2017 07 19

wangkuiyi

• Expose Tensor to Python
  • https://github.com/PaddlePaddle/Paddle/pull/2928#pullrequestreview-50416254
  • https://github.com/PaddlePaddle/Paddle/pull/2928#pullrequestreview-50444255
• Rewrite Tensor
  • https://github.com/PaddlePaddle/Paddle/pull/2953
• PADDLE_ENFORCE
  • https://github.com/PaddlePaddle/Paddle/pull/2883#pullrequestreview-50436482
  • https://github.com/PaddlePaddle/Paddle/pull/2883#pullrequestreview-50450264
• Python style
  • https://github.com/PaddlePaddle/Paddle/pull/2785#pullrequestreview-50443488

gongweibao

• Do the integration test of PaddleCloud and fix bugs:
  • https://github.com/PaddlePaddle/cloud/pull/225
  • https://github.com/PaddlePaddle/cloud/pull/211
  • https://github.com/PaddlePaddle/cloud/pull/148
• Code review:
  • https://github.com/PaddlePaddle/Paddle/pull/2885
  • https://github.com/PaddlePaddle/Paddle/pull/2848
  • https://github.com/PaddlePaddle/cloud/pull/223
  • https://github.com/PaddlePaddle/cloud/pull/231
• other:
  • Repair kubernetes system on America machines

Gang Liao

• Add CPU/GPU Memcpy in memory folder https://github.com/PaddlePaddle/Paddle/pull/2970
• FIX: cpplint code style https://github.com/PaddlePaddle/Paddle/pull/2888
• ENH: unify PADDLE_ENFORCE for cublas, cudnn, curand https://github.com/PaddlePaddle/Paddle/pull/2883

Review:

• cmake: fix problem that go_library is never rebuilt. #2954
• enable Intel MKLDNN library and Intel MKL small package #2940

helinwang

• PR:
  • Paddle Cloud:
    • Save model: https://github.com/PaddlePaddle/Paddle/pull/2973
    • Fix optimizer: https://github.com/PaddlePaddle/Paddle/pull/2855
    • Make save model examples no longer compress model: https://github.com/PaddlePaddle/book/pull/342
  • Go, cmake and CI:
    • Turn on race detector in CI for Go: https://github.com/PaddlePaddle/Paddle/pull/2856
    • Reduce cmake build time for Go: https://github.com/PaddlePaddle/Paddle/pull/2884
    • Add Go linter pre commit hook: https://github.com/PaddlePaddle/Paddle/pull/2885
    • Fix travis Go pre commit check: https://github.com/PaddlePaddle/Paddle/pull/2887
    • cmake: fix problem that go_library is never rebuilt: https://github.com/PaddlePaddle/Paddle/pull/2954
• Review:
  • Paddle Cloud:
    • https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-51000329
    • https://github.com/PaddlePaddle/Paddle/pull/2072#pullrequestreview-50478810
    • https://github.com/PaddlePaddle/Paddle/pull/2849#pullrequestreview-50220786
    • https://github.com/PaddlePaddle/Paddle/pull/2848#pullrequestreview-49864606
    • https://github.com/PaddlePaddle/Paddle/pull/2869#pullrequestreview-50126894
  • Refactor:
    • https://github.com/PaddlePaddle/Paddle/pull/2972#pullrequestreview-50989251

Qiaolongfei

refactoring

• change net to operator
  • https://github.com/PaddlePaddle/Paddle/pull/2846
• Optimize ptr
  • https://github.com/PaddlePaddle/Paddle/pull/2851
• Add Init to OperatorBase
  • https://github.com/PaddlePaddle/Paddle/pull/2838
• Add sgd op
  • https://github.com/PaddlePaddle/Paddle/pull/2950
• add ADD_OP_CPU to enable add op with only cpu kernel
  • https://github.com/PaddlePaddle/Paddle/pull/2896
• check duplicate of ProtoAndCheckerMaker
  • https://github.com/PaddlePaddle/Paddle/pull/2903
• support multiple template parameter for REGISTER_OP_XPU
  • https://github.com/PaddlePaddle/Paddle/pull/2932
• [wip]add cross-entropy-op
  • https://github.com/PaddlePaddle/Paddle/pull/2965
• [wip]use var name to get variable instead of index
  • https://github.com/PaddlePaddle/Paddle/issues/2980

code review

• Feature/add op test
  • https://github.com/PaddlePaddle/Paddle/pull/2947
• Skeleton Of fully connected operator
  • https://github.com/PaddlePaddle/Paddle/pull/2945
• 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

Yu Yang

• Make Python can create variable from a default scope.
  • PR #2808
• Design of OperatorWithKernel class
  • PR #2815
• Add a sample operator skeletons, add_op
  • PR #2827
• Make OpProto's VarProto support multiple and temporary.
  • PR #2860
• Python OpCreation methods
  • PR #2864
  • PR #2893
• Polish operator cmake file, add op_library function
  • PR #2897
• Override Net::DebugString to make it better for reading.
  • PR #2912
• Add skeletons of mul, rowwise_add, sigmoid, softmax ops
  • PR #2914
• Make Python can create, read, write Tensor
  • PR #2928
• Change Others PR to make them better
  • PR #2935 Use shared_ptr for in_out_idxs_.
  • PR #2939 Tensor should use PADDLE_ONLY_CPU, not CUDACC
• Make CreateOp can be invoked by plain C++ params
  • PR #2936
• Add fully connected operator.
  • PR #2945
  • FullyConnectedOperator is not a plain operator with kernel. It is a composed/grouped operator.
• Add a operator unittest framework in Python
  • Developer only need to write how numpy calculate this operator.
  • PR #2947, 2968

Yan Chunwei

• paddle refactor

  • RNN Operator with @qingqing @luotao

    • https://github.com/PaddlePaddle/Paddle/pull/2890
    • C++ code finished
    • writing unittest in python
    • waiting for some other dependency operators to be ready.

  • Named getter for operator arguments

    • https://github.com/PaddlePaddle/Paddle/pull/2901
    • to make argument retrive more convenient
    • to hide details of arguments with multiple inputs

  • PADDLE_ENFORCE add a switch for glog/CHECK

    • https://github.com/PaddlePaddle/Paddle/pull/2850
    • to better display the stack info when ENFORCE failed.

• paddle models

  • DSSM

    • merged
      • https://github.com/PaddlePaddle/models/pull/141
    • an enhancement
      • https://github.com/PaddlePaddle/models/pull/172
      • add infer.py
      • add more command line arguments support to help configure model structure
      • fix bugs

  • CTR model enhancement

    • https://github.com/PaddlePaddle/models/pull/138
    • decouple data format and model structure
    • add infer script
    • support classification/regression

• Others

  • http://models.paddlepaddle.org site

    • check all html pages and fix some models markdown file format
    • support math equation display
    • add a navigation bar

  • visualDL with @longfei @yongfeng

    • discussion with FE
    • determine what features first version support.

qijun

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

Implement some basic OpKernels

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

Tensor to EigenTensor

• https://github.com/PaddlePaddle/Paddle/pull/2805
• https://github.com/PaddlePaddle/Paddle/pull/2953
• https://github.com/PaddlePaddle/Paddle/pull/2978

Dang Qingqing

• PaddlePaddle Refactoring
  • refactor recurrent group forward operator with @chunwei @luotao
  • implement recurrent group backward operator with @chunwei @luotao
    • https://github.com/PaddlePaddle/Paddle/pull/2890
  • Code Review
    • Add Tensor::CopyFrom and Tensor::mutable_data(Place place)
      • https://github.com/PaddlePaddle/Paddle/pull/2825#pullrequestreview-49752498
    • Op varient inputs
      • https://github.com/PaddlePaddle/Paddle/pull/2901#pullrequestreview-50213829
• Code Review
  • FCN
    • add crop layer
      • https://github.com/PaddlePaddle/Paddle/pull/2490#pullrequestreview-50304712
    • Add pixel softmax layer for FCN model
      • https://github.com/PaddlePaddle/Paddle/pull/2788#pullrequestreview-50328846
    • Add cross entropy layer for pixel classification:
      • https://github.com/PaddlePaddle/Paddle/pull/2876#issuecomment-315784515
    • Add voc2012 dataset for image segment
      • https://github.com/PaddlePaddle/Paddle/pull/2785#pullrequestreview-50356316
  • models: SSD
    • https://github.com/PaddlePaddle/models/pull/137#pullrequestreview-50597896

xingzhaolong

• Mobilenet

  • optimize the gradInput in the gpu implementation
  • support inputchannels != outputChannels
  • naive forward implementaion

• investigate

  • shuffleNet

• review

  • https://github.com/PaddlePaddle/Paddle/pull/2977#pullrequestreview-51097570

luotao

• RecurrentOp: update the SegmentInput and ConcatOutput function, discuss some functions(slice_ddims, etc) in dims.h with @fengjiayi.
• book:
  • add video link in chinese version: https://github.com/PaddlePaddle/book/pull/341
• code review:
  • SSD model:
  • https://github.com/PaddlePaddle/Paddle/pull/2865#pullrequestreview-49982208
  • Cluster train doc for v2 API: https://github.com/PaddlePaddle/Paddle/pull/2072#pullrequestreview-50290075
• update yixiu with @lizhao
• wechat PaddlePaddle: 324 fans -> 389 fans

caoying

• Paddle
  • Add a gated unit helper:
    • https://github.com/PaddlePaddle/Paddle/pull/2865
  • enable recurrent group to output a sequence inside it during generation:
    • https://github.com/PaddlePaddle/Paddle/pull/2871
  • fix recurrent_group parsing bug:
    • https://github.com/PaddlePaddle/Paddle/pull/2944
• Models
  • Add a trained text generation model to generate Chinese poetry into Paddle Models. Not finished yet. The learning performance of model trained by V2 API is not comparable with that triained by v1 API. Still checking it.
    • https://github.com/PaddlePaddle/models/compare/develop...lcy-seso:gen_chinese_poetry
  • PR review: https://github.com/PaddlePaddle/models/pull/36#pullrequestreview-50802703

livc(Zhao Li)

• Add models repo automated test. PR:
• https://github.com/PaddlePaddle/models/pull/167
• https://github.com/PaddlePaddle/models/pull/170

Yancey1989(Yanxu)

• Fault Tolerant
  • Fix read record failed,https://github.com/PaddlePaddle/Paddle/pull/2848
• PaddleCloud
  • Fix notebook session request 404
  • PR Review: https://github.com/PaddlePaddle/cloud/pull/219#pullrequestreview-50250328
  • PR Review: https://github.com/PaddlePaddle/cloud/pull/212#pullrequestreview-50896371
  • Bug fix, https://github.com/PaddlePaddle/cloud/pull/231
  • Bug fix, https://github.com/PaddlePaddle/cloud/pull/227
• MPI
  • [WIP] Paddle on P40 MPI cluster.

hedaoyuan

• PaddlePaddle Mobile
  • The current progress, integrating Paddle into the face SDK. Performance issues, in the mobile big.LITTLE architecture, multi-threaded computing performance is not as high as expected. And at present, on the Huawei P9 multi-threaded performance test has not yet reached the requirements (less than 450ms).
  • https://github.com/PaddlePaddle/Paddle/pull/2816 -- merge
  • https://github.com/PaddlePaddle/Paddle/pull/2824 -- merge
• Bug fix
  • https://github.com/PaddlePaddle/Paddle/pull/2862
• Review
  • Mobile:
    • https://github.com/PaddlePaddle/Paddle/pull/2776
    • https://github.com/PaddlePaddle/Paddle/pull/2874
    • https://github.com/PaddlePaddle/Paddle/pull/2845
  • Other:
    • https://github.com/PaddlePaddle/Paddle/pull/2940

wanghaoshuang

• FCN
  • Add switch order layer for FCN model
    • https://github.com/PaddlePaddle/Paddle/pull/2788
  • Add voc2012 dataset for image segment[merged]
    • https://github.com/PaddlePaddle/Paddle/pull/2785
  • Refine crop layer[merged]
    • https://github.com/PaddlePaddle/Paddle/pull/2490
  • Debug and tune FCN model demo[WIP]
• Fix bugs[merged]
  • https://github.com/PaddlePaddle/Paddle/pull/2977
• Code review
  • https://github.com/PaddlePaddle/Paddle/pull/2828#pullrequestreview-49705565

fengjiayi

• Add Interfaces for Tensor and simply is implementation
  • https://github.com/PaddlePaddle/Paddle/pull/2891
  • https://github.com/PaddlePaddle/Paddle/pull/2900
  • https://github.com/PaddlePaddle/Paddle/pull/2926
  • https://github.com/PaddlePaddle/Paddle/pull/2964
• [WIP] Backward implementation with @zhihong
  • https://github.com/PaddlePaddle/Paddle/pull/2949

dongzhihong

• PR:
  • Refactor:
    • Add Backward Operator: https://github.com/PaddlePaddle/Paddle/pull/2949
  • Fixes:
    • fix unrar package in docker: https://github.com/PaddlePaddle/Paddle/pull/2902
    • add start script for ltr model : https://github.com/PaddlePaddle/models/pull/173
    • fix dense shape in dataset: https://github.com/PaddlePaddle/Paddle/pull/2905
    • add skip layer of loading model interface: https://github.com/PaddlePaddle/Paddle/pull/2904
    • fix checkpoint init error: https://github.com/PaddlePaddle/Paddle/pull/2839
• Review:
  • Refactor:
    • fully connected op: https://github.com/PaddlePaddle/Paddle/pull/2945
    • sgd operator: https://github.com/PaddlePaddle/Paddle/pull/2950
  • Paddle Cloud:
    • add go testing into cmake: https://github.com/PaddlePaddle/Paddle/pull/2811
    • Implement distribute save model: https://github.com/PaddlePaddle/Paddle/pull/2973
    • add gometaliner checker: https://github.com/PaddlePaddle/Paddle/pull/2885
    • fix optimizer buffer allocation size: https://github.com/PaddlePaddle/Paddle/pull/2855
    • race detector on golang https://github.com/PaddlePaddle/Paddle/pull/2856

guosheng

• PaddlePaddle Image Enhancement
  • Faster R-CNN
    • ROIPoolLayer https://github.com/PaddlePaddle/Paddle/pull/2982
    • ProposalTargetLayer https://github.com/PaddlePaddle/Paddle/pull/2983
    • RCNNLossLayer

Xinghai Sun

• DS2
  • Decayed ADAM (-1.3 WER), decoder refinement (with @Yibing), Mandarin training debug (with @Yaming)
  • PR Review https://github.com/PaddlePaddle/models/pull/168#pullrequestreview-51093780
• Investigating ConvS2S, Transformer, SliceNet.

Yibing Liu

DS2:

• Add the support of mfcc feature and retrain the model

  • https://github.com/PaddlePaddle/models/pull/168

• CTC Beam search decoder in C++: covert the compuation of probs in log format

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

Conv seq2seq:

• Read related papers

yangyaming

• Align LM scoring logic with @yibing
• Cooperate with @yiqun to find numerical bug for wctc
• Polish SSD demo
  https://github.com/PaddlePaddle/models/pull/137
• Correct formation for nce_cost/train.py
  https://github.com/PaddlePaddle/models/pull/169
• Code review
  • https://github.com/PaddlePaddle/models/pull/168#pullrequestreview-51088968
  • https://github.com/PaddlePaddle/models/pull/157#pullrequestreview-51090221

typhoonzero(wuyi)

• Done Fault tolerant on paddle cloud: https://github.com/PaddlePaddle/cloud/pull/212
• Fault tolerant job task fixes: https://github.com/PaddlePaddle/Paddle/pull/2948
• Enable notebook access dlnel.org: https://github.com/PaddlePaddle/cloud/pull/219
  • BFE dosn’t support websocket need a workaround
• Clusters: Formal cluster for dlnel on Baidu Cloud, CPU cluster resource exchange request
• MPI cluster mantainous

Liu Yiqun

• Add build_android task on Travis CI
  • https://github.com/PaddlePaddle/Paddle/pull/2845
• Update the dockerfile to build cross-compiling image for Android.
  • https://github.com/PaddlePaddle/Paddle/pull/2874
• Fix a bug of warp-ctc, that cause float-pointing exception
  • https://github.com/gangliao/warp-ctc/pull/3
• Search for a toolchain to cross-compiling Paddle for iOS device on Linux
• genome want to use Paddle instead of caffe in their Android application, and they have some demands, https://github.com/PaddlePaddle/Paddle/issues/2966
  • Test the inference speed of Paddle
• Code Review
  • nnpack, https://github.com/PaddlePaddle/Paddle/pull/2824
  • mkldnn, https://github.com/PaddlePaddle/Paddle/pull/2940

juliecbd

• Pull request

  • https://github.com/PaddlePaddle/Paddle/pull/2907

• Issue

  • https://github.com/PaddlePaddle/Paddle/issues/2908
  • https://github.com/PaddlePaddle/Paddle/issues/2906