2017 09 07

wangkuiyi

• Distributed training and graph serialization:

  https://github.com/PaddlePaddle/Paddle/pull/3811#pullrequestreview-60279820

• Should/how we allow users to write the train loop:

  https://github.com/PaddlePaddle/Paddle/pull/3811#pullrequestreview-60280583

• RNN design doc complementary:

  https://github.com/PaddlePaddle/Paddle/pull/3727#pullrequestreview-60284275

• RNN and block:

  https://github.com/PaddlePaddle/Paddle/pull/3708#pullrequestreview-60283836

• Functions, operators, and layers:

  https://github.com/PaddlePaddle/Paddle/pull/3827

• IfElseOp

  https://github.com/PaddlePaddle/Paddle/pull/3828#pullrequestreview-60288081

• VarDesc

  https://github.com/PaddlePaddle/Paddle/pull/3835#pullrequestreview-60305827

• Block design

  https://github.com/PaddlePaddle/Paddle/pull/3708#pullrequestreview-60456742

• Graph construction design

  https://github.com/PaddlePaddle/Paddle/pull/3860

• Refactorize operator unit test framework

  https://github.com/PaddlePaddle/Paddle/pull/3882#pullrequestreview-60818278

Yu Yang

• Change GetAttr -> Attr
  • https://github.com/PaddlePaddle/Paddle/pull/3903
• Describe problems for Functor
  • https://github.com/PaddlePaddle/Paddle/pull/3893
  • We will not give a unify way for functor and GPU kernel fusion right now
• Make operator Input/Output can return nullptr
  • https://github.com/PaddlePaddle/Paddle/pull/3831
• Reviewed PR https://github.com/PaddlePaddle/Paddle/pull/3857
  • It implement gradient of MulOp and RowwiseAddOp correct.
  • Give a Patch to not modify the unit-test framework
• Make operator can return nullptr & empty
  • https://github.com/PaddlePaddle/Paddle/pull/3831
• Reviewed and discussed the VarDesc protobuf
  • https://github.com/PaddlePaddle/Paddle/pull/3835
• Reviewed and discussed the axis for mul and rowwise_add
  • https://github.com/PaddlePaddle/Paddle/issues/3722
  • Using num_flatten_dims instead of axis
• Reviewed https://github.com/PaddlePaddle/Paddle/pull/3708

Cao Ying

• GNR in Paddle Models:
  • Tune the hyper-parameters fix the gradient explosion problem. Now the model can converge in 1 pass and runs quite stable. The decoding results look quite good to me (just check them by myself), but still, need an evaluation script, I will add it later.
  • clean codes, add comments, create a PR to Paddle Models:
    • https://github.com/PaddlePaddle/models/pull/220
• some bug fixes in Paddle:
  • Rename the mse_cost into square_error_cost to fix issue #2246
    • https://github.com/PaddlePaddle/Paddle/pull/3845
    • https://github.com/PaddlePaddle/models/pull/222
    • https://github.com/PaddlePaddle/book/pull/370
  • Fix the problem CTC error evaluator cannot print information in v2 API in issue #3802
    • https://github.com/PaddlePaddle/Paddle/pull/3844
  • Fix the problem that long message will be truncated in issue #3809 and issue #2421
    • https://github.com/PaddlePaddle/Paddle/pull/3810
  • Add the missing protostr file for sub_nested_seq_layer to fix issue #3853
    • https://github.com/PaddlePaddle/Paddle/pull/3856
• Port layers:
  • Work on softmax_with_loss_operator, but not finished yet.
  • Update the how to write op doc：
    • https://github.com/PaddlePaddle/Paddle/pull/3873
    • https://github.com/PaddlePaddle/Paddle/pull/3921
  • Update the doc for softmax operator:
    • https://github.com/PaddlePaddle/Paddle/pull/3887
    • when poring layers, I found the doc in our codes are over-simple, so I create this issue to hope we can make a standard to write the doc: https://github.com/PaddlePaddle/Paddle/issues/3885.
  • Fix the naming style and doc for some operators:
    • https://github.com/PaddlePaddle/Paddle/pull/3928

chengduoZhao

• Porting the Vnet model from pytorch to paddle(discussed code optimization)
• Adapting to the cudnnBatchNorm structure for 3D data
  • https://github.com/PaddlePaddle/Paddle/pull/3594
• bug fixes
  • https://github.com/PaddlePaddle/book/pull/357
  • https://github.com/PaddlePaddle/Paddle/pull/3930
• merge
  • https://github.com/PaddlePaddle/book/pull/357
  • https://github.com/PaddlePaddle/Paddle/pull/3504
  • https://github.com/PaddlePaddle/Paddle/pull/3874
• Operators Porting(conv, pool)

helinwang

• Refactor discussion:
  • Dependency graph without Var, to prevent cycle
  • Multiple-GPU design doc review: https://github.com/PaddlePaddle/Paddle/pull/3769#pullrequestreview-60469741
  • Operator based parameter server, design doc discussions: https://github.com/PaddlePaddle/Paddle/pull/3747 (WIP)
  • Should we make everything an OP, design doc discussion: https://github.com/PaddlePaddle/Paddle/pull/3811 (WIP)

Xinghai Sun

• Layer Port
  • Add cos-sim operator (merged: https://github.com/PaddlePaddle/Paddle/pull/3815)
  • Add dropout operator (https://github.com/PaddlePaddle/Paddle/issues/3816)
  • Add cos-sim-vec operator (https://github.com/PaddlePaddle/Paddle/pull/3918)
• DeepSpeech
  • Adjust network structure -- weight sharing (merged: https://github.com/PaddlePaddle/models/pull/224)
  • Clean & reduce codes for configuration arguments (merged: https://github.com/PaddlePaddle/models/pull/227)
  • Reorganize folder structure & hierarchy (merged: https://github.com/PaddlePaddle/models/pull/233)
  • Data preprocess acceleration (https://github.com/PaddlePaddle/models/pull/231)
  • New ctc-decoder review (https://github.com/PaddlePaddle/models/pull/139#pullrequestreview-60832482)
  • Re-write README doc (in progress, 10%)

qijun

Operators Porting:

• Refactor operator python test framework and add sum operator
  • https://github.com/PaddlePaddle/Paddle/pull/3882

LODTensor:

• Refine LODTensor using composition instead of derivation
  • https://github.com/PaddlePaddle/Paddle/pull/3784

Enhance and fix:

• fix clang build and run error
  • https://github.com/PaddlePaddle/Paddle/pull/3919
• make dim int to int64_t
  • https://github.com/PaddlePaddle/Paddle/pull/3900
• fix tensor copyfrom bug
  • https://github.com/PaddlePaddle/Paddle/pull/3788

wanghaoshuang

• Layer Port
  • Refine pad op:
    • https://github.com/PaddlePaddle/Paddle/pull/3765
  • Crop op:
    • https://github.com/PaddlePaddle/Paddle/pull/3906
  • Clip op:
    • https://github.com/PaddlePaddle/Paddle/pull/3937
  • Make attribute support for std::vector<std::pair<int, int>>
    • https://github.com/PaddlePaddle/Paddle/pull/3878

fengjiayi

• Add axis for mul_op and rowwise_add_op
  • https://github.com/PaddlePaddle/Paddle/pull/3888
• Update design doc of batch_norm_op
  • https://github.com/PaddlePaddle/Paddle/pull/3748
• fix:
  • https://github.com/PaddlePaddle/Paddle/pull/3892
  • https://github.com/PaddlePaddle/Paddle/pull/3897
  • https://github.com/PaddlePaddle/Paddle/pull/3940
• code review:
  • https://github.com/PaddlePaddle/Paddle/pull/3827
  • https://github.com/PaddlePaddle/Paddle/pull/3831
  • https://github.com/PaddlePaddle/Paddle/pull/3851
  • https://github.com/PaddlePaddle/Paddle/pull/3878
  • https://github.com/PaddlePaddle/Paddle/pull/3900

qiaolongfei

refactoring

• add VarDesc design
  • https://github.com/PaddlePaddle/Paddle/pull/3835
• add op() to InferShapeContext
  • https://github.com/PaddlePaddle/Paddle/pull/3819
• add var desc proto
  • https://github.com/PaddlePaddle/Paddle/pull/3933
• [wip] Impl vardesc
  • https://github.com/PaddlePaddle/Paddle/pull/3946
• many discussion about python api and the work for next Phase

code review

• Update graph construction design doc
  • https://github.com/PaddlePaddle/Paddle/pull/3862
• Change Op::GetAttr to Op::Attr
  • https://github.com/PaddlePaddle/Paddle/pull/3903
• Make operator Input/Output can return nullptr
  • https://github.com/PaddlePaddle/Paddle/pull/3831

gongweibao

• PR: https://github.com/PaddlePaddle/Paddle/pull/3787
  • caffe2:https://caffe2.ai/docs/operators-catalogue.html#mul

Chen xi

• Keep reading through design docs under mentor of Helin
• Read and understand tensorflow white paper

hedaoyuan

• Operator
  • https://github.com/PaddlePaddle/Paddle/pull/3753
  • Reviewed
    • https://github.com/PaddlePaddle/Paddle/pull/3791
• Mobile
  • https://github.com/PaddlePaddle/Paddle/pull/3792
  • Reviewed
    • https://github.com/PaddlePaddle/Paddle/pull/3843
• Other Reviewed and bug fix
  • https://github.com/PaddlePaddle/Paddle/pull/3348
  • https://github.com/PaddlePaddle/Paddle/pull/3843
  • https://github.com/PaddlePaddle/Paddle/issues/3908

guosheng

• Layer port
  • RowL2NormOp: https://github.com/PaddlePaddle/Paddle/pull/3871.
  • ReduceSum OP and ReduceMean OP in developing.
• Cooperation with Video Analysis Team in IDL:
  • Complete the TSN networks(feature extractor) of their final network by V2 API and model training.
• Review
  • https://github.com/PaddlePaddle/Paddle/pull/3845
  • https://github.com/PaddlePaddle/models/pull/222
  • https://github.com/PaddlePaddle/book/pull/370

dongzhihong

• Refactorization:
  • Dependency graph without Var, refine the graph representation.
  • discuss Multi-GPU design doc
    • https://github.com/PaddlePaddle/Paddle/pull/3769
  • [WIP] NCCL Operator support
    • https://github.com/PaddlePaddle/Paddle/pull/3948
• Reviews
  • Accuracy operator, Topk operator
    • https://github.com/PaddlePaddle/Paddle/pull/3907
  • Add axis for mul_op and rowwise_add_op
    • https://github.com/PaddlePaddle/Paddle/pull/3888
  • Add squared_l2_distance_op
    • https://github.com/PaddlePaddle/Paddle/pull/3768
  • cosine similarity operator
    • https://github.com/PaddlePaddle/Paddle/pull/3815

luotao

• don't need to specify the path of libwarpctc.so, refine the python api doc: https://github.com/PaddlePaddle/Paddle/pull/3783
• op related compiler:
  • Simplify the CMakeLists for paddle/operator and paddle/pybind: https://github.com/PaddlePaddle/Paddle/pull/3801
  • package a new USE_NO_KERNEL_OP for USE_OP_ITSELF: https://github.com/PaddlePaddle/Paddle/pull/3851
  • Op refine for making build system more automatic: https://github.com/PaddlePaddle/Paddle/pull/3881
• [WIP] bilinear_interp_op：implement CPU version

Dang Qingqing

• op porting: https://github.com/PaddlePaddle/Paddle/projects/22
  • 17 PR ( not including many other PRs): https://github.com/PaddlePaddle/Paddle/pulls?q=is%3Apr+is%3Aopen+label%3AOpPorting
    • reviewing: 16
    • merged: 1
  • cudnn wrapper: https://github.com/PaddlePaddle/Paddle/issues/3731 [update and merge]
  • unit testing for gradients: https://github.com/PaddlePaddle/Paddle/issues/3797 [merged]
  • pybind for LoDTensor : WIP
  • op code review
    • norm op: https://github.com/PaddlePaddle/Paddle/pull/3871
    • dropout op: https://github.com/PaddlePaddle/Paddle/pull/3817
    • elementwise-op:https://github.com/PaddlePaddle/Paddle/pull/3787
    • concat-op:https://github.com/PaddlePaddle/Paddle/pull/3775
    • squared-l2-distance: https://github.com/PaddlePaddle/Paddle/pull/3768
    • pad op: https://github.com/PaddlePaddle/Paddle/pull/3765
    • topk: https://github.com/PaddlePaddle/Paddle/pull/3760
    • im2col: https://github.com/PaddlePaddle/Paddle/pull/3753
    • cosine sim op: https://github.com/PaddlePaddle/Paddle/pull/3815
    • Add axis for mul_op and rowwise_add_op: https://github.com/PaddlePaddle/Paddle/pull/3888
    • element-wise op: https://github.com/PaddlePaddle/Paddle/pull/3678

Liu Yiqun

• Update the version of OpenBLAS and fix the compiling error using clang on Mac
  • PR: https://github.com/PaddlePaddle/Paddle/pull/3794
  • issue: OpenBLAS under third_party built failed on Mac #3790
• Add the support of the cross-compiling for iOS Devices
  • update: add the find and link of vecLib to use iOS's optimized BLAS, and add a build_ios task in travis
  • PR: https://github.com/PaddlePaddle/Paddle/pull/3764
• Enable the use of clang to cross-compiling for Android, because the binaries built by clang is faster than that of gcc
  • PR: https://github.com/PaddlePaddle/Paddle/pull/3843
• Port fully connected operator, it is still ongoing
  • PR: https://github.com/PaddlePaddle/Paddle/pull/3927
  • Met some problem
    • The puzzling error message: https://github.com/PaddlePaddle/Paddle/issues/3899
    • The Op's input/output name: https://github.com/PaddlePaddle/Paddle/issues/3922

Yan Chunwei

• paddle refactor
  • PR, update design of RNNOp
  • PR, update design of block
  • PR, add offset implementation to lod tensor design
  • PR, rename LOD to LoD for short of "Level of Details"
  • issue, argue about namespace
  • some discussion with if-else-op
  • much discussion about many thing
  • review
    • https://github.com/PaddlePaddle/Paddle/pull/3835#pullrequestreview-61071261

Yancey1989(yanxu)

• [Reviewing] Concat Op: https://github.com/PaddlePaddle/Paddle/pull/3775
• [Developing] Split Op.
• [Testing] Fix submit predict job on MPI with V1 API.

typhoonzero(wuyi)

• port op: topk, accuracy
• some review:
  • https://github.com/PaddlePaddle/Paddle/pull/3881
• ctr model for feed, large model training problem.

Yibing Liu

Deep Speech 2:

• Evaluate mozilla/DeepSpeech

• Refine C++ decoders

  • https://github.com/PaddlePaddle/models/pull/139

• Code Review:

  • https://github.com/PaddlePaddle/models/pull/233

  • https://github.com/PaddlePaddle/models/pull/227

Paddle Refactor:

• Scaling operator

  • https://github.com/PaddlePaddle/Paddle/pull/3942/files

• [WIP] Reshape operator

  • https://github.com/PaddlePaddle/Paddle/pull/3949

• [WIP] multiplex operator

yangyaming

• DS2:
  • Discuss and review re-organize folder structure
    • https://github.com/PaddlePaddle/models/pull/214#pullrequestreview-60324297
    • https://github.com/PaddlePaddle/models/pull/223#pullrequestreview-60336602
    • https://github.com/PaddlePaddle/models/pull/221#pullrequestreview-60331123
    • https://github.com/PaddlePaddle/models/pull/233#pullrequestreview-60889711
• Operator porting:
  • Refine Squared l2 distance op
    • https://github.com/PaddlePaddle/Paddle/pull/3768
  • Smooth l1 loss op
    • https://github.com/PaddlePaddle/Paddle/pull/3913

Zhuoyuan Chen

Dynamic Conditional Op:

• With the Help of Chunwei and Yi, Design Conditional Op API (https://github.com/PaddlePaddle/Paddle/pull/3828)
• Implemented Conditional Op, now builds well with no error. I will test it in Python soon.