2017 07 26

wangkuiyi

• paddel::memory
  • https://github.com/PaddlePaddle/Paddle/pull/2970#pullrequestreview-51267827
• versioning
  • https://github.com/PaddlePaddle/Paddle/pull/2985#pullrequestreview-51270917
• infer shape
  • https://github.com/PaddlePaddle/Paddle/pull/2958#pullrequestreview-51271848
  • https://github.com/PaddlePaddle/Paddle/pull/3024#pullrequestreview-51935409
• operators
  • https://github.com/PaddlePaddle/Paddle/pull/2965#pullrequestreview-51275598
  • https://github.com/PaddlePaddle/Paddle/pull/3059#pullrequestreview-52434425
  • https://github.com/PaddlePaddle/Paddle/pull/3011#discussion_r129693908
• tensor
  • https://github.com/PaddlePaddle/Paddle/pull/2964#pullrequestreview-51277486
  • https://github.com/PaddlePaddle/Paddle/pull/3014
• net
  • https://github.com/PaddlePaddle/Paddle/pull/2967#pullrequestreview-51539619

helinwang

• Paddle Cloud:
  • Master and Pserver graceful shutdown: https://github.com/PaddlePaddle/Paddle/pull/3062
  • Fix client discover pserver: https://github.com/PaddlePaddle/Paddle/pull/2999
  • Improve logging: https://github.com/PaddlePaddle/Paddle/pull/2997
  • https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-51537224
  • https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-51897010
  • https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-52489318
  • https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-52503132

Gang Liao

• Add CopyFrom for CPU/GPU in Tensor and its Unit Test https://github.com/PaddlePaddle/Paddle/pull/3056
• In memory, for clarity, change DLOG to VLOG https://github.com/PaddlePaddle/Paddle/pull/3057
• Replace pool allocator to buddy allocator in Paddle https://github.com/PaddlePaddle/Paddle/pull/3030
• Qianmo Docker for internal users

Review:

• Make PADDLE_ENFORCE and PADDLE_THROW catchable https://github.com/PaddlePaddle/Paddle/pull/3055
• Fix memory deps on device context https://github.com/PaddlePaddle/Paddle/pull/3018
• Move Copy out from memory.h into memcpy.h https://github.com/PaddlePaddle/Paddle/pull/3016

Yancey1989(yanxu)

• Fault Tolerant
  • PR Review: https://github.com/PaddlePaddle/Paddle/pull/3053#discussion_r129291480
  • PR Review: https://github.com/PaddlePaddle/cloud/pull/212#pullrequestreview-50896371
  • PR Review: https://github.com/PaddlePaddle/Paddle/pull/2948
• Paddle version API: https://github.com/PaddlePaddle/Paddle/pull/2985
• MPI
  • Paddle on MPI P40 cluster, done with tcp mode.
• PaddleCloud
  • Release 0.1.1-beta.3
  • update doc: https://github.com/PaddlePaddle/cloud/pull/248
  • bug fix: https://github.com/PaddlePaddle/cloud/pull/234

Yu Yang

• Working in process a major PR Backward with Feng jiayi, Dong Zhihong.

  • framework::Backward can generate Backward Operators of a Fast-Forward network.
    • It can ignore any gradient variable in forward operator.
    • It can handle shared variable's gradient
  • https://github.com/PaddlePaddle/Paddle/pull/3068

• Remove ScopePtr, OpPtr,

  • https://github.com/PaddlePaddle/Paddle/pull/3027

• Fix a bug in SequenceActivation.

  • https://github.com/PaddlePaddle/Paddle/pull/3058

• Make framework::Net -> framework::NetOp, remove framework::PlainNet

  • Also add common using header in operators module.
  • https://github.com/PaddlePaddle/Paddle/pull/3067

• Make PADDLE_ENFORCE throw same type exception EnforceNotMet

  • https://github.com/PaddlePaddle/Paddle/pull/3055

• Fix a import error when WITH_GOLANG=OFF

  • https://github.com/PaddlePaddle/Paddle/pull/3045

Caoying

• Some bug fix:

  • https://github.com/PaddlePaddle/Paddle/pull/3071
  • https://github.com/PaddlePaddle/Paddle/pull/3025
  • https://github.com/PaddlePaddle/Paddle/pull/3071

• About Paddle Models:

  • fix some diff with previous v1 version.
  • after the bug of sequence_softmax has been fixed, the training process works well. I trained the model again. It now can generate interesting Chinese poetry correctly. I will add the documentation and update it to Paddle models later.

• About GNR:

  • Read some paper on learning to search.
  • make clear every computation details of GNR model.
    • I will add four new layers: subsequence_select, sequence_slice, cross_entroy_over_beam (including decoding)
    • subsequence_select is in progress
  • work with Andrew to first make some simplified model running. https://github.com/lcy-seso/paddle_experiment

Yan Chunwei

• paddle refactor

  • RecurrentOp implementation with @qingqing01 @luotao @longfei
    • https://github.com/PaddlePaddle/Paddle/pull/2890
    • python unittest with forward completed
    • ready to merge the first version
  • design of RecurrentOp with varient length sequence support with @luotao @longfei @qingqing01
    • https://github.com/PaddlePaddle/Paddle/pull/3011

• others

  • visualDL, add design of data structures
    • https://github.com/VisualDL/VisualDL/pull/9
  • Community construction
    • paddle blog with @helin ready.
      • http://blog.paddlepaddle.org/
      • write blog and PR to https://github.com/PaddlePaddle/blog/tree/develop

luotao

• RecurrentOp:
  • clean the code
  • use the latest add_op and mul_op to test the forward
  • remove ScopePtr/OperatorPtr
• deep speech2:
  • Make setup.sh more robust: https://github.com/PaddlePaddle/models/pull/177
  • Follow-up Intel Optimization on DeepSpeech2 with @liuyibing
• code review:
  • design doc of variable sequence
  • https://github.com/PaddlePaddle/Paddle/pull/3009#pullrequestreview-51414637
  • https://github.com/PaddlePaddle/Paddle/pull/2965#pullrequestreview-51376698

qijun

Operator --> OpKernel --> Tensor/DeviceContext --> Eigen

Implement some basic OpKernels and add python unittest

• https://github.com/PaddlePaddle/Paddle/pull/2971

[WIP]enable operator gpu unittest

• https://github.com/PaddlePaddle/Paddle/pull/3050

[WIP] Elementwise Operator Functor

• https://github.com/PaddlePaddle/Paddle/pull/3029

fix bug in OpKernel register macro

• https://github.com/PaddlePaddle/Paddle/pull/3059

Qiaolongfei

refactoring

• use operator context and infer context
  • https://github.com/PaddlePaddle/Paddle/pull/3024
• 讨论RNN变长输入设计
  • https://github.com/PaddlePaddle/Paddle/pull/3011
• 讨论ViaulDL一些事情

code review

• fix bug in OpKernel register macro
  • https://github.com/PaddlePaddle/Paddle/pull/3059
• Add type_alias to import framework into ops
  • https://github.com/PaddlePaddle/Paddle/pull/3048
• Make network op
  • https://github.com/PaddlePaddle/Paddle/pull/3067
• Make CreateOp in Plain C++ params
  • https://github.com/PaddlePaddle/Paddle/pull/2936
• Python Generate OpCreation Methods by OpProto
  • https://github.com/PaddlePaddle/Paddle/pull/2893

hedaoyuan

• PaddlePaddle Mobile
  • Face Model
    Multithreading performance optimization for recognition model prediction(doning). Study of the face detection model deployed based on Paddle, and paddle is missing slice layer(todo).
  • https://github.com/PaddlePaddle/Paddle/issues/3034
  • https://github.com/PaddlePaddle/Paddle/issues/3037
  • https://github.com/PaddlePaddle/Paddle/pull/2449 -- merge
• Bug fix
  • https://github.com/PaddlePaddle/Paddle/pull/3044

gongweibao

• PaddleCloud
  • PR:
    • Integration testing pfs and pcloud :https://github.com/PaddlePaddle/cloud/pull/148
  • Fix bugs:
    • env bugs:https://github.com/PaddlePaddle/cloud/pull/249
    • Chinese character bugs:https://github.com/PaddlePaddle/cloud/pull/240
    • Fix docs:https://github.com/PaddlePaddle/cloud/pull/250
• PaddlePaddle:
  • Embed etcd into Go for easy unit test: https://github.com/PaddlePaddle/Paddle/issues/2504
• Code review:
  • https://github.com/PaddlePaddle/Paddle/pull/2948
  • https://github.com/PaddlePaddle/Paddle/pull/3053
  • https://github.com/PaddlePaddle/cloud/pull/248

dongzhihong

• Refactor:
  • Add Backward Operator registry : https://github.com/PaddlePaddle/Paddle/pull/2949
  • Add Random Operator https://github.com/PaddlePaddle/Paddle/pull/3060
  • Feature/Backward https://github.com/PaddlePaddle/Paddle/pull/3068
• bug fix:
  • fix gen data shape https://github.com/PaddlePaddle/Paddle/pull/3063
• reviews
  • polish gradient registry https://github.com/PaddlePaddle/Paddle/pull/3036
  • implement distribute save model https://github.com/PaddlePaddle/Paddle/pull/2973

fengjiayi

• backward
  • Update gradient op registry mechanism. With @dongzhihong
    • https://github.com/PaddlePaddle/Paddle/pull/3036
    • https://github.com/PaddlePaddle/Paddle/pull/2949
  • [WIP] Refine depends of op_registry to make it support general DAG net. With @yuyang and @dongzhihong
• review:
  • https://github.com/PaddlePaddle/Paddle/pull/3056
  • https://github.com/PaddlePaddle/Paddle/pull/3060

chengduoZhao

• Learn the design philosophy and overall architecture of paddle
• Learn the design patterns of Op in paddle
  • Contrast the differences in Op design between paddle and caffe2
  • Try to update the design document for Op

Dang Qingqing

• PaddlePaddle Refactoring
  • RNN op
    • forward InferShape https://github.com/PaddlePaddle/Paddle/pull/2890/commits/74fd16cf1b6e22ef1dad389c2fa4fea22cf3a3b4
• DS2
  • WarpCTC debug with Yaming: https://github.com/PaddlePaddle/Paddle/issues/3076
  • Bug when Multi-GPU infering: https://github.com/PaddlePaddle/Paddle/issues/3073
• Two days off

wanghaoshuang

• FCN
  • Debug and tune FCN model demo[WIP]
• Fix bugs[merged]
  • https://github.com/PaddlePaddle/Paddle/pull/3043
• DeepSpeech2
  • Profiling of DS2 on paddle cloud
    • https://github.com/PaddlePaddle/models/issues/179
  • Refine DS2 paddle cloud submit scripts

Yibing Liu

DS2:

• Train the model based on mfcc audio feature
• Profile the whole training pipline (with@lutao)
• Compare the beam search decoder with that in speech-dl (with@yaming)
• Refine the decoder for deployment: add best_path_decoder, unify the interface
  • https://github.com/PaddlePaddle/models/pull/139

yangyaming

• DS2:
  • Debug WarpCTCLayer with @qingqing (WIP)
    https://github.com/PaddlePaddle/Paddle/issues/3076
  • Wrap python interface for decoder of dl_speech (Done)
  • Process 3k hours mandarin data (WIP)
  • Help to validate performance for latest PaddlePaddle version (Done)
• Code Review:
  • https://github.com/PaddlePaddle/Paddle/pull/3044

xingzhaolong

• pr

  • https://github.com/PaddlePaddle/Paddle/pull/3009 add param_attr for img_conv_groups(...) in /python/trainer_config_helpers/networks.py

• investigate

  • ncnn (a light forward framework)

livc(Zhao Li)

• survey on supporting Genome Group to use paddle with Guo Sheng.
• bug fix https://github.com/PaddlePaddle/Paddle/pull/2992

typhoonzero(wuyi)

PaddleCloud

• Sync master client between passes and fix recordio split https://github.com/PaddlePaddle/Paddle/pull/2948
• paddlecloud domain bug fixes
• dlnel notebook release
• dlnel 10*P40 machines deploy WIP

guosheng

• PaddlePaddle Image Enhancement

  • Faster R-CNN: fix bug for the before PR, including ROIPoolLayer and ProposalTargetLayer; Complete the RCNNLossLayer:
    • https://github.com/PaddlePaddle/Paddle/pull/2982
    • https://github.com/PaddlePaddle/Paddle/pull/2983

• Cooperation with Video Analysis Team in IDL:

  • Communication about the model for ActivityNet2017, and some enhancements are needed to satisfy the model. Layers for L2-normalization and clipping will be added. RowL2NormLayer is in developing.

Xinghai Sun

• DS2: task list, experiments, detail verification, and bug fixing.
  • https://github.com/PaddlePaddle/models/issues/176
  • https://github.com/PaddlePaddle/models/issues/178
  • https://github.com/PaddlePaddle/Paddle/issues/3073