2017 07 12

wangkuiyi

Pthon/C++ interface

• https://github.com/PaddlePaddle/Paddle/issues/2750
• https://github.com/PaddlePaddle/Paddle/pull/2759#discussion_r125996741

Operator

• https://github.com/PaddlePaddle/Paddle/pull/2725#discussion_r126044899
• https://github.com/PaddlePaddle/Paddle/pull/2775#pullrequestreview-48655409

helinwang

• Add Go testing into cmake (CI will now run Go unit tests): https://github.com/PaddlePaddle/Paddle/pull/2811
• Improve Go cmake: https://github.com/PaddlePaddle/Paddle/pull/2832
• Review:
  • Paddle Cloud:
    • https://github.com/PaddlePaddle/Paddle/pull/2719#pullrequestreview-49070191
    • https://github.com/PaddlePaddle/Paddle/pull/2741#pullrequestreview-49278005
    • https://github.com/PaddlePaddle/Paddle/pull/2741#pullrequestreview-49399922
  • cmake:
    • https://github.com/PaddlePaddle/Paddle/pull/2735#issuecomment-312969596
• Fix TeamCity CI.

Gang Liao

• Buddy Memory Allocator Pull Request
• Buddy Allocator Documentation
• CMAKE Generic Documentation

Others:

• go_binary: remove hardcoded library link path, add pserver client test #2832
• FIX: add -lrt for link #2823
• FIX: explicitly specify glog install path #2763

gongweibao

• Paddle Cloud:

  • ISSUE: https://github.com/PaddlePaddle/Paddle/issues/2752
  • Add Go pre-commit support:https://github.com/PaddlePaddle/cloud/pull/204
  • Add task fail interface: https://github.com/PaddlePaddle/Paddle/pull/2719

• PaddlePaddle:

  • Fix cmake error:
    • https://github.com/PaddlePaddle/Paddle/pull/2735
    • https://github.com/PaddlePaddle/Paddle/pull/2745

caoying

• refine machine translation models and fix the problem that training process goes to NaN or explosion.

  • https://github.com/PaddlePaddle/book/pull/339

• modifications to recurrent layer group to output attention weights for each generated sequence in each time step during beam search.

  • this feature is required by OCR team.
  • the codes are finished and under test with colleagues from OCR. I will create a PR later.
  • I am also writing an example to show how to use this feature to Paddle models， but some bugs of V2 APIs are found and I haven't fixed them yet, that this feature cannot be used in V2 API currently.
    • recurrent_group cannot return more than one layer as its outputs: https://github.com/PaddlePaddle/Paddle/issues/2834
    • beam search in the recurrent group gives wrong results by using multi-threads: https://github.com/PaddlePaddle/Paddle/issues/2836

luotao

• RecurrentOp: implement SegmentInput and ConcatOutput function.
• fix bug:
  • pybind compiler question in CentOS: https://github.com/PaddlePaddle/Paddle/pull/2819
  • mistaken: Folk -> Fork: https://github.com/PaddlePaddle/Paddle/pull/2768
• MKL-DNN: find two bugs, create issues, all have been resolved.
  • make fails in master branch: https://github.com/01org/mkl-dnn/issues/87
  • test_reorder fails in v0.9: https://github.com/01org/mkl-dnn/issues/88
• code review:
  • models/SSD: https://github.com/PaddlePaddle/models/pull/137#pullrequestreview-48747183
  • update docs url: https://github.com/PaddlePaddle/Paddle/pull/2786#pullrequestreview-48807140
• wechat PaddlePaddle: 279 fans -> 324 fans

Yu Yang

• Expose paddle.framework C++ --> Python

  • Give Cython+C-API and PyBind11 as demo, we dicided use PyBind11.
  • #2793

• Refine OpRegistry.

  • #2782
  • Defined static variable in .cc
  • Refined C++ syntax
  • Fix static variable init order problems

• Fix slow parsing a recursive depends topolgy in trainer_config_helper

  • #2802

• Refine CUDA Related Libraries.

  • Fix compile error in cuda.h
  • #2806

• Define the interface about OpWithKernel

  • #2815

• [WIP] add a sample op, add_op

  • #2827

• [WIP] Default scope function in Python

  • 2808
  • Need Review

• [TODO] Generate Python OpCreation Code

• [TODO] Python Model concept.

fengjiayi

• finish op_registry development
  • https://github.com/PaddlePaddle/Paddle/pull/2739
• add interfaces for tensor and renew its implementation
  • https://github.com/PaddlePaddle/Paddle/pull/2809
  • https://github.com/PaddlePaddle/Paddle/pull/2813
  • https://github.com/PaddlePaddle/Paddle/pull/2825
• discuss Op design

qijun

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

Operator

• review Operator/OpKernel design
• https://github.com/PaddlePaddle/Paddle/pull/2725#discussion_r126335399

DeviceContext

• Refine DeviceContext
• https://github.com/PaddlePaddle/Paddle/pull/2814

Tensor and Eigen

• Add method converting Tensor to Eigen TensorMap
• https://github.com/PaddlePaddle/Paddle/pull/2805

Some fix

• https://github.com/PaddlePaddle/Paddle/pull/2799
• https://github.com/PaddlePaddle/Paddle/pull/2837

wanghaoshuang

• Image
  • Add pixel softmax layer for FCN model
    • https://github.com/PaddlePaddle/Paddle/pull/2788
  • Add voc2012 dataset for image segment
    • https://github.com/PaddlePaddle/Paddle/pull/2785
  • Add pixel cross entropy cost layer.#2840(WIP)
  • Add FCN model. (WIP)

typhoonzero(wuyi)

• Paddle cloud:
  • new pserver fixes: https://github.com/PaddlePaddle/Paddle/pull/2774
  • Notebook update https://github.com/PaddlePaddle/cloud/pull/210
  • Fault tolerant on paddle cloud https://github.com/PaddlePaddle/cloud/pull/212
  • Job termination: https://github.com/PaddlePaddle/cloud/pull/206
  • Survey on https://github.com/PaddlePaddle/Paddle/wiki/Comparison-of-Distributed-Training-Implementation
  • Meet mellanox, talk about RDMA in deep learning
  • Review https://github.com/PaddlePaddle/Paddle/pull/2674#discussion_r127118366
• MPI:
• reduce job package size by auto removing cuda package when non-GPU training

yangyaming

• DS2
  • Prepare VoxForge dataset
  • Compress LM model
  • Track reason of float overflow
• SSD
  • Performance validation
    https://github.com/PaddlePaddle/models/issues/136
• Other
  • Refine ConvShiftLayer
    https://github.com/PaddlePaddle/Paddle/pull/2133
  • Code Review
    https://github.com/PaddlePaddle/models/pull/154

guosheng

• Code Review:
  • VOC dataset https://github.com/PaddlePaddle/Paddle/pull/2785
• PaddlePaddle Image Enhancement
  • Faster R-CNN: add Region Proposal Network https://github.com/PaddlePaddle/Paddle/pull/2828

Qiaolongfei

Operator

• Add OperatorBase
  • https://github.com/PaddlePaddle/Paddle/pull/2725
• OpKernel Design
  • https://github.com/PaddlePaddle/Paddle/issues/2790
• OpKernel test
  • https://github.com/PaddlePaddle/Paddle/pull/2820
• Add Init for operator
  • https://github.com/PaddlePaddle/Paddle/pull/2838

code review

• Add a sample op, add_op
  • https://github.com/PaddlePaddle/Paddle/pull/2827
• OperatorWithKernel
  • https://github.com/PaddlePaddle/Paddle/pull/2815
• refine device_context
  • https://github.com/PaddlePaddle/Paddle/pull/2814
• Fix slow parsing a recursive depends topology
  • https://github.com/PaddlePaddle/Paddle/pull/2802
• Rnn Opeartor
  • https://github.com/PaddlePaddle/Paddle/pull/2775

Xinghai Sun

• DS2
  • Prepare CHiME-test dataset for evaluating noisy speech (https://github.com/PaddlePaddle/models/pull/157)
  • Experimenting with data augmentation.
• Models
  • Polish model "machine translation with external memory (neural turning machine)" (https://github.com/PaddlePaddle/models/pull/36)
• RL
  • Play with OpenAI-Gym and implement Tabular Q-learning agents.

dongzhihong

• paddle cloud develop

  • continue revise pserver saving checkpoint https://github.com/PaddlePaddle/Paddle/pull/2716

• paddle cloud fix

  • pserver Serivce error https://github.com/PaddlePaddle/Paddle/pull/2839
  • [WIP] pserver return error https://github.com/PaddlePaddle/Paddle/pull/2702

• paddle cloud reviews

  • pserver recover from checkpoint https://github.com/PaddlePaddle/Paddle/pull/2741
  • add go testing to cmake https://github.com/PaddlePaddle/Paddle/pull/2811
  • add pserver client test to go_binary https://github.com/PaddlePaddle/Paddle/pull/2832

• paddle refactoring develop

  • network implement https://github.com/PaddlePaddle/Paddle/pull/2730
  • [WIP] look into the autograd, caffe2 IR module. add GradientOperator to Network Implement.

• paddle refactoring reviews

  • rnn op implement https://github.com/PaddlePaddle/Paddle/pull/2775

Yibing Liu

DS2:

• Refine CTC beam search decoder in C++: performance evaluation, adding the support of loading language model in trie tree.

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

• Distribute the language model

  • https://github.com/PaddlePaddle/models/pull/154

• Improve parameters tuning

  • https://github.com/PaddlePaddle/models/pull/156

hedaoyuan

Fixed two issues of convolution calculation performance last week. At present, the inference of face model with Paddle can reach 400ms, better than their own implementation(750ms).

Convolution Reconstruction and Mobile Optimization

• https://github.com/PaddlePaddle/Paddle/pull/2824
• https://github.com/PaddlePaddle/Paddle/pull/2816

Review

• https://github.com/PaddlePaddle/Paddle/pull/2709#pullrequestreview-48091788
• https://github.com/PaddlePaddle/Paddle/pull/2776#pullrequestreview-48798896
• https://github.com/PaddlePaddle/Paddle/pull/2776#pullrequestreview-49188400
• https://github.com/PaddlePaddle/Paddle/pull/2776#pullrequestreview-49666837

Dang Qingqing

• PaddlePaddle Refactoring
  • implement recurrent group operator
    • https://github.com/PaddlePaddle/Paddle/pull/2775/commits
  • Issue:
    • Whether the OperatorBase needs an initial interface?
      • https://github.com/PaddlePaddle/Paddle/issues/2830
    • The usage of parent scope in the scope design. #2781
      • https://github.com/PaddlePaddle/Paddle/issues/2781
• Code Review
  • https://github.com/PaddlePaddle/Paddle/pull/2771

xingzhaolong

• pr
  • mobilenet gpu acceleration https://github.com/PaddlePaddle/Paddle/pull/2776
• in process
  • mobilent cpu acceleration.

Yancey1989(yanxu)

• Fault Tolerant
  • PServer recovery from checkpint file, https://github.com/PaddlePaddle/Paddle/pull/2741
  • PR review: https://github.com/PaddlePaddle/Paddle/pull/2719
  • Integration test for fault tolerant
• PaddleCloud
  • DLNEL index page: https://github.com/PaddlePaddle/cloud/pull/194
  • Bug fix: https://github.com/PaddlePaddle/cloud/pull/205
  • Bug fix: https://github.com/PaddlePaddle/cloud/pull/202

Yan Chunwei (superjom)

• paddle models

  • DSSM (Deep Semantic Similarity Model)
    • fix first PR

• paddle refactor

  • RecurrentOp with @qingqing and @luotao
    • forward nearly finished
  • reviews
    • https://github.com/PaddlePaddle/Paddle/pull/2725#pullrequestreview-48549349
    • https://github.com/PaddlePaddle/Paddle/pull/2808#pullrequestreview-49177724

• VisualDL with @longfei @yongfeng

  • had a discussion with FE

Liu Yiqun

• Compile Arm‘s ComputeLibrary, and try the calling of gemm implementation
• Benchmark several computing libraries benchmark
• Code review:
  • auto pruning: https://github.com/PaddlePaddle/Paddle/pull/2603
  • android compiling for armv8: https://github.com/PaddlePaddle/Paddle/pull/2816