2018 05 30

tonyyang-svail

• Automatical Differentiation with the Tape: https://github.com/PaddlePaddle/Paddle/pull/11019

Lei Wang

• Fix build related documents after scripts clean up
  • https://github.com/PaddlePaddle/Paddle/pull/10721
• Set up CI for paddlepaddle/PARL repo ( Travis and Teamcity work now)
  • https://github.com/PaddlePaddle/PARL/pull/10
• Fix dependencies for test_paddle_inference_api_impl
  • https://github.com/PaddlePaddle/Paddle/pull/11064
• Fix teamcity ccache

luotao

• inference engine:
  • refine fluid deploy on teamcity: https://github.com/PaddlePaddle/Paddle/pull/10904
  • obtain the correct git commit id in teamcity with @yanxu https://github.com/PaddlePaddle/Paddle/pull/10991
  • fix compiler error when do not have TensorRT library https://github.com/PaddlePaddle/Paddle/pull/11051
• MKLDNN:
  • request for the static library of MKLML with intel @huying.
  • help OCR to fix (doing): when use fluid version with MKLML shared library, other CPU jobs (with MKL static libary) on the same machine core, other GPU jobs become slower.
• code review:
  • mkldnn
    • [merge] Blas optimized elementwise_add forward and backward passes (10% speedup on elemtwise_add op of OCR CRNN_CTC model): https://github.com/PaddlePaddle/Paddle/pull/10913
    • [merge] Top K algorithm parallel version: https://github.com/PaddlePaddle/Paddle/pull/10941
    • [merge] Withdraw MKLDNN Mul operator: https://github.com/PaddlePaddle/Paddle/pull/10703#pullrequestreview-124659139
  • https://github.com/PaddlePaddle/Paddle/pull/10934
  • https://github.com/PaddlePaddle/Paddle/pull/10988
  • https://github.com/PaddlePaddle/Paddle/pull/10841

gongweibao

• Move sync_mode device ctx from grpc server
  • https://github.com/PaddlePaddle/Paddle/pull/10881

fengjiayi

• Test ReacordIO Reader with Resnet50(P40):

  • 1 GPU:139 imgs/sec 8 GPU:896 imgs/sec
  • speed-up ratio: 6.44

• Training Data Preprocessor:

  • https://github.com/PaddlePaddle/Paddle/pull/10872

• A preprocessing op(random crop op):

  • https://github.com/PaddlePaddle/Paddle/pull/10970

wanghaoshuang

• Slice op.
  • https://github.com/PaddlePaddle/Paddle/pull/11052
• Add shape op to get the shape of a variable.
  • https://github.com/PaddlePaddle/Paddle/pull/11048
• Make bilinear_interp_op support attrs from the input.
  • https://github.com/PaddlePaddle/Paddle/pull/11041
• Add python wrapper for gather op.
  • https://github.com/PaddlePaddle/Paddle/pull/11033
• Discussion of polygon NMS op :
  • https://github.com/PaddlePaddle/Paddle/issues/10935
• Review:
  • https://github.com/PaddlePaddle/Paddle/pull/10923

Chenxi

• aws integration with CE https://github.com/PaddlePaddle/paddle-ce-latest-kpis/pull/27
  • now it supports multi cluster config to work with KPI matrix introduced by #10944
  • cluster spec chunk execution to make the most use of aws quota
  • aws_runner now works with /benchmark/fluid
  • going to integrate /benchmark/fluid to CE

dongzhihong

• memory optimize
  • [WIP]a MxNet like test of memory optimize
    • https://github.com/PaddlePaddle/Paddle/pull/11049
  • accelerate the reduce op compile
    • https://github.com/PaddlePaddle/Paddle/pull/11029
  • reduce the Se-ResNeXt-152 ops
    • https://github.com/PaddlePaddle/Paddle/pull/10977
  • fix build scripts
    • https://github.com/PaddlePaddle/Paddle/pull/11011

Bai Yifan

• [WIP] Reading face detection caffe code
• issue
  • https://github.com/PaddlePaddle/models/issues/937
• PR
  • https://github.com/PaddlePaddle/models/pull/945
  • https://github.com/PaddlePaddle/Paddle/pull/10925
  • https://github.com/PaddlePaddle/Paddle/pull/11067

guosheng

• NMT:
  • Fix and enhance beam_search_op and beam_search_decode_op (WIP).
  • Continue the work on beamsearch decoder in Transformer (WIP).
    • https://github.com/PaddlePaddle/models/pull/915
  • Validate the beam-search related fix by comparing with original python beam searh

tangwei

• Incremental Learning Support
  • https://github.com/PaddlePaddle/Paddle/pull/10878
• update fluid Train API param_path to checkpoint_config
  • https://github.com/PaddlePaddle/Paddle/pull/10878
• checkpoint feature on Paddle Cloud

Dang Qingqing

• Face Detection:
  • Fix bug in model and data sampling .
    • https://github.com/PaddlePaddle/Paddle/pull/11005
    • https://github.com/PaddlePaddle/models/pull/939
    • https://github.com/PaddlePaddle/models/pull/940
    • https://github.com/PaddlePaddle/models/pull/944
    • https://github.com/PaddlePaddle/models/pull/947
• Fixed-point quantization:
  • Develop a fake dequantized op for fixed-point quantization training framework.
    • https://github.com/PaddlePaddle/Paddle/pull/10965
  • Update design doc of fixed-point quantization.
    • https://github.com/PaddlePaddle/Paddle/pull/10553
• Others:
  • Fix fill_constant_batch_size_like_op when input is LoDTensor.
    • https://github.com/PaddlePaddle/Paddle/pull/10943
  • Check label range in cross entropy calculation.
    • https://github.com/PaddlePaddle/Paddle/pull/10954

Qingsheng Li

• No Attention Model
  • Training: Cannot converge on wmt14 or wmt16 dataset
    • Encoder is correct now
    • Decoder has some unnatural behaviors
  • Inference:
    • Added Accumulated Prob
    • Issue caused by beam search op, fixed by Sheng Guo
• Attention Model
  • Training: Cannot overfit small dataset (1~4 lines of data)

Yan Xu

• performance
  • merge overlap rpc send and backward ops, performance improve 20%, https://github.com/PaddlePaddle/Paddle/pull/10550
• stablity
  • discuess dist train KIPs on CE, https://github.com/PaddlePaddle/Paddle/issues/10944
  • bug fix, https://github.com/PaddlePaddle/Paddle/pull/10937
  • bug fix, https://github.com/PaddlePaddle/Paddle/pull/10932
• feature
  • support RecordIO in distributed training, https://github.com/PaddlePaddle/Paddle/pull/11066
• PR review
  • https://github.com/PaddlePaddle/Paddle/pull/11039#pullrequestreview-124664458
  • https://github.com/PaddlePaddle/Paddle/pull/10881#pullrequestreview-123673797

wuyi

• finish resnet50 performance comparations with TF
• transpiler cleanup, put weight decay and clipping on pserver:
  • https://github.com/PaddlePaddle/Paddle/pull/11039
• Discussion with @weibao about rpc interface, and a prototype:
  • https://gist.github.com/typhoonzero/fc2489231c8e3a29d9ac7c895c8f4aae
• Look into TF codes to see how it deals with performance issue
• Testing LARS on imgnet

zhaochengduo

• Analysis of the reasons for the low acceleration ratio of PE 8 cards on Resnext152 with @yuyang
• PR
  • https://github.com/PaddlePaddle/Paddle/pull/11062
  • https://github.com/PaddlePaddle/Paddle/pull/11046
  • https://github.com/PaddlePaddle/Paddle/pull/11006
  • https://github.com/PaddlePaddle/Paddle/pull/10906
• Review
  • https://github.com/PaddlePaddle/Paddle/pull/11017

qiaolongfei

• distributed trianing

  • distribute lookup table support async training https://github.com/PaddlePaddle/Paddle/pull/10973
  • add auto_grown_mutex for selected_rows https://github.com/PaddlePaddle/Paddle/pull/11012
  • run dist lookup table on dlnel cluster[Done]

• AbacusToPaddle

  • build paddle CPP lib on centos4.3/6.3 kernel 2.x/3.x http://agroup.baidu.com/abacus2paddle/md/article/933770
  • make the plan http://agroup.baidu.com/abacus2paddle/md/article/936930
  • build Abacus with paddle static lib.
    • fix protobuf problem(Abacus: protobuf 2.4.1 Paddle: protobuf 3.1.0)
    • fix eigen build problem.(both use eigen)

• fengchao reinforcement learning with Paddle

  • operators alignment
  • make plan

• code clean

  • fix inference build on mac https://github.com/PaddlePaddle/Paddle/pull/11007

weixing

• Hsigmoid

  • https://github.com/PaddlePaddle/Paddle/pull/11063

• Polish and add fluid_guide to fluid/get_started:

  • https://github.com/PaddlePaddle/Paddle/pull/10961
  • https://github.com/PaddlePaddle/Paddle/pull/10979

• [WIP]paddlepaddle jupyter notebook raises error

Xin Pan

• Inference API implementation
• https://github.com/PaddlePaddle/Paddle/pull/11036
• https://github.com/PaddlePaddle/Paddle/pull/11009
• https://github.com/PaddlePaddle/Paddle/pull/10983
• profiler API and dist benchmark fix
• https://github.com/PaddlePaddle/Paddle/pull/11036

guochaorong

• CE frameworks

  • issue
    • https://github.com/PaddlePaddle/continuous_evaluation/issues/61
  • pr
    • https://github.com/PaddlePaddle/continuous_evaluation/pull/59
    • https://github.com/PaddlePaddle/continuous_evaluation/pull/60
    • https://github.com/PaddlePaddle/continuous_evaluation/pull/55
    • https://github.com/PaddlePaddle/continuous_evaluation/pull/58
    • https://github.com/PaddlePaddle/continuous_evaluation/pull/54
  • wiki
    • https://github.com/PaddlePaddle/continuous_evaluation/wiki/CE-%E5%80%BC%E7%8F%AD%E6%89%8B%E5%86%8C

• benchmark

  • environment built for Random model testing (Monitoring illegal access and convergence of model)

qiuxuezhong

• NMT
  • running transformer with mutil-trainers and mutil-psservers on one machine
  • try to running transformer on paddlecloud
• AbacusToPaddle
  • familiar with look-up table code

kexinzhao

• Add create LoDTensor from list option and simplify recommender book example:
  • https://github.com/PaddlePaddle/Paddle/pull/10946
• Modify machine translation example using new LoDTensor API
  • https://github.com/PaddlePaddle/Paddle/pull/11018
• Modify RNN encoder decoder example using new LoDTensor API
  • https://github.com/PaddlePaddle/Paddle/pull/11021
• Fix float16 demo location issue
  • https://github.com/PaddlePaddle/Paddle/pull/10948
• Review
  • https://github.com/PaddlePaddle/Paddle/pull/10721#pullrequestreview-123478779

sidgoyal78

• PR:

  • Fix order of attributes in label_semantic_roles example: https://github.com/PaddlePaddle/Paddle/pull/10993
  • consistent attribute naming in new API: https://github.com/PaddlePaddle/Paddle/pull/10947
  • train() function API in fit_a_line example: https://github.com/PaddlePaddle/Paddle/pull/11020

• Review:

  • new LoDTensor API: #10886, #10888, #10889, #10946, #11018, #11021
  • Translation with new API: https://github.com/PaddlePaddle/Paddle/pull/10895
  • Re-write book chapter with new API: https://github.com/PaddlePaddle/book/pull/524

daming-lu

• Finished re-writing book tests to use new Fluid API
  • https://github.com/PaddlePaddle/Paddle/pull/10894
  • https://github.com/PaddlePaddle/Paddle/pull/10892
  • https://github.com/PaddlePaddle/Paddle/pull/10890
• Re-writing PaddlePaddle book, demo code and Jupyter notebook based on new Fluid API
  • https://github.com/PaddlePaddle/book/pull/524
  • https://github.com/PaddlePaddle/book/pull/525
• Reviewed PRs and created issues:
  • https://github.com/PaddlePaddle/Paddle/pull/11020
  • https://github.com/PaddlePaddle/Paddle/pull/10895
  • https://github.com/PaddlePaddle/Paddle/pull/10993
  • https://github.com/PaddlePaddle/Paddle/pull/10947
  • https://github.com/PaddlePaddle/Paddle/issues/11057

jetfuel(Jeff)

Working with Daming, Nicky and Sid on the book documentation. Currently all changes will go to https://github.com/PaddlePaddle/book/tree/high-level-api-branch first to ensure stability.

[WIP] working on Recognize Digit example.

• PR:

  • Update sentiment analysis text: https://github.com/PaddlePaddle/PaddlePaddle.org/pull/482

• Issues:

  • Unable to import paddle fluid in latest docker: https://github.com/PaddlePaddle/Paddle/issues/11059

PR reviews:

• https://github.com/PaddlePaddle/Paddle/pull/11056#event-1654718201
• https://github.com/PaddlePaddle/book/pull/524#pullrequestreview-124636834

Nicky

• Rewrite Book chapter 5 Recommendation System documentation https://github.com/PaddlePaddle/book/pull/526
• Simply Machine translation function name https://github.com/PaddlePaddle/Paddle/pull/11056
• Discussed with Daming, Jeff, Siddarth for book documentation plan
• Follow up with Longfei for Machine Translation infer issue
• Review:
  • https://github.com/PaddlePaddle/book/pull/524
  • https://github.com/PaddlePaddle/book/pull/525

varunarora

• Initial version of menu editor, and several updates and simplification to new documentation building system https://github.com/PaddlePaddle/PaddlePaddle.org/pull/481