* update benchmark for int8v2, QAT1, QAT2 accuracy and performance
test=document_fix
* change according to reviews
test=develop test=document_fix
* improve some descriptions and some models
test=develop test=document_fix
* update models benchmark data
test=develop test=document_fix
* update int8v2 and qat2 performance
test=develop test=document_fix
Add tests to use dy/dx to make sure the gradient values calculated by the control flow backward is correct. Also fixed bugs detected by those tests.
Fix bugs:
1. Unlike sum_op, optimizer ops don't allow uninitialized input tensor. But in conditional_block_grad_op, since the conditional_block may not run, the output gradient tensor may be uninitialized, which will cause the optimizer op error. To fix it, we should let optimizer ops support uninitialized input like sum_op or assign the uninitialized gradient to 0 when the conditional_block_grad_op doesn't run. I found there are about 10+ optimizer ops. **To be simpler, I just assign output gradient of the conditional_block_grad_op to 0 in this PR**. But it can be further explored whether we can make optimizer ops like sum_op to support uninitialized input tensor because theoretically we can speed up without the assigning in conditional_block_grad_op.
2. Infer parameter shapes during append_backward. I didn't know that all our parameters are in global block. When op_desc is inferring shapes at the sub-block, it may not know the shape of gradients of parameters whose shape information is at global block. I fixed it by inferring shapes of gradients from forward var.
This PR also did some code clean up:
1. Print the var name when sgd_op catches shape error so that it is easier to debug
2. Fix a typo: dicta -> dict
* add file check_op_desc.py and add interface to get default value. test=develop
* add test for c++ coverage rate. test=develop
* Correct typo. test=develop
* test=develop, fix docker with paddle nccl problem
* don't expose numerous Tensor.set(), test=develop
* fix condition, test=develop
* fix float16 bug, test=develop
* feed should be Tensor or np.array, not Variable or number, test=develop
* use forcecast to copy numpy slice to new array, test=develop
* remove float16-uint16 hacking, test=develop
* add variable method to varbase and refactor to_variable to support return varbase
* support kwargs in varbase constructor
* add VarBase constructor to support default python args
* refine varbase initial method
* reset branch
* fix ut for change VarBase error info to PaddleEnforce
* cherry is parameter change before
* overload isinstance to replace too many change of is_variable
* rm useless files
* rm useless code merged by git
* test=develop, fix some ut failed error
* test=develop, fix test_graph_wrapper
* add some tests, test=develop
* refine __getitem__, test=develop
* add tests, test=develop
* fix err_msg, test=develop
* add data_type register CI test=develop
* add op add test case test=develop
* add test case for register op kernel test=develop
* fix shell script bug test=develop
* fix checkout branch test=develop
* remove test case code test=develop
* fix op_type.spec name test=develop
* fix default grad op maker ci bug, test=develop, test=document_fix
* remove some ops from paddle/fluid/op_use_default_grad_op_maker.spec, test=develop, test=document_fix
* fix the device supported of the op unique and unique_with_counts.
test=develop
test=document_fix
* Fix the precision of test in the op of unique and unique_with_counts.
test=develop
test=document_fix
* add param & grad shape check for sgd op
* add _reshape_inplece interface for dygraph parallel
* refine unittest based paddle/models scripts, test=develop
* add unittest for parallel grad fuse, test=develop
* Commit before merging develop
test=develop
* Backup after working with Huihuang logs
* Commit before deleting Huihuang debug loggings
* Commit before debug
test=develop
* Fix bug commit
test=develop
* Backup of fixing bugs
test=develop
* Clean up code
test=develop
* Fix a bug in sum_op
test=develop
* Add ascending for argsort
* Refine api doc description.
* Refine descending description
* Add int32 logic to speedup when data is small size.
* Remove int32 opt as not support in python
* add ut for comparing FP32 and QAT INT8
* add save qat transformed model python script
test=develop
* updated
* added missing file
* add "with_label"
test=develop
* performance benchmark as unit test
test=develop
* change names of unnecessary thing
* Change CMakeList.txt for model downloading and UT
test=develop
* change names of functions and params for more readable code
test=develop
* Change PADDLE_ENFORCE messages
test=develop
* fix indent problems
test=develop
* indent problems
test=develop
* Implement Int8 FC
* Integrate FC into INT8v2
test=develop
* int8 FC: transpose weights before computing scales
test=develop
* Add support for activation_type string in FC
test=develop
* Disable MKL-DNN's FC in VGG16 and 19
test=develop
* Disable FC quantization when mkldnn FC is disabled
test=develop
* Solve PADDLE_ENFORCES in FC int8
* Fix Paddle enforces and remove const cast
test=develop
* Fix style changes
test=develop
* Fix quantizer_tester test and add fc quantization
test=develop
* Fix FC test fail on CUDA
* Remove unnecessary log from quantize placement pass
test=develop
* Add Thread ID to FC hash key
test=develop
* Add comments to MKL-DNN FC Kernel
test=develop
* Refactor quantizer
test=develop
* Fix linter issues
test=develop
* Fix crash in slim googlenet
test=develop
* Fix PADDLE_ENFORCE messages
test=develop
* Add fc padding to solve mkl performance
test=develop
* fix gpu pass and error information
test=develop
* fix fc_fuse_pass_test
test=develop
* fix error information
test=develop
* fix error information
test=develop
* fix name and add fc op padding test
test=develop
* fix attributes
test=develop
* optimize fc padding
test=develop
* fix test
test=develop
* Refactor MKL-DNN ElementwiseMul
remove manual fallback, remove format attrs
test=develop
* Refine PADDLE_ENFORCEs in eltwise_mul_op.h
test=develop
* Make ElementwiseMulOp inherit from ElementwiseOp
* Change type of simd_width to int
test=develop
* Remove Constructor extensions in ElementwiseOp and ElementwiseMulOp
test=develop
* Restore attributes
test=develop
* Fix test coverage for mkldnn eltwise mul
test=develop
* Conform to new is_run_common_broadcast API
test=develop
* Add UT for AreDimsAndFormatCorrect
test=develop
* Improve argsort performance.
- Give 200000 data to compute argsort on v100,
can speed up ~190x
before opt cost: 0.53s
after opt cost:0.0027s
- Add fp16 support
* Refine error message
* Refine code
test=develop
Signed-off-by: zhaoyuchen <zhaoyuchen01@baidu.com>
* fix fetch handler problem and refactor
when a user define FetchHandler class, he or she should initialize a handler
with variable dict. the key of a variable dict is a user defined name,
the value of a variable dict is a Varaible generated from python API.
For each fetching, a user should implement handler function in which
fetched_result_dict will be available and the user can access the fetched value
with user defined keys.
* Disable fusion_group pass for windows and mac. We will do some experiments on Linux first.
test=develop
* Print the subgraph when check failed.
test=develop
* add int8 kernel to lookup_table op and add dequantize op test=develop
* change paddle_enforce to paddle_enforce_eq test=develop
* change copyright and change some not suitable code test=develop
* remove debug log test=develop
* replace GetInputType with IndicateVarDataType test=develop
* fix EmptyGradMaker test=develop
* fix diff between cpu and gpu test=develop
* use memcopy when int8_t test=develop
* open dygraph op test, test=develop
* modify to_variable, test=develop
* modify input and output for dygraph, test=develop
* modify input and output for dygraph(fix bug), test=develop
* fix input processing of dygraph op test, test=develop
* fix bug, test=develop
* fix op test, test=develop
* fix forward bug for dygraph, test=develop
* fix mkldnn op test for forward, test=develop
* update nn.py for dygraph, test=develop
* fix crop_tensor_op, test=develop
* fix elementwise_mul_op, test=develop
* fix fill_op, test=develop
* fix some mkldnn op, test=develop
* open backward op test for dygraph, test=develop
* delete log, test=develop
* close backward op test for dygraph, test=develop
* fix bug for edit_distance_op and test_lstm_cudnn_op, test=develop
* fix optest backward bug for dygraph, test=develop
* fix optest backward bug for dygraph, test=develop
* close backward op test for dygraph, test=develop
* close backward op test for dygraph, test=develop
* open dygraph op test, test=develop
* fix op test for dygraph, fix GradOpDescMaker, test=develop
* fix bug for linear_chain_crf_op.h, test=develop
* remove log, test=develop
* remove log, test=develop
* remove log for op_test.py, test=develop
* remove log for op_test.py, test=develop
* fix bug for var_conv_2d_op, change PADDLE_ENFORCE, test=develop
* fix PADDLE_ENFORCE_EQ for hierarchical_sigmoid_op.cc, test=develop
* fix bug for test_increment_ngraph_op.py, test=develop
* fix lod for op test in dygraph, test=develop
* refactor op_test.py to reduce redundant code, test=develop
* fix lod optest, modify InputVar/OutputVar to HasInput/HasOutput, test=develop
* remove debug log, test=develop
* remove redundant code in base.py, test=develop
* fix some error in optest, test=develop
* fix ClearNoNeedBufferInputs function's bug for LoDTensor, test=develop
* refactor op_test.py, test=develop
* remove redundant writing, test=develop
* fix error(get tensor of the grad variable), test=develop
* fix test_concat_mkldnn test_conv2d_mkldnn, test=develop
* fix optest.py for get tensor of LoDTensor, test=develop
* fix optest.py for get tensor of LoDTensor, test=develop
* fix optest.py for get tensor of LoDTensor, test=develop
* fix some redundant code, test=develop
* reslove conflict and rewrite paddle error message, test=develop
* fix the CAPI ZeroCopy shape error and reconstruct the output obtain
* use an anonymous namespace to cover the functor
* fix unit tests because of the output of typeid(T).name() is different from linux and windows, test=develop
* Enable generating code for a given subgraph.
* Support sorting the subgraph.
* Remove the rearange of expressions because we use the sorted subgraph directly.
* Enable generating code for a subgraph which is composed of grad ops.
* Use expression information to check the accuracy in unittest.
* Separate load and store from computation expressions.
test=develop
* Improve the loading statements in generated codes.
test=develop
* Remove unused arguments from formal list.
test=develop
* fix auc drop first commit test=develop
* update datanorm op
* update datanorm with enforce test=develop
* update test=develop
* update format test=develop
* update format
* update format test=develop
* add unit test test=develop
* update unit test test=develop
* update format test=develop
* update format test=develop
* update API description test=develop
* update API description test=develop
* update format test=develop
* fix codes as comments test=develop
* fix description as comments test=develop
* fix description as comments test=develop
* update codes.. test=develop
* modified error message for conv and conv_transpose, test=develop
* modified doc of conv and conv_transpose op, test=develop
* modified the expression for error message, test=develop
* modified error message for group_norm op, test=develop
* modified detail of Attr(data_format) or Attr(data_layout)
* add ValueError in API doc for maxout op, test=develop
* copy some feasigns and corresponding embeddings from one sparse table to another
* copy all feasigns and corresponding embeddings from one sparse table to another
* copy all dense params from one table to another
* copy some local vars to other local vars
* Add Asypadding for conv fusion.
test=develop
reference: pr/20042
* Fix eigen build link error
* Change back file mode
* Use math function & add more checks.
* set the default value of alpha for prelu to 0.25, test=develop
* add the call to __syncthreads(), test=develop
* fix the implementation of cpu prelu, test=develop
* repair the implementation of element mode prelu, test=develop
* modify test_prelu_op.py, test=develop
* Add the check of lod_level between compile-time and runtime.
test=develop
* Fix bug in check_compile_vs_runtime.
test=develop
* Fix the check of output when it is dispensiable or intermediate.
test=develop
* Share lod of x to out in match_matrix_tensor op in compile-time.
* Implement GetLoDLevel in InferShapeContext.
* Set the default value of check_compile_vs_runtime to False and enable it in test_sequence_pad_op.
test=develop
* Enable check_compile_vs_runtime in test_match_matrix_tensor.
* Add the implementation of SetLoDLevel in InferShapeContext.
* Remove the implementation of IncreaseLoDLevel and call Get/SetLoDLevel instead.
* Remove the implementation of DecreaseLoDLevel and call Set/GetLoDLevel instead.
* Refine some ops and unittests.
test=develop
* Fix a typo.
test=develop
* Remove the check of var type, and change int to int32_t.
test=develop
* Add unittest for Get/SetLoDLevel.
test=develop
* Add the definition of operation in fusion_group.
* Use operations in OperationMap to detect fusion_group of elementwise pattern.
* Add namespace fusion_group in code_generator.
* Use operations recorded in OperationMap to generate code.
* Remove implementation codes to .cc file.
* Refine Operation and CodeGenerator to make it easier to generate code for grad_op.
Refine the unittest for better reuse.
* Avoid recording the template's keyword in a array.
* Support the generating of code for grad_op and add unittest.
test=develop
* Remove replaced_element_in_order and use use number instead.
test=develop
* fix bug in pool/conv/conv_transpose:
1. It should be stride[i] not stride[0] in UpdatePaddingAndDilation;
2. fix bug of func _get_padding_with_SAME in test_conv/conv_transpose_op.py;
3. fix bug of the computation process in function conv2dtranspose_forward_naive.
test=develop
* change test to make the data of different dimensions different. test=develop
* Add ernie unit test
test=develop
* Add ernie unit test
test=develop
* Add ernie unit test
test=develop
* remove ngraph
* optimize gpu test
test=develop
* optimize codes
test=develop
* Enrich the type of error and declare the error type interfaces, test=develop
* adjust tests to adapt new form, test=develop
* add inference deps with error_codes.pb.h, test=develop
* restore stack iter start pos, test=develop
* polish code based review comments, test=develop
* Add asymetric padding support for mkldnn pooling
test=develop
* Add asymetric padding support for mkldnn conv
test=develop
* Add asymetric padding support for mkldnn conv_transpose
test=develop
* Add c++ global current tracer for dygraph, test=develop
* add tracer property in c++, test=develop
* support different place, test=develop
* add unittest for tracer, test=develop
* remove duplicate code and duplicate config of master+patch
* drop all ins which has conflict slot or size < merge_size
* user only need to set merge size，if ins num of same id is not equal to merge size, just drop these ins
* user must make sure master data and patch data has no same slot whose feasigns are both non-zero, otherwise these ins will be dropped. (slot list should still be the same of both master and patch)
* test=develop
* support no need buffer vars in dygraph, test=develop
* fix inference compilation error, test=develop
* update no_need_buffer_vars_inference, test=develop
* add unittests for no_need_buffer_vars_context, test=develop
* refine no_need_buffer_vars by return ref, test=develop
* polish some codes, test=develop
fix the bug of conv_transpose cudnn kernel: before version 1.6, the data_format is AnyLayout in inference model. When use version 1.6 and load the model which is saved by previous version, the error occurs. This is because the cudnn kernel in version 1.6 is not compitable with Anylayout setting.
* don't expose numerous Tensor.set(), test=develop
* fix condition, test=develop
* fix float16 bug, test=develop
* feed should be Tensor or np.array, not Variable or number, test=develop
* use forcecast to copy numpy slice to new array, test=develop
* remove float16-uint16 hacking, test=develop
Adam
rensilin
xujiaqi01
wangchaochaohu
Leo Chen
lidanqing
liym27
Zeng Jinle
Aurelius84
zhouwei25
Huihuang Zheng
hutuxian
Jacek Czaja
Pei Yang
Youwei Song
Zhaolong Xing
WangXi
hong
GaoWei8
zhongpu
Tao Luo
lilong12
tangwei12
Lv Mengsi
wawltor
wangguanzhong
Jie Fang
Chen Weihang
zhaoyuchen2018
Kaipeng Deng
tianshuo78520a
ShenLiang
Michał Gallus
silingtong123
Thunderbrook
Leo Zhao
Dong Daxiang
gongweibao
Yiqun Liu
Yihua Xu
Zhang Ting
Liufang Sang
danleifeng
liu zhengxi
yaoxuefeng
liuwei1031
guofei
ruri
whs
Zhang Ting
Aurelius84
joanna.wozna.intel
LielinJiang
Wojciech Uss
mapingshuo
ddokupil
Chengmo
Yucheng
Wilber
123malin