* Refine the cache of program, context and scope in executor.
test=develop
* Refine the unittest test_executor_and_use_program_cache.
* Add the test the PaddingRNN with use_program_cache=True.
test=develop
* Remove a check.
test=develop
* Refine the unittest to check whether it is correct when setting use_program_cache=True.
test=develop
* Refine the InferShape of ReadFrom and WriteTo op, and add comment to explain why not call ShareLoD for runtime.
test=develop
* Add comment for ReorderLoDTensorByRank op.
* Add comment for lod_tensor_to_tensor_array op to explain why only call DecreaseLoDLevel for compile time.
test=develop
* ShrinkRNNMemory op should call ShareLoD for compile time.
test=develop
* Add the implementation of IncreaseLoDLevel and add the compile-time check of lod_level in InferShape of sequence_pool.
test=develop
* Refine the unittest of DynamicRNN.
test=develop
* Change PADDLE_ENFORCE to PADDLE_ENFORCE_NE.
test=develop
* Add fusion_group_pass and elementwise pattern.
* Rewrite the detector of elementwise group.
test=develop
* Add a comment in codegen.
* Add more unittest cases.
test=develop
* Move code_generator related code to fusion_group directory.
* Correct the including path.
* Add the definition of SubGraph and finish the insert of fusion_group op in pass.
* Insert graph_vis_pass in tester to visualize the graph for debug.
* replace part of the old implementation, test=develop
* restore concat op, test=develop
* update all ops implemention & delete GetDataTypeOfVar func, test=develop
* no longer need to define all embedding layers (no one less) of all slots in each program. make trainer_param repeated in ps.proto.
* add find_distributed_lookup_table_grads instead of hard code GRAD
* support embedding stop gradient. push sparse has error before fix this.*
* fix fill sparse, skip slots which do not have embedding. each slot's embedding in a sparse table should be used in all training programs before fix this.
* fix pull sparse, skip slots which do not have embedding.
* fix collect feasign label info, skip slots which do not have embedding.
* support when there are multi sparse tables in one or multi training programs, each program can pull/push its own related sparse tables instead of all sparse tables.
* test=develop
* - Flushing mkl-dnn cache
test=develop
- Disabled clearing cache for LoadModel
- Added clearing of mkl-dnn cache when Executor is created
test=develop
- Do not clear for GPU places
test=develop
- compilation fix
test=develop
* - Moved clearing of mkl-dnn cache in destructor of executor
test=develop
* - Compilation fix
test=develop
- Reverted conditional clearing of mkl-dnn cache in Executors's
destructor
test=develop
- compilation fix
* How to write custom op needs to follow framework OP spec.
* Package fluid_framework.so and headers into whl.
* Add paddle.sysconfig.get_include() and paddle.sysconfig.get_lib() to get include dir and lib dir.
* Export some C-APIs to merge OpInfo between core.so and custom_op.so.
* Add unit testing.
* Update API.spec.
* Follow Wangzhen's comment in PR 18970, test=develop
* Review comments, test=develop
* Leave fake quantization around mul
test=develop
* Replace Fake with Real Quantized Mul
test=develop
* Fix bug in quantize placement pass
Nodes in the graph now have checked type instead of node name when they are to be marked for quantization test=develop
The new "fluid.data" changes old "fluid.layers.data":
1. Add shape and dtype check.
2. Remove "append_batch_size" parameter. We won't offer this in the new data layer because other deep learning platforms don't have this kind of data layer pre-processing. It may confuse users.
3. Remove "stop gradient" parameter because the data layer doesn't do back-propagation
TODO：
Now data layer feeded by executor is checked, will we want to check the feed data of readers in the future?
* support change shuffle thread num
* support change train thread num
* fix receive shuffle data of each channel
* data norm stop gradient
* add check thread_tensor type and root_tensor type when merge metric
* remove sleep in shuffle, add config
* add config of pslib client to client communication
* fix xbox str
* add data norm op testcase
* add flush in trainer finalize
* Set states of recurrent op as dependent vars in prune of save inference model
This PR will fix the save/load inference model problem of RNN models.
The reason of the bug is that save_inferenc_model will prune OPs that doesn't contribute to Output. But in recurrent_op, States are not Output, OPs refers States will be pruned.
This fix adds States of recurrent_op as dependent var so that OPs referring States won't be pruned.
* Add fc_elementwise_layernorm_fuse pass and unittest.
* Add fused_fc_elementwise_layernorm op and its GPU kernel.
test=develop
* Apply fc_elementwise_layernorm_fuse_pass to GPU inference.
* Add the setting of attrs in the definition of binary_op.
test=develop
* Add comment.
* Implement the unittest.
test=develop
* Change the unittest name of layer_norm.
test=develop
* refactor dygraph,test=develop
* fix failed unittest,test=develop
* polish code,test=develop
* check windows ci error,test=develop
try to fix windows ci error by np.allclose,test=develop
* polish vlog and profiler, test=develop
* try to fix preceding ops order,test=develop
* test transformer in windows ci, test=develop
* use python c-api to speed up tracer.trace,test=develop
* test=develop, fix docker with paddle nccl problem
* test=develop, add ut for debug string and gradient_accumulator
* test=develop, add tests for layer/gradient_accumulator/prepared_op
* test=develop, fix complie error for test_prepared_op
* test=develop, add more ut for dygraph
* test=develop, create API.spec for dygraph api change
* add transform_data to dygraph
* test=develop, refoctor name to make it easier to understand
* test=develop, refoctor name to make it easier to understand
* add test and change input to const ref for safety
* test=develop, fix multi-gpu failed problem , add Tracer tests, change PADDLEENFORCE to PADDLEENFORCE_EQ
* add ut for data transform
* refine ut for data_transform
* test=develop, fix ut failed on parallel se-resnext
* test=develop, change one more PADDLE_ENFORCE
* add test_tracer on multiple devices
* test=develop, change place to mutable for data transform
* test=develop, add transform data on same place test and remove useless log
* test=develop, Add to do for data layout and and ut for conv2d with no bias
* Refine the codes related to fc op.
* Add GPU implementation for fc functor.
* Apply fc_fuse_pass in GPU inference.
test=develop
* Change the cmake for fc op.
* Change PADDLE_ENFORCE to PADDLE_ENFORCE_EQ.
* Add an attribute to set the activation type in fc_op.
* Enhance the unittest of fc_op.
test=develop
* Remove the declaration of FCOpGrad back to the header file.
test=develop
* Set default value for newly added arguments in test_fc_op.
test=develop
* Enhance fc_fuse_pass to enable fusing relu.
* Allow print the shapes of var_desc in graph.
test=develop
* Enhance fc_fuse_pass_tester.
* Remove the use of PADDLE_ENFORCE.
test=develop
* Correct the number of ops after fusing.
test=develop
* Fix a typo.
test=develop
* Set activation_type to null when there is no relu in fc.
test=develop
* Refine fc_fuse_pass's codes.
* Enable the set of shape for tensor.
* Refine repeated_fc_relu_pass and add unittest.
test=develop
* Open fuse all reduce op
test=develop
* Add Fuse optimization op log
* Add log in fuse_optimizer op pass and fuse all_reduce op pass
* replace with boost::optional<bool>
test=develop
* Polish code
test=develop
* fix code coverage
test=develop
TemporaryAllocator is a singleton used for allocating memory for Cudnn. Since it is a singleton, we can delete it for better performance in memory.
We replace TemporaryAllocator by CUDADeviceContextAllocator and CUDADeviceContextAllocation, which uses stream callback to delete the memory allocated for the stream to avoid singleton.
Also added data_feed_proto to operator to fix CI in CPU compilation
* Refine the codes related to fc op.
* Add GPU implementation for fc functor.
* Apply fc_fuse_pass in GPU inference.
test=develop
* Change the cmake for fc op.
* Change PADDLE_ENFORCE to PADDLE_ENFORCE_EQ.
* Add an attribute to set the activation type in fc_op.
* Enhance the unittest of fc_op.
test=develop
* Remove the declaration of FCOpGrad back to the header file.
test=develop
* Set default value for newly added arguments in test_fc_op.
test=develop
* Add a interface to enable cudnn for inference.
* Add cudnn_placement_pass.
test=develop
* Set the default value of cudnn_enabled_op_types to null.
test=develop
* Write the common basic class, placement_pass_base, to refine the codes.
test=develop
* Call EnableCUDNN in unittest.
test=develop
* Refine cudnn_placement_pass tester.
* Enable the testing of cudnn_placement_pass in inference's unittest.
test=develop
* Add the check of op kernels.
test=develop
* Support looking up embeddings from BoxPS.
* Add a _pull_box_sparse op, for now this op is not exposed to users.
* Add a BoxHelper class, providing 'BeginPass', 'EndPass', 'FeedPass' functions and so on.
* Add 'BoxPSDataset' in python code.
* Add a compile options WITH_BOX_PS and a MACRO PADDLE_WITH_BOX_PS.
* Add UT.
* More concrete information pls refer to: https://github.com/PaddlePaddle/Paddle/pull/18982
- Refactor step 1
- Compilation fix
- Yet another compilation fix
- Even more compilation fix
- Lint fixes
test=develop
- Removed deprectaed PADDLE_ENFORCE occurance
test=develop
- Candidate fix to BN forward
- Lint fixes
test=develop
- Refactoring in data_layout_transform
- compilation fix
- Another comppilation fix
- Step further into darkness
- Yet another compilation fix
- Yet another compilation fix
- missing header
- compilation fix
- Added MKLDNN -> Paddle conversion in fetch op
test=develop
- Compilation fix
test=develop
- Lint
test=develop
- Mul fix
- Fix to MKLDNN MUL op and Elementwise MUL UT
test=develop
- Workaround for diffrent weights with groups representation Paddle vs
MKL-DNN.
test=develop
- Candidate fix for 5D convolution with groups
- Refactor of fix for conv3d and conv2d in fetch op
test=develop
- Compilation fix
- Still same compilation fix
- Compilation fix
- Compilation fix
- Reverted refactoring of fixes
- Adapted test_conv2d_int8_mkldnn so it exects data in NCHW format
not NHWC
test=develop
- minor fix in UT
test=develop
- Lint fixes
test=develop
* Add simplify_with_basic_ops_pass to replace dropout_op with scale_op when is_test is true.
test=develop
* Delete dropout_op directly when upscale_in_train is true.
test=develop
* Improve the debug string, adding the print of op_desc information.
* Fix the case when dropout's input x is reused as the next op's output.
* Add the pass to inference.
test=develop
* Change the log level.
test=develop
* Add unittest for inplace case.
* Add comment to explain the pass.
* Apply the pass for CPU inference.
test=develop
* Fix the typo.
test=develop
* Add the check of AttrType.
test=develop
* fix correctness of the communicator
* fix a bug in send thread when sending var context is empty, test=develop
* add lookup_table_prefetch_op and prefetch optimize, test=develop
* remove remote prefetch GPU supported
* word2vec force with CPU, test=develop
* test dist remote lookup table force with CPU, test=develop
* add pybind interface to get all inplace ops, test=develop
* enhance OpTest to check whether the consistency of operator when using and not using inplace, test=develop
* handle corner cases in op_test, test=develop
* support outputs without tensor holder_, like XShape in reshape_op, test=develop
* fix bug, some op has GradOpMaker, but actually no grad_op in OpInfoMap, test=develop
* use reshape_grad instead of reshape in FlattenGradOp, test=develop
* fix error debug dims info for variables like XShape, test=develop
* change computational order in sum_op to relieve computation difference using inplace, test=develop
* add inplace_atol to check group_norm, and skip inplace_grad for mkldnn, test=develop
* follow sneaxiy's comments, test=develop
* remove unused DefaultGradOpDescMaker in mkldnn op, test=develop
* Enhance the error message when GrapOpMaker is null.
test=develop
* Call Proto() instead of directly using proto_ pointer.
test=develop
* Rollback to use proto_ directly, because some sepecial grad ops, such some double grad ops, donot have proto.
test=develop
* add a place field in DataFeed to denote which place it will feed data to.
* abstract the copy process in CopyToFeedTensor function
* add UT for float32 type and for CUDAPlace
* Add call stack info during runtime and compile time
test=develop
* Rename operator_call_stack
test=develop
* Add unit test
test=develop
* follow comment
test=develop
* fix QueueDataset queue size，set queue size = batch size * 100, to avoid too many instances in channel when training is much slower than reading data.
* open gc by default, test=develop
* fix test_train_recognize_digits and disable gc when ngraph is enabled, test=develop
* fix conditional_block op eager deletion bug, test=develop
* add some comments to reviewers, test=develop
* support filelist size < trainer num
* pull dense when stop, to make sure local dense params are same as pserver, so save paddle model will save dense model same as pserver
* enable QueueDataset train same filelist for serveral times
* update paddle-trt for:
1. fix bug: when batch > 2, core in split plugin.
2. add leaky_relu trt5.0 support (yolov3 from 65ms to 42ms.)
3. add new attr to dropout.
4. shuffle channel, swish, relu6 support
test=develop
* 1. fix ci
test=develop
The change includes 2 things:
1. save delta model and shrink table are control by the same parameter before, now add delete_after_unseen_days to control shrink table.
2. value in sparse table has no slot before, now add slot in sparse table, and add DownpureCtrAccessor to support the new meta.
test=develop
（1）support patch data （merge slots of instances of same line id, modify dense layer which
changes its size）
（2）add fleet load_one_table interface, support load from paddle model and load from pslib model
（3）fix push sparse bug which cause push sparse cost more time（about 10% in my testcase）
（4）when some slots are not in one of your network (join/update, etc.)，data feed、collect label info、push/pull sparse will skip these slots， instead of throw error.
（5）add more debug info in TrainFilesWithProfiler
Test PaddingRNN on V100 GPU device.
Test configuration: large model, padding mode (which is the mode using recurrentOp), one GPU.
GPU memory (MiB): 6414 (this PR) vs 6837 (without this PR)
Speed (steps/s): 10.28 (this PR) vs 9.89 (without this PR)
* feature/auto_growth_allocator, test=develop
* add unittest of AlignedAllocator, test=develop
* try to turn on auto_growth to test on CI, test=develop
* fix segmentation fault in mixed_vector.h, test=develop
* add unittests, test=develop
* not use transferscope cache in cpu case
test=develop
* adjust variable name and add comments
test=develop
* use correct format for class member in operator.h
* use correct format for class member in operator.cc
test=develop
* Fix Mask rcnn predictor
1. refine memory optim algorithm to support the model with the block op.
2. output diff : modify the affine channel fuse
3. add condition_block_infer op
add interface for setting trt calib table dir
test=develop
* add the missing files.
test=develop
1. Since allreduce op has 4 reduce types, We split these four reduce types into four ops
2. We also refined the collective op code, e.g. we separated the collective op kernel into CPUKernel and CUDAKernel, and remove the device specified DeviceContext parameter in template as we already knew the target DeviceContext
3. We remove the newly added Collective op role to reduce the complexity of program and graph analysis
* fix prepare context redundant code problem, optimize executor by caching create_varaiables
test=develop
* supports collective training in executor
* make fetch_list runable with variables, add more unittest for use_program_cache
test=develop
* fix comment
test=develop
* use unique name for nccl_id
* supports output to stream in program_to_code
* insert sync_comm_stream before regularization; add skip_op_callstack capability in program_to_code
* set op role in collective training
* add collective op role
* remove orig file
* add build optimizer by strategy
* add collective strategy
* refine collective strategy
* add multi-process role maker
* refine strategy building factory so that we can easily plugin more strategy
* scale loss grad in collective sgd transpiler
* add support for distributed fc
* code format
* revert some features for dist fc
* add support for distributed fc training
* fix prepare context redundant code problem, optimize executor by caching create_varaiables
test=develop
* supports collective training in executor
* make fetch_list runable with variables, add more unittest for use_program_cache
test=develop
* use unique name for nccl_id
* supports output to stream in program_to_code
* insert sync_comm_stream before regularization; add skip_op_callstack capability in program_to_code
* set op role in collective training
* add collective op role
* fix comment
test=develop
* remove orig file
* add build optimizer by strategy
* add collective strategy
* refine collective strategy
* add multi-process role maker
* refine strategy building factory so that we can easily plugin more strategy
* scale loss grad in collective sgd transpiler
* add support for distributed fc
* code format
* revert some features for dist fc
* add support for distributed fc training
* test=develop
add collective op unittest standard
* test=develop
remove the test_collective directory
* test=develop
remove the test_collective directory
* remove slicegather test
* code format for reducescatter
* update attr of shard_index_op
* Modify macro nccl_helper
* remove test without distribute
* macro collective_helper
* marcro update
* test=develop
update support python3.5
* test=develop change gpu memory use to 0.1 when test
* test=develop
update ut equal func
* test=develop
set flags to 1.5
* test=develop fix pickle dumple py35
* test=develop
fix divide in slice and add sync_comm_stream
update atol and rtol to 1e-05
rm shard_index op and test
modify read input from file to read from memory
remove origin_program in framework and add i/o in c_sync_calc_stream
* test=develop update unittest sync operator I/O
(1) use channel instead of vector/BlockingQueue in Dataset，to keep same with existing implementation, and make code more readable and flexible (dataset single output channel or multi output channel). one previous memory out of limit problem is cause by not release memory after training.
(2) add Record because MultiSlotType costs too much memory (80B)，fix memory out of limit problem.
(3) add Channel, Archive in paddle/fluid/framework
(4) change dataset from shared_ptr to unique_ptr in pybind
(5) move create/destroy readers from trainer to dataset
(6) move shuffle from datafeed to dataset. dataset holds memory, datafeed is only for load data and feed data to network.
(7) fix thread num bug of Dataset when filelist size < thread num
(8) support set_queue_num in InMemoryDataset
Add Pipeline Concurrency Train Mode:
- Cpp: pipeline_trainer & section_worker
- Python: PipelineOptimizer
- Add a new data_feed type: PrivateInstantDataFeed
- Add a test demo of pipeline trainer and the test model is gnn
- Do not support win32 now
* fix prepare context redundant code problem, optimize executor by caching create_varaiables
test=develop
* cache sub_scope, program, var when use_program_cache=True is set
* make fetch_list runable with variables, add more unittest for use_program_cache
* Enhance fused_elementwise_activation op.
test=develop
* Move the api fused_elementwise_activation to contrib.
test=develop
* Add including files.
test=develop
* Add the support of sigmoid in fused_elementwise_activetion op.
* Update API.spec.
test=develop
* add Concat quantization
add unit test for quantizing concat
fix for wrong value when the input is not in map of calculated scales
add use_quantizer to concat_op.cc
add scale_algo rules for concat
test=develop
* missing fix for multiple inputs quantize-squash
* wojtuss review fix: adding comment
test=develop
* fluid int8 train and trt int8 predict align.
trt int8 predict init
op converter
* 2. align fluid int8 train and trt int8 inference.
enhance quant dequant fuse pass
enhance op converter, trt engine, trt engine op, trt subgraph pass.
* 3. add delete_quant_dequant_pass for trt
test=develop
* 4. add the missing file
test=develop
* 5. i modify the c++ interface, but forget to modify the pybind code
fix the IS_TRT_VERSION_GE bug, and fix elementwise op converter
test=develop
* fuse mul and elementwise add to fc
* Reimplement the FC forward operator
* Fix FC MKLDNN integration by transposing weights
* Add FC MKLDNN Pass
test=develop
* FC MKLDNN Pass: change memcpy to std::copy
* Fix MKLDNN FC handling of mismatch input and weights dims
* Lower tolerance for MKL-DNN in resnet50 test
test=develop
* Adjust FC to support MKLDNN Op placement
test=develop
* Adjust Placement Op to set use_mkldnn attribute for graph
test=develop
* MKLDNN FC: fix weights format so that gemm version is called
test=develop
* FC MKLDNN: Remove tolerance decrease from tester_helper
* FC MKL-DNN: Refactor the code, change input reorder to weight reorder
* MKL-DNN FC: Introduce operator caching
test=develop
* FC MKL-DNN: Fix the tensor type in ExpectedKernelType
test=develop
* FC MKL-DNN: fix style changes
test=develop
* FC MKL-DNN: fallback to native on non-supported dim sizes
test=develop
* FC MKLDNN: fix CMake paths
test=develop
* FC MKLDNN: Refine placement pass graph mkldnn attribute
test=develop
* Fix Transpiler error for fuse_conv_eltwise
test=develop
* Fix missing STL includes in files
test=develop
* FC MKL-DNN: Enable new output size computation
Also, refine pass to comply with newest interface.
test=develop
* FC MKL-DNN: enable only when fc_mkldnn_pass is enabled
* FC MKL-DNN: Allow Weights to use oi or io format
* FC MKL-DNN: Adjust UT to work with correct dims
test=develop
* Enable MKL DEBUG for resnet50 analyzer
test=develop
* FC MKL-DNN: Improve Hashing function
test=develop
* FC MKL-DNN: Fix shape for fc weights in transpiler
* FC MKL-DNN: Update input pointer in re-used fc primitive
* Add log for not handling fc fuse for unsupported dims
test=develop
* FC MKL-DNN: Move transpose from pass to Op Kernel
test=develop
* FC MKL-DNN: Disable transpose in unit test
test=develop
* FC MKL-DNN: Remove fc_mkldnn_pass from default list
* Correct Flag for fake data analyzer tests
test=develop
* FC MKL-DNN: Add comment about fc mkldnn pass disablement
test=develop
* FC MKL-DNN: Disable fc in int8 tests
test=develop
* add conv_concat_relu fuse
test=develop
* add test code
test=develop
* added missing include with unordered_map
test=develop
* review fixes for wojtuss
test=develop
* remove 'should (not) be fused' comment statements
one of them was invalid anyway
test=develop
* fix quantize_squash_pass segfault when there is no tensor linked do Bias input
test=develop
* add googlenet test
test=develop
* fix concat CreateKey not using input format
test=develop
* Relu6 is the bottleneck op for Mobilenet-v2. As the mkldnn supports the conv/relu6 fusion, we implement it fusion via cpass way. Due to the int8 enabling for this fusion will be supported in MKLDNN v0.20, so this PR is focused on the fp32 optimization.
Below table shows the benchmark(FPS) which measured on skx-8180(28 cores)
Batch size | with fusion | without fusion
-- | -- | --
1 | 214.7 | 53.4
50 | 1219.727 | 137.280
test=develop
* Fix the format issue
test=develop
* Add the missing nolint comments.
test=develop
* Fix the typos.
test=develop
* Register the conv_brelu_mkldnn_fuse_pass for the MKLDNN engine.
test=develop
* Adjust the indentation.
test=develop
* Add the test_conv_brelu_mkldnn_fuse_pass case.
test=develop
* Slightly update the code per Baidu comments.
Let the parameter definition embedded into the code.
That's will make the code easy to understand.
test=develop
* Add conv2d_grad_grad_op
* Extracte the cuDNN conv algo searching code in conv_cudnn_helper.h.
- Now use it in conv2d_grad_grad.
- Will simply the searching code in conv2d and conv2d_grad in next PR.
* Enhance and fix bug in unit testing of gradient_checker.
* Support to fetch empty variables，return None in Python.
* fix train_from_dataset and infer_from_dataset example
* add inductive dim for data_reader, example: shape=[-1, 1], then -1 will be inducted through run-time reading of number of elements
* add use_cuda to inplace pass,test=develop
* add test softmax_with_xe_inplace test,test=develop
* fix potential inplace bug
test=develop
* add more skip vars in mem opt pass,test=develop
* follow comment,test=develop
* follow comments,move duplicate out arg check to program->graph,test=develop
* fix bn fuse vardesc and add model saver
test=develop
* unify save model in test helper
test=develop
* fix mkdir on windows
test=develop
* remove magic number use bn bias var desc
test=develop
* speedup gc and inplace softmax_with_cross_entropy_grad
test=develop
* refine models gpu mem
Merge skip vars and warning messages of mem opt
remove relu mem opt
test=develop
* follow comments
test=develop
123malin
hong
Thunderbrook
Yiqun Liu
Huihuang Zheng
Zeng Jinle
Chen Weihang
xujiaqi01
wangchaochaohu
zhongpu
石晓伟
wopeizl
WangXi
Jacek Czaja
Chengmo
633WHU
Pei Yang
zhaoyuchen2018
Adam
tangwei12
chengduo
qingqing01
bingyanghuang
joanna.wozna.intel
Leo Chen
Jiabin Yang
Huihuang Zheng
Tao Luo
wangguanzhong
hutuxian
mapingshuo
tensor-tang
baojun
gongweibao
yaoxuefeng
Zhaolong Xing
liuwei1031
juncaipeng
Leo Zhao
chengduo
fuyinno4
Huihuang Zheng
guru4elephant
pkpk
Yi Liu
Michał Gallus
HaoRen
Sylwester Fraczek
gongweibao
wopeizl
mozga-intel
Qiao Longfei
guomingz
Zhen Wang
songhao
wopeizl
Yibing Liu