2017 08 30

weekly meeting

wangkuiyi

• Lod Tensor design doc https://github.com/PaddlePaddle/Paddle/pull/3746
• Add shape inference design doc https://github.com/PaddlePaddle/Paddle/pull/3693
• Add docker build doc https://github.com/PaddlePaddle/Paddle/pull/3676
• Make homepage points to documents in develop branch by default https://github.com/PaddlePaddle/Paddle/issues/3686
• Add a CONTRIBUTING.md file https://github.com/PaddlePaddle/Paddle/issues/3683
• What does it mean WITH_STYLE_CHECK https://github.com/PaddlePaddle/Paddle/issues/3674
• Bug in RecurrentGradienMachine's fillGenOutputs https://github.com/PaddlePaddle/Paddle/issues/3673
• Remove "About" tab in "Documentation" https://github.com/PaddlePaddle/Paddle/issues/3671
• Merge two tabs in our homepage https://github.com/PaddlePaddle/Paddle/issues/3670

Yu Yang

Done

• Writing a skeleton of refactoring notes
  • https://github.com/PaddlePaddle/Paddle/pull/3705
• Done for operator notes
  • https://github.com/PaddlePaddle/Paddle/pull/3719

Doing

• [WIP] Notes for backward
  • https://github.com/PaddlePaddle/Paddle/pull/3738

Should be done in next week

• Notes for sparse matrix and sparse update
• Notes for optimization of GPU memory and kernel

typhoonzero(wuyi)

• paddle cloud fixes and reviews:
  • https://github.com/PaddlePaddle/cloud/pull/359
  • https://github.com/PaddlePaddle/cloud/pull/366
  • https://github.com/PaddlePaddle/cloud/pull/363
  • https://github.com/PaddlePaddle/cloud/issues/340
• topk op: https://github.com/PaddlePaddle/Paddle/pull/3760
• review: https://github.com/PaddlePaddle/Paddle/pull/3747
• feed large model training support

qiaolongfei

refactoring

1. register rowwise add gpu kernel
   https://github.com/PaddlePaddle/Paddle/pull/3660
2. fix mnist, support GPU train https://github.com/PaddlePaddle/Paddle/pull/3694
3. optimizer design.
   https://github.com/PaddlePaddle/Paddle/pull/3711
4. add compile vs runtime discussion
   https://github.com/PaddlePaddle/Paddle/pull/3728
5. why need namespace in python
   https://github.com/PaddlePaddle/Paddle/issues/3749
6. discussion about Scope
7. discussion about InferShape

fengjiayi

• complete document of backward with dongzhihong
  • https://github.com/PaddlePaddle/Paddle/pull/3709
• a survey on bacth_norm_op in Caffe2 and TensorFlow
  • https://github.com/PaddlePaddle/Paddle/issues/3684
• design doc of bacth_norm_op
  • https://github.com/PaddlePaddle/Paddle/pull/3748
• discuss Scope, InferShape and LoDTensor
• review:
  • https://github.com/PaddlePaddle/Paddle/pull/3688
  • https://github.com/PaddlePaddle/Paddle/pull/3697

dongzhihong

• Refactorization:
  • Multi-GPU Survey in mxnet, tensorflow, caffe2&&NCCL2 https://github.com/dzhwinter/Paddle/wiki/Survey-on-Multi-GPU-Feature
  • Multi-GPU design doc https://github.com/PaddlePaddle/Paddle/pull/3769
  • fix backward doc https://github.com/PaddlePaddle/Paddle/pull/3697
• Reviews
  • add compile time vs runtime discussion https://github.com/PaddlePaddle/Paddle/pull/3728
  • refine backward implement https://github.com/PaddlePaddle/Paddle/pull/3725
  • complete backward design doc https://github.com/PaddlePaddle/Paddle/pull/3709
  • docker build doc https://github.com/PaddlePaddle/Paddle/pull/3676

wanghaoshuang

• Add pad op:
  • https://github.com/PaddlePaddle/Paddle/pull/3765
• Add crop op
• Fix img_pool_layer bug:
  • https://github.com/PaddlePaddle/Paddle/pull/3682

qijun

Eigen related:

• [WIP]add tensor sum axis sample
  • https://github.com/PaddlePaddle/Paddle/pull/3774
• add how to use eigen cn doc
  • https://github.com/PaddlePaddle/Paddle/pull/3750

Enhance and fix:

• refine backward
  • https://github.com/PaddlePaddle/Paddle/pull/3725
• fix clang build error
  • https://github.com/PaddlePaddle/Paddle/pull/3762

LODTensor:

• [WIP]replace framework::Tensor with framework::LODTensor
  • https://github.com/PaddlePaddle/Paddle/pull/3793
  • https://github.com/PaddlePaddle/Paddle/issues/3717#issuecomment-325279303

Yancey1989(yanxu)

• [WIP] Concat Op:https://github.com/PaddlePaddle/Paddle/pull/3775
• PaddleCloud on teamCity, https://paddleci.ngrok.io/project.html?projectId=PaddleCloud&tab=projectOverview
• PaddleCloud support suername env, https://github.com/PaddlePaddle/cloud/pull/361
• Bug Fix,https://github.com/PaddlePaddle/cloud/pull/355
• PR Review, https://github.com/PaddlePaddle/cloud/pull/343
• PR Review, https://github.com/PaddlePaddle/cloud/pull/359
• PR Review, https://github.com/PaddlePaddle/Paddle/pull/3747#discussion_r135963936

Yan Chunwei

• paddle refactor
  • PR, design of RNN for fix-length-setence
  • PR, design of block
  • PR, design of namespace
  • update python user interface
  • discuss with other about InferShape, LoTTensor, multi level RNN

helinwang

• Paddle Refactor
  • Discussions
  • Design doc: operator based parameter server. https://github.com/PaddlePaddle/Paddle/pull/3747

Yibing Liu

DS2: CTC beam search decoder in C++

• Finish all the main logic

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

• Retune parameters with a larger acoustic model

  • https://github.com/PaddlePaddle/models/issues/218

  • Current optimal WER 7.87%, reducing 0.8% compared to the decoder in Python

Paddle Refactor:

• [WIP] port scaling layer

xingzhaolong

• assist @daoyuan with the neon implementation of depthwise layer. merged
• help genome group with the Memory leak problem of paddle
• assist @yiqun with the ios cross compile of paddle
• transpose operation implementaion of new paddle, in process
• review
  • https://github.com/PaddlePaddle/Paddle/pull/3718

Dang Qingqing

• PaddlePaddle Refactoring

  • Operators Porting

    • Add a doc about how to write operators in new framework.
      • https://github.com/PaddlePaddle/Paddle/issues/3703
    • Update doc about how to write new operators.
      • https://github.com/PaddlePaddle/Paddle/pull/3734
    • Write some operator code notes
    • Issue: https://github.com/PaddlePaddle/Paddle/issues/3650
    • cuDNN Wrapper
      • https://github.com/PaddlePaddle/Paddle/compare/develop...qingqing01:cudnn_wrapper?expand=1

  • code review:

    • Add im2col functor : https://github.com/PaddlePaddle/Paddle/pull/3753#pullrequestreview-59518905
    • Lod Tensor design doc: https://github.com/PaddlePaddle/Paddle/pull/3746
    • Add skeleton of refactor notes: https://github.com/PaddlePaddle/Paddle/pull/3705#pullrequestreview-58845560

• Other:

  • Code Review:
    • 3DConv/Deconv:
      • https://github.com/PaddlePaddle/Paddle/pull/3348
    • 3DPooling:
      • https://github.com/PaddlePaddle/Paddle/pull/3504

luotao

• refine Huber loss, add huber_regression_cost: https://github.com/PaddlePaddle/Paddle/pull/3571
• [WIP] interpolation op
• document:
  • Remove "About" tab in "Documentation": https://github.com/PaddlePaddle/Paddle/pull/3715
  • add "develop" item to the drop-down menu "Document": https://github.com/PaddlePaddle/portal/pull/3
  • remove unused ubuntu Debian install doc: https://github.com/PaddlePaddle/Paddle/pull/3754
• Intel MKLDNN
  • add MKLDNN_DEVICE, similar as parallel_nn (Code review): https://github.com/PaddlePaddle/Paddle/pull/3712#pullrequestreview-58886485
  • fix download mklml error: https://github.com/PaddlePaddle/Paddle/pull/3761

Chen xi

• Keep reading through design docs under mentor of Helin
• Update learning notes for future new comers’s reference.
• Doc review and update
  • cloud train https://github.com/PaddlePaddle/Paddle/pull/3669
  • docker build https://github.com/PaddlePaddle/Paddle/pull/3676

caoying

• Update the PR cross entropy over beam, and fix some small bug:
  • https://github.com/PaddlePaddle/Paddle/pull/3639
  • https://github.com/PaddlePaddle/Paddle/pull/3679
  • https://github.com/PaddlePaddle/Paddle/pull/3701
• add a python beam decoding script for GNR model
  • https://github.com/lcy-seso/models/tree/GNR/globally_normalized_reader
  • Currently, the model can train and coverage. It gives quite reasonable decoding results on SQuAD reading comprehension task after only one pass training, but I haven't well evaluated it yet, and the model needs better tuning.

yangyaming

• Port operator
  • MSE operator
    • Learn new framework and eigen
    • https://github.com/PaddlePaddle/Paddle/pull/3768
• DS2
  • Continue training unicom

hedaoyuan

• Convolution Operator
  • https://github.com/PaddlePaddle/Paddle/pull/3753 -- doing
• Mobile, Optimize depthwise convolution calculation.
  • DepthwiseConv has a significant acceleration relative to GemmConv, and the mobilenet model on the mobile side can be accelerated by around 30%.
  • https://github.com/PaddlePaddle/Paddle/pull/3718 -- merged
• Bug fix & Review
  • https://github.com/PaddlePaddle/Paddle/pull/3667
  • https://github.com/PaddlePaddle/Paddle/pull/3348

guosheng

• Layer port
  • Read the code of GatedRecurrentLayer and other GRU related code.
• Cooperation with Video Analysis Team in IDL:
  • Communicate about the details of the third module(feature extractor) of their final network.
• Code Review
  • https://github.com/PaddlePaddle/Paddle/pull/3571

Xinghai Sun

• Layer Port: cos_sim op (https://github.com/PaddlePaddle/Paddle/pull/3778)
• DS2 network: https://github.com/PaddlePaddle/models/pull/219
• RL: https://github.com/xinghai-sun/deep-rl/blob/master/docs/selfplay_pong.md

chengduoZhao

• Complete Conv3D, DeConv3D, Pooling3D
  • https://github.com/PaddlePaddle/Paddle/pull/3594
  • https://github.com/PaddlePaddle/Paddle/pull/3504
• Add config parse for Batch_Norm_3D
  • https://github.com/PaddlePaddle/Paddle/pull/3594
• Porting the Vnet model from pytorch to paddle
• Operators Porting(conv, pool)

Zhuoyuan Chen

• CondOp: Much harder than my previous ops... Implementing Condition_op for dynamic if-else. Have done design with Chunwei, will get it done soon.
• Get familiar with pybind11 and Paddle system design.

gongweibao

• PFS:
  • Fix upload and download large file timeout bug.
  • Add upload and download rate info.
  • Add "-chunksize" parameters to speed up upload or download speed.
• CAPI:
  • Package build environment into docker image and put it on teamcity.
  • Modify wiki about CAPI.
• Element-wise multiple operator:
  • Learn how to write operator.
  • Complete code and test it now:https://github.com/gongweibao/Paddle/commit/0a3fd429710b03b9b1aa39d8925bc0ac2264767d

Liu Yiqun

• Add the support of the cross-compiling for iOS devices
  • https://github.com/PaddlePaddle/Paddle/pull/3764