1. Bcast the params to devs.
// Create local scopes
for (size_t i = 0; i < member_->places_.size(); ++i) {
member_->local_scopes_.push_back(&scope->NewScope());
if (local_scopes.empty()) {
for (size_t i = 0; i < member_->places_.size(); ++i) {
member_->local_scopes_.push_back(&scope->NewScope());
}
} else {
PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size());
for (size_t i = 0; i < member_->places_.size(); ++i) {
member_->local_scopes_.push_back(local_scopes[i]);
}
}
// Bcast Parameters to all GPUs
#ifdef PADDLE_WITH_CUDA
member_->nccl_ctxs_.reset(new platform::NCCLContextMap(member_->places_));
#endif
if (platform::is_gpu_place(places[0]) &&
member_->local_scopes_.size() != 1) { // Is CUDA
BCastParamsToGPUs(startup_program);
if (platform::is_gpu_place(places[0]) && member_->local_scopes_.size() != 1 &&
local_scopes.empty()) { // Is CUDA
BCastParamsToGPUs(bcast_vars);
}
// Startup Program has been run. All local scopes has correct parameters.
@ -99,48 +109,47 @@ ParallelExecutor::ParallelExecutor(
}
void ParallelExecutor::BCastParamsToGPUs(
const ProgramDesc &startup_program) const {
const std::unordered_set<std::string> &vars) const {
#ifdef PADDLE_WITH_CUDA
auto *main_scope = member_->local_scopes_[0];
for (auto *var_desc : startup_program.Block(0).AllVars()) {
size_t idx = var_desc->Name().find("@GRAD");
if (idx != std::string::npos) continue;
if (var_desc->GetType() == proto::VarType::LOD_TENSOR) {
auto &main_tensor =
main_scope->FindVar(var_desc->Name())->Get<LoDTensor>();
auto &dims = main_tensor.dims();
if (paddle::platform::is_gpu_place(main_tensor.place())) {
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;
if (i == 0) {
buffer = const_cast<void *>(main_tensor.data<void>());
} else {
auto local_scope = member_->local_scopes_[i];
auto *t =
local_scope->Var(var_desc->Name())->GetMutable<LoDTensor>();
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());
}
} else {
platform::CPUPlace cpu;
for (size_t i = 1; i < member_->places_.size(); ++i) {
for (auto &var : vars) {
auto *main_var = main_scope->FindVar(var);
if (!main_var->IsType<LoDTensor>()) {
continue;
}
auto &main_tensor = main_var->Get<LoDTensor>();
auto &dims = main_tensor.dims();
if (paddle::platform::is_gpu_place(main_tensor.place())) {
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;
if (i == 0) {
buffer = const_cast<void *>(main_tensor.data<void>());
} else {
auto local_scope = member_->local_scopes_[i];
auto *t = local_scope->Var(var_desc->Name())->GetMutable<LoDTensor>();
auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
t->Resize(dims);
t->mutable_data(cpu, main_tensor.type());
paddle::framework::TensorCopy(main_tensor, cpu, t);
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());
}
} else {
platform::CPUPlace cpu;
for (size_t i = 1; i < member_->places_.size(); ++i) {
auto local_scope = member_->local_scopes_[i];
auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
t->Resize(dims);
t->mutable_data(cpu, main_tensor.type());
paddle::framework::TensorCopy(main_tensor, cpu, t);
}
}
member_->nccl_ctxs_->WaitAll();

7
paddle/fluid/framework/parallel_executor.h
Unescape View File

@ -36,11 +36,14 @@ class ParallelExecutor {
explicit ParallelExecutor(size_t num_threads, bool use_event,
const std::vector<platform::Place>& places,
const std::unordered_set<std::string>& params,
const ProgramDesc& startup_program,
const std::unordered_set<std::string>& bcast_vars,
const ProgramDesc& main_program,
const std::string& loss_var_name, Scope* scope,
const std::vector<Scope*>& local_scopes,
bool allow_op_delay);
std::vector<Scope*>& GetLocalScopes();
void Run(const std::vector<std::string>& fetch_tensors,
const std::string& fetched_var_name,
const std::unordered_map<std::string, LoDTensor>& feed_tensors);
@ -51,7 +54,7 @@ class ParallelExecutor {
ParallelExecutorPrivate* member_;
void BCastParamsToGPUs(const ProgramDesc& startup_program) const;
void BCastParamsToGPUs(const std::unordered_set<std::string>& vars) const;
};
} // namespace framework

5
paddle/fluid/framework/scope.cc
Unescape View File

@ -15,7 +15,6 @@ limitations under the License. */
#include "paddle/fluid/framework/scope.h"
#include <memory> // for unique_ptr
#include <mutex> // for call_once
#include <set>
#include "glog/logging.h"
#include "paddle/fluid/framework/threadpool.h"
@ -39,6 +38,7 @@ Scope::~Scope() {
}
Scope& Scope::NewScope() const {
std::unique_lock<std::mutex> lock(mutex_);
kids_.push_back(new Scope(this));
return *kids_.back();
}
@ -92,6 +92,7 @@ std::vector<std::string> Scope::LocalVarNames() const {
}
void Scope::DeleteScope(Scope* scope) {
std::unique_lock<std::mutex> lock(mutex_);
auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
this->kids_.erase(it);
@ -103,7 +104,7 @@ void Scope::DeleteScope(Scope* scope) {
}
}
void Scope::EraseVars(std::vector<std::string>& var_names) {
void Scope::EraseVars(const std::vector<std::string>& var_names) {
std::set<std::string> 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()) {

8
paddle/fluid/framework/scope.h
Unescape View File

@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <list>
#include <mutex> // NOLINT
#include <string>
#include <unordered_map>
#include <vector>
@ -51,13 +52,13 @@ class Scope {
/// Create a variable with a scope-unique name.
Variable* Var(std::string* name = nullptr);
void EraseVars(std::vector<std::string>& var_names);
void EraseVars(const std::vector<std::string>& var_names);
/// Find a variable in the scope or any of its ancestors. Returns
/// nullptr if cannot find.
Variable* FindVar(const std::string& name) const;
const Scope& parent() const { return *parent_; }
const Scope* parent() const { return parent_; }
/// Find the scope or an ancestor scope that contains the given variable.
const Scope* FindScope(const Variable* var) const;
@ -88,6 +89,9 @@ class Scope {
Scope const* parent_{nullptr};
DISABLE_COPY_AND_ASSIGN(Scope);
private:
mutable std::mutex mutex_;
};
} // namespace framework
} // namespace paddle

2
paddle/fluid/inference/CMakeLists.txt
Unescape View File

@ -1,4 +1,4 @@
set(FLUID_CORE_MODULES proto_desc paddle_memory lod_tensor executor prune init)
set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor prune init)
cc_library(paddle_fluid_api
SRCS io.cc

8
paddle/fluid/inference/tests/book/test_inference_image_classification.cc
Unescape View File

@ -46,8 +46,8 @@ TEST(inference, image_classification) {
// Run inference on CPU
LOG(INFO) << "--- CPU Runs: ---";
TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds, cpu_fetchs1,
FLAGS_repeat);
TestInference<paddle::platform::CPUPlace, false>(dirname, cpu_feeds,
cpu_fetchs1, FLAGS_repeat);
LOG(INFO) << output1.dims();
#ifdef PADDLE_WITH_CUDA
@ -57,8 +57,8 @@ TEST(inference, image_classification) {
// Run inference on CUDA GPU
LOG(INFO) << "--- GPU Runs: ---";
TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds, cpu_fetchs2,
FLAGS_repeat);
TestInference<paddle::platform::CUDAPlace, false>(dirname, cpu_feeds,
cpu_fetchs2, FLAGS_repeat);
LOG(INFO) << output2.dims();
CheckError<float>(output1, output2);

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save