-
Notifications
You must be signed in to change notification settings - Fork 5.6k
2017 08 09
Code style
Intel MKLDNN Design
-
https://github.com/PaddlePaddle/Paddle/pull/3241#pullrequestreview-54421884
-
Refactorize enforce_test.cc #3344
-
Fix some build warnings #3341
-
Remove curl progress info in Dockerfile #3326
-
Remove wget progress of downloading Go from build log #3324
-
Refactorize framework/*.proto #3322
-
Change git tag of any #3313
-
Use PaddlePaddle fork of any #3286
-
Use boost::any other than linb::any #3284
-
Add --expt-relaxed-constexpr to surppress Eigen warnings #3282
-
Fix comparison between signed and unsigned values #3279
-
Remove paddle/utils/Flags.h warnings #3278
-
Add environment variable WITH_DEB to trigger the build of .deb file #3276
-
Reformat paddle/operators/* strictly following Google Style Guide #3262
-
Add explicit to some constructors #3256
-
Simplify test_matrixCompare #3253
-
Disable MKL and OpenMP by default #3230
-
Re-structure source code of type Attribute #3225
- Rename PROJ_ROOT to PADDLE_SOURCE_DIR https://github.com/PaddlePaddle/Paddle/pull/3378
- Enable Python Unit Test before make install https://github.com/PaddlePaddle/Paddle/pull/3360
- [Merge] Accelerate CI process under android environment https://github.com/PaddlePaddle/Paddle/pull/3299
- [Merge] enhance cpplint cmake https://github.com/PaddlePaddle/Paddle/pull/3295
- [Merge] Import HPC linear algebra libs as cblas target for cc_library deps https://github.com/PaddlePaddle/Paddle/pull/3290
- [Merge] ClangFormat for proto and cuda https://github.com/PaddlePaddle/Paddle/pull/3228
- [Merge] cpplint code style for *.h and *.cu https://github.com/PaddlePaddle/Paddle/pull/3227
- [Merge] py_test in generic cmake
One Day: MKLML RPATH for paddle_pybind -> pip install paddleapddle
Review:
- Add --expt-relaxed-constexpr to suppress Eigen warnings
- Enforce add more helpers
- Add unittests for
grad_op_builder
: - [WIP] Refactor
op_registry
: - [WIP] Compelete backward doc
- code review:
- Paddle
- Add some to Paddle documents in Intranet
- Guide the user to supplement the document
- Fix C-API compilation bugs
- PaddleCloud
- Fix ls public dataset bugs:
- Blog
- GNR
-
[Merged] Enable defining sub-sequence data in test layer gradients:
-
[Merged] Add gradient check for sequence softmax activation.
-
[Merged] Add a Kmax sequence score layer:
-
[Merged] Add a nest sequence select layer:
-
Add a sequence slice layer:
-
Up to now, all the layers involved in expanding the beam are finished. I am working on writing the last layer: globally normalized cross-entropy over beam.
-
BUT ONE PROBLEM:
- In PaddlePaddle, the currently available matrixes all have real-typed data fields, but the selected indices information is actually int-typed (with -1 as a special flag token).
- The selected indices is an int-typed matrix and have sequence start position and sub sequence start position information. This means the selected indices are nested sequence. the So it cannot be stored into Argument.ids with sequence start position and sub-sequence start position information to act as an int-typed Matrix.
- Storing indices information in Matrix leads to converting real to int. This is very dangerous if the user fills this matrix himself, invalid data may occur. The selected indices should be stored in a matrix with int-typed value. I will fix this later.
-
What has been done.
- Reduce Python unittest.
- Change Operator Creation API in Python
- op.fc(input=xxx, size=xxx) to Operator("fc", input=xxx, size=xxx)
- https://github.com/PaddlePaddle/Paddle/pull/3237
- Add uniform random operator
- Make Our C++ code fit google C++ style.
- Several Enhancements for others PR
- Implement to_string, instead of
std::to_string
to make PADDLE_ENFORCE_EQ extendable. - Better message for Gradient Check
- Git Ignore enhancement https://github.com/PaddlePaddle/Paddle/pull/3334
- remove unnecessary C++ unit tests https://github.com/PaddlePaddle/Paddle/pull/3355
- Implement to_string, instead of
What under doing.
- [WIP] Refactorize framework/*.proto
- https://github.com/PaddlePaddle/Paddle/pull/3322
- co-work with Wang Yi, Dang QingQing.
- reduce travis-ci/teamcity time:
- reduce travis-ci time of build_doc (with @liuyiqun): https://github.com/PaddlePaddle/Paddle/pull/3310
- reduce time of test_TrainerOnePass: https://github.com/PaddlePaddle/Paddle/pull/3296
- code review:
- Intel MKLDNN:
- Enable mkldnn_fc for general format: https://github.com/PaddlePaddle/Paddle/pull/3337#pullrequestreview-55141837 https://github.com/PaddlePaddle/Paddle/pull/3337#pullrequestreview-55183349
- MKLDNN design doc: https://github.com/PaddlePaddle/Paddle/pull/3241#pullrequestreview-54311590
- other:
- help @caoying extract a modification in RecurrentLayerGroup: https://github.com/PaddlePaddle/Paddle/pull/3208
- Add nest sequence select: https://github.com/PaddlePaddle/Paddle/pull/3297#pullrequestreview-54575049
- Add a Kmax sequence score layer: https://github.com/PaddlePaddle/Paddle/pull/3249#pullrequestreview-54602368
- Fix curdir error in book: https://github.com/PaddlePaddle/book/pull/363#pullrequestreview-55161671
- Intel MKLDNN:
- MPI: receiver monitor
- paddle whl package all in one, can run
pip install paddlepaddle
now - Go math lib experiments, code-generation/type reflect
- New pserver use OptimizerConfig.proto
- PaddlePaddle cloud share/Blog
- large model for v1 bug fixing
- PaddleCloud
- [Merged]Add demo of understand_sentiment on paddlecloud
- [Merged]Add demo of recognize_digits on paddlecloud
- FCN
- [WIP]Debug reason of diff between paddle and caffe in FCN model's convolution layer backward
- DeepSpeech2
- [WIP]Implement uploading data automatically and fix some issues
Operator --> OpKernel --> Tensor/DeviceContext --> Eigen
-
port gemm to new framework
-
refine the dependencies of BLAS related libraries
-
fix and enhance:
- remove eigen tensor header file in dddim.h
- use isclose instead of assert_almost_equal
- PaddleCloud
- Run fit-a-line demo on PaddleCloud with Fault tolerant mode, https://github.com/PaddlePaddle/cloud/pull/278 https://github.com/PaddlePaddle/cloud/pull/272
- Prepare public recordio dataset,https://github.com/PaddlePaddle/cloud/pull/269/files
- DLNEL PaddleCloud cluster is ready
- PaddleCloud PPT
- Blog outline:https://github.com/PaddlePaddle/blog/issues/5#issuecomment-321406453
- command-line interface design doc: https://github.com/PaddlePaddle/cloud/pull/291/files
- Paddle
- Following up save large parameters bug: https://github.com/PaddlePaddle/Paddle/issues/2895
- PaddlePaddle Refactoring
- Refine unit test in op_test_util: https://github.com/PaddlePaddle/Paddle/pull/3238
- Modify rnn op unit test after refactoring framework proto: https://github.com/wangkuiyi/Paddle/pull/5
- Update grad_op_builder after refactoring framework proto: https://github.com/wangkuiyi/Paddle/pull/7
- code review:
- Other:
- add a batch norm inference kernel: https://github.com/PaddlePaddle/Paddle/pull/3309
- fix test_KmaxSeqScore for only CPU compile: https://github.com/PaddlePaddle/Paddle/pull/3361
-
paddle refactor
- PR
- framework related
- rnn related
- PR
-
others
- VisualDL with @qiaolongfei
- a discussion with FE team, determine the architecture of backend, workflow and so on.
- Community construction
- blog theme with @liuyongfeng
- VisualDL with @qiaolongfei
-
Refactorization:
- add mul op backward implement https://github.com/PaddlePaddle/Paddle/pull/3379
- add rowwise op backward implement https://github.com/PaddlePaddle/Paddle/pull/3183
- remove type_alias to follow google c++ style https://github.com/PaddlePaddle/Paddle/pull/3300
- refine device context https://github.com/PaddlePaddle/Paddle/pull/3289
- add no gradient https://github.com/PaddlePaddle/Paddle/pull/3166
- refine gaussian random. https://github.com/PaddlePaddle/Paddle/pull/3060
-
discussion some discussion in paddle cloud blog. https://github.com/PaddlePaddle/blog/issues https://github.com/PaddlePaddle/blog/issues/2
-
reviews Refactorization:
- refine variable.h const define https://github.com/PaddlePaddle/Paddle/pull/3350
- uniform random op https://github.com/PaddlePaddle/Paddle/pull/3293
- fix comparision in test https://github.com/PaddlePaddle/Paddle/pull/3279
- simplify gradient operator building https://github.com/PaddlePaddle/Paddle/pull/3192
- implement sigmoid backward op https://github.com/PaddlePaddle/Paddle/pull/3181
- unittest of fill_zeros_like op https://github.com/PaddlePaddle/Paddle/pull/3142
Paddle Cloud
- fix Pserver save/load checkpoint https://github.com/PaddlePaddle/Paddle/pull/3222
- golang pserver use OptimizerConfig https://github.com/PaddlePaddle/Paddle/pull/3358
DS2 on Paddle
- Data Provider
- Noise&Impulse Augmentor: https://github.com/PaddlePaddle/models/pull/140
- Efficiency
- Multi-GPU Acceleration Rate: https://github.com/PaddlePaddle/Paddle/issues/3137#issuecomment-320996396
- MFCC Acceleration: https://github.com/PaddlePaddle/models/pull/189
- Data Acceleration: https://github.com/PaddlePaddle/models/pull/188
- https://github.com/PaddlePaddle/models/pull/192#pullrequestreview-55430256
- Mandarin Adaptation
- Cloud Training
- Reduce test_CompareSparse unit test time: https://github.com/PaddlePaddle/Paddle/pull/3196
- Paddle cloud improvements and fixes:
- Implement trainer init parameters election with etcd: https://github.com/PaddlePaddle/Paddle/pull/3321
- Fix pserver save / load checkpoint: https://github.com/PaddlePaddle/Paddle/pull/3222
- Master persist more states to etcd: https://github.com/PaddlePaddle/Paddle/pull/3223
- Add recommender system demo of fault tolerant mode: https://github.com/PaddlePaddle/cloud/pull/284
- Blog "Go在分布式机器学习系统中的应用" outline: https://github.com/PaddlePaddle/blog/issues/1
- PaddlePaddle Image Enhancement
- Faster R-CNN: RCNNDetectionLayer in developing.
- Cooperation with Video Analysis Team in IDL:
- Debug, Config and train the Attention Net with Li Fu.
- Up-to-now, the classification error is 0.240 which is one percent heigher than the Pytorch benchmark.
- Add Conv3DLayer, DeConv3DLayer
- CPU implementation.
- GPU implementation.
- Add test_Conv3D and Test_DeConv3D unit test.
- Pass CPU unit test
- Testing GPU implementation
- https://github.com/PaddlePaddle/Paddle/pull/3348
- Add 3D-Pooling
-
- add gradient test framework(with @yuyang and @qingqing)
- https://github.com/PaddlePaddle/Paddle/pull/3226
- check INFINITY in cross_entropy
- add support_gpu
- merge InferShapeContext and OperatorContext
- support python test without installation python package
- Remove unnecessary C++ operator test: 3355
- A better error message for gradient checker: 3354
- Feature/change op creation: 3237
- VisualDL with @yanchunwei a discussion with FE team, determine the architecture of backend, workflow and so on.
- Adapt ds2 to mandarin data
- Debug gradient problem with @haoshuang
- Wrap and analyze dl-speech decoder
- Code Review
-
a tool that merge batch normalization with fc or conv https://github.com/NHZlX/PaddleModelZoo/tree/master/Merge_bn_paddle
-
test 3 * 3 conv neon acceleration of paddle on android
in process
-
Mobile
-
Bug fix & Review
- Use paddle instead of caffe in genome's album classification application, and test the runtime.
- Review
Deep Speech 2:
-
Retrain model based on mfcc feature
-
Catch up with some key tricks in speech-dl decoder (with@yaming)
-
Improve params tuning strategy for CTC beam search decoder
-
Some minor fixes
-
Code Review:
Get Familiar with our Paddle framewok
- Implement the gather function (can be later integrated into a gather op), merged
- Implement the scatter function (wait for review) To-do: use gather and scatter to implement the control-flow (if-else/switch) op