2018 03 14

weixing

• Fix some outdated links in doc/fluid/read_source.md

  • https://github.com/PaddlePaddle/Paddle/pull/8886

• Fix some outdated contents in Contribute Documentations

  • https://github.com/PaddlePaddle/Paddle/pull/9016

• Fix bugs about building documentations failed.

  • https://github.com/PaddlePaddle/Paddle/pull/9093

qijun

Memory optimization

• [WIP] Recompute policy in memory optimization

  • https://github.com/PaddlePaddle/Paddle/pull/9089

• Runtime Operator and Variable should take OpDesc and VarDesc as private data members

  • https://github.com/PaddlePaddle/Paddle/issues/9069

Fix and enhance:

• Refine cast operator
  • https://github.com/PaddlePaddle/Paddle/pull/8923

Yu Yang

• [WIP] A prototype of ParallelExecutor in C++
  • Link: https://github.com/PaddlePaddle/Paddle/pull/9080/files#diff-564dec854cf4f37015001783f71e06cb
  • We have converted ProgramDesc into DependencyGraph with SSA form and analysis which operators can be paralleled.
• Add Readers and decorators
  • RecordIO Reader
  • Shuffle Decorator
  • DoubleBuffer Decorator

tangwei

• ISSUE:
  • https://github.com/PaddlePaddle/DeepSpeech/issues/170
• EDL
  • DeepSpeech2 on Sys Kubernetes Cluster
    • Distribute with CPU, done
    • Distribute with CPU and EDL, training
    • Distribute with GPU, debugging

Yan Xu

• document
  • enhancement dist train doc, https://github.com/PaddlePaddle/Paddle/pull/8893
  • sparse update doc, https://github.com/PaddlePaddle/Paddle/pull/8997
• bug fix
  • fix save model bug, https://github.com/PaddlePaddle/Paddle/pull/8928
  • fix dist compile failed, https://github.com/PaddlePaddle/Paddle/pull/8987
• design doc: distributed training with large parameter, https://github.com/PaddlePaddle/Paddle/pull/9068
• review:
  • PaddleCloud CRD, https://github.com/PaddlePaddle/cloud/pull/634
  • https://github.com/PaddlePaddle/Paddle/pull/9075/files#r174669158

Dang Qingqing

• Object Detection
  • Fix detection_map_op for multi-device.
    • https://github.com/PaddlePaddle/Paddle/pull/8845
  • Clipping bbox in the mAP evaluator calculation.
    • https://github.com/PaddlePaddle/Paddle/pull/8872
  • Fix bug in detection_output and mAP calculation in SSD.
    • https://github.com/PaddlePaddle/Paddle/pull/8985
  • Refine network config for MobileNet-SSD.
    • https://github.com/PaddlePaddle/models/pull/721
    • https://github.com/PaddlePaddle/models/pull/722
  • Train MobileNet-SSD on Pascal VOC dataset and do some contrastive experiments with Caffe.

yangyaming

• Adapt attention seq2seq
  https://github.com/PaddlePaddle/models/pull/675
• [WIP] Enhance sequence expand operator
  https://github.com/PaddlePaddle/Paddle/pull/9100
• Improve speed of async reader
  https://github.com/PaddlePaddle/models/pull/726

wanghaoshuang

• OCR
  • Add sum accumulator with window for model average[WIP]
    • https://github.com/PaddlePaddle/Paddle/pull/9082
• Review
  • https://github.com/PaddlePaddle/models/pull/721

Yang Yang (tonyyang-svail)

• [Speed] feature/ParallelExecutor: https://github.com/PaddlePaddle/Paddle/pull/8891
• Review:
  • Cpp parallel executor: https://github.com/PaddlePaddle/Paddle/pull/9080
  • Refine parallel_do_grad: https://github.com/PaddlePaddle/Paddle/pull/9072
  • Repair nccl op test: https://github.com/PaddlePaddle/Paddle/pull/8575

helinwang

• Participating Parallel Executor discussion and development with YangYang
• clean up EDL repo and fix CI
  • https://github.com/PaddlePaddle/cloud/pull/641
  • https://github.com/PaddlePaddle/edl/pull/5
  • https://github.com/PaddlePaddle/edl/pull/10
• review and issue discussions
  • https://github.com/PaddlePaddle/Paddle/pull/8891#pullrequestreview-103229732
  • https://github.com/PaddlePaddle/Paddle/issues/8903#issuecomment-371909647
  • https://github.com/PaddlePaddle/Paddle/issues/8929#issuecomment-371918111
  • https://github.com/PaddlePaddle/Paddle/pull/8997#pullrequestreview-103235570
  • https://github.com/PaddlePaddle/cloud/pull/634#pullrequestreview-103244068
  • https://github.com/PaddlePaddle/Paddle/pull/9033#pullrequestreview-103603905
  • https://github.com/PaddlePaddle/Paddle/issues/9034
  • https://github.com/PaddlePaddle/Paddle/pull/9068#issuecomment-373181797
  • https://github.com/PaddlePaddle/Paddle/pull/9079#pullrequestreview-104011829
  • https://github.com/PaddlePaddle/Paddle/pull/9080#pullrequestreview-104028959

ranqiu

• Fluid API Std (merged)

  https://github.com/PaddlePaddle/Paddle/pull/8927

• Fluid API doc problems

  https://github.com/PaddlePaddle/Paddle/issues/9056

  https://github.com/PaddlePaddle/Paddle/issues/9060

  https://github.com/PaddlePaddle/Paddle/issues/9062

  https://github.com/PaddlePaddle/Paddle/issues/9063

  https://github.com/PaddlePaddle/Paddle/issues/9064

  https://github.com/PaddlePaddle/Paddle/issues/9070

  https://github.com/PaddlePaddle/Paddle/issues/9092

  https://github.com/PaddlePaddle/Paddle/issues/9095

  https://github.com/PaddlePaddle/Paddle/issues/9097

• PR Review:

  https://github.com/dzhwinter/benchmark/pull/85

  https://github.com/dzhwinter/benchmark/pull/80

Chenxi

• CI

  • get 198 back online
  • added WITH_UBUNTU_MIRROR in CI configuration parameter https://github.com/PaddlePaddle/Paddle/issues/9073

• EDL Guide update

  • https://github.com/PaddlePaddle/edl/pull/9

• Document translation

  • https://github.com/PaddlePaddle/Paddle/issues/8911

abhinavarora

• Paddle CSP
  • Add modification to Channel to support Select OP https://github.com/PaddlePaddle/Paddle/pull/9084
  • Implement Select OP https://github.com/PaddlePaddle/Paddle/pull/9088#commits-pushed-b107575
  • Implement non blocking CanSend and CanReceive in Channels for Select OP https://github.com/PaddlePaddle/Paddle/issues/8815
  • Expose methods to create and add QueueMessage to Send or Receive Queue https://github.com/PaddlePaddle/Paddle/issues/8863
  • Add ability to provide callbacks in QueueMEssage Notify https://github.com/PaddlePaddle/Paddle/issues/8864
  • Add ability to provide condition variable externally to QueueMessage https://github.com/PaddlePaddle/Paddle/issues/9083
  • Fluid Channels should support both move and copy semantics for data transfer https://github.com/PaddlePaddle/Paddle/issues/9085
  • Move all Concurrency operators to paddle/fluid/operators/concurrency https://github.com/PaddlePaddle/Paddle/issues/9086
  • Channel Destroy should inform Select callback of destruction https://github.com/PaddlePaddle/Paddle/issues/9087
• Fix the CPP Data Feeding design document https://github.com/PaddlePaddle/Paddle/pull/9033
• Correct language inLarge parameter distributed training https://github.com/PaddlePaddle/Paddle/pull/9068/commits/5948fd27170d91e2728e9366e9852f32dc17fd2f
• PR Reviews
  • https://github.com/PaddlePaddle/ModelCE/pull/1#pullrequestreview-104119184
  • https://github.com/PaddlePaddle/Paddle/pull/9079#pullrequestreview-103954554
  • https://github.com/PaddlePaddle/Paddle/pull/9018#pullrequestreview-103976955
  • https://github.com/PaddlePaddle/Paddle/pull/8934#pullrequestreview-103640702

Liu Yiqun

• Inference Framework
  • [Reviewing] Remove unnecessary clone of program in C++ Executor.Run, a performance gain of 1.3% ~ 20%
    • https://github.com/PaddlePaddle/Paddle/pull/9043
  • Measure the performance gain of separate Executor.Prepare and Executor.RunPreparedContext, 1% ~ 9%
  • [Reviewing] Limit the symbol table of fluid shared library
    • https://github.com/PaddlePaddle/Paddle/pull/9065
• Mobile
  • Make a schedule with image group

Yibing Liu

• Train the DeepASR model on whole training dataset
  • https://github.com/PaddlePaddle/models/issues/737
• Add script for inference by using checkpoint
  • https://github.com/PaddlePaddle/models/pull/730
• Add pybind11 wrapper for decoder
  • https://github.com/PaddlePaddle/models/pull/736
• Fix the profiling bug in multi-gpu mode
  • https://github.com/PaddlePaddle/Paddle/pull/9077

Code Review:

• Refine data reader and move data_reader to async_data_reader
  • https://github.com/PaddlePaddle/models/pull/726
• Enhance reshape
  • https://github.com/PaddlePaddle/Paddle/pull/9008
• Add a test to ensure profiler works on multi-gpu
  • https://github.com/PaddlePaddle/Paddle/pull/9051
• Fix models #725
  • https://github.com/PaddlePaddle/Paddle/pull/9058

gongweibao

• Distributed GPU Version:
  • https://github.com/typhoonzero/grpc_zerocopy_async_example/pull/2
  • https://github.com/gongweibao/Paddle/tree/optsend
• Debugger: https://github.com/PaddlePaddle/Paddle/pull/9025
• Review:
  • https://github.com/PaddlePaddle/Paddle/pull/9027
  • https://github.com/PaddlePaddle/Paddle/pull/8839
  • https://github.com/PaddlePaddle/Paddle/pull/8900

dongzhihong

• refine nccl ops, reduce streamSychronize call
  • https://github.com/PaddlePaddle/Paddle/pull/9009
• export scatter op to python
  • https://github.com/PaddlePaddle/Paddle/pull/9038
• [Speed] sequence_softmax_op, sequence_softmax_op_grad and some others.
  • https://github.com/PaddlePaddle/Paddle/pull/8978
• try to fix memory bug/ parallel.do bug
  • https://github.com/PaddlePaddle/Paddle/issues/8621
  • https://github.com/PaddlePaddle/models/issues/732

wuyi

• fluid dist train:
  • https://github.com/PaddlePaddle/Paddle/pull/9027
  • update prototype bugs run in Linux: https://github.com/typhoonzero/grpc_zerocopy_async_example
• EDL:
  • make TPR version workable: https://github.com/PaddlePaddle/edl/pull/7
  • EDL PR reviews
• reviews:
  • https://github.com/PaddlePaddle/Paddle/pull/9075
  • https://github.com/PaddlePaddle/Paddle/pull/9068 TODO: async SGD fix simple transpiler that don’t split variables by default to boost dist train perf

fengjiayi

• C++ readers refining and potential bugs fix:
  • https://github.com/PaddlePaddle/Paddle/pull/8943
• C++ readers for multi-threads and multi-GPU
  • design doc: https://github.com/PaddlePaddle/Paddle/pull/9079
• readers and RecordIO review:
  • https://github.com/PaddlePaddle/Paddle/pull/8991
  • https://github.com/PaddlePaddle/Paddle/pull/8830

guosheng

• NMT:
  • Add beam search and inference program for Transformer.
    • https://github.com/PaddlePaddle/models/pull/727
  • Add initializers for Transformer and the corresponding cost curve.
    • https://github.com/PaddlePaddle/models/pull/701
    • https://github.com/PaddlePaddle/models/issues/700
• Review:
  • https://github.com/PaddlePaddle/models/pull/689
  • https://github.com/PaddlePaddle/models/pull/686
• Other:
  • Add guide for documentation-9-"RNN模型"
    • https://github.com/PaddlePaddle/Paddle/pull/8882

zhangchao

• Fix the Accuracy interface in benchmark.

  • https://github.com/dzhwinter/benchmark/pull/85

• Profiling for Transformer model.

  • https://github.com/PaddlePaddle/models/issues/697

• Add rnn_search in models.

  • https://github.com/PaddlePaddle/models/pull/729

• Reviews:

  • Remove Accuracy: https://github.com/PaddlePaddle/models/pull/704

Xin Pan

• Continue to improve profiler. Added nested block, nested event support, more tests coverage.

  • https://github.com/PaddlePaddle/Paddle/pull/9037
  • https://github.com/PaddlePaddle/Paddle/pull/9051

• Other improvements

  • https://github.com/PaddlePaddle/Paddle/pull/8901
  • https://github.com/PaddlePaddle/Paddle/pull/8897

• Work with Chunwei on Model CI and Contrib

• Work with Input Pipeline Design

• Work with Performance team and profiling

• Multi-threading prototype -https://github.com/panyx0718/Paddle/commit/0c581b1df1a7acec331129604a9fcfe6566951a7

zhaochengduo

• Optimization
  • [Speed]Refine parallel_do_grad
    • https://github.com/PaddlePaddle/Paddle/pull/9072
  • [Speed]fuse optimize op transpiler (with @qiaolongfei)
    • https://github.com/PaddlePaddle/Paddle/issues/8941
    • https://github.com/PaddlePaddle/Paddle/pull/8869
    • https://github.com/PaddlePaddle/Paddle/pull/8940
  • SE-ResNeXt-150 Optimization
    • https://github.com/PaddlePaddle/Paddle/issues/8990#issuecomment-372571527
• Enhancement regularization for sparse parameter
  • Enhance regularizer.py
    • https://github.com/PaddlePaddle/Paddle/pull/8934
  • Enhance look_up_table op
    • https://github.com/PaddlePaddle/Paddle/pull/8932
  • Add ElementwiseOpInferVarType for Elementwise_op
    • https://github.com/PaddlePaddle/Paddle/pull/8890
• Review
  • Better timeline
    • https://github.com/PaddlePaddle/Paddle/pull/9037
  • Repair nccl op test
    • https://github.com/PaddlePaddle/Paddle/pull/8575
  • Refine network config for MobileNet-SSD.
    • https://github.com/PaddlePaddle/models/pull/722
  • A little optimize of optimizer
    • https://github.com/PaddlePaddle/Paddle/pull/8874

luotao

• compiler: enable WITH_FLUID option: https://github.com/PaddlePaddle/Paddle/pull/9067
• inference:
  • add MKL for fluid static and shared library: https://github.com/PaddlePaddle/Paddle/pull/8887
  • [WIP] fuse batch norm
• doc:
  • fix document deployment: https://github.com/PaddlePaddle/Paddle/pull/8894
  • [Discuss] Collect and reclassify fluid documentation: https://github.com/PaddlePaddle/Paddle/issues/9031
• code review:
  • [doc guide]:
    • https://github.com/PaddlePaddle/Paddle/pull/8882
    • https://github.com/PaddlePaddle/Paddle/pull/8803
  • API standard: https://github.com/PaddlePaddle/Paddle/pull/8927
  • Contribute Documentations: https://github.com/PaddlePaddle/Paddle/pull/9016
  • Refine the profile codes for inference: https://github.com/PaddlePaddle/Paddle/pull/8910

qiaolongfei

fluid

• Profile

  • fuse optimize op transpiler https://github.com/PaddlePaddle/Paddle/pull/8940
  • [wip]add executor.prepare https://github.com/PaddlePaddle/Paddle/pull/9022
  • Discuss the method to support regularization on sparse parameter with @chenduo
  • Benchmark
    • https://github.com/PaddlePaddle/Paddle/issues/8818
    • https://github.com/PaddlePaddle/Paddle/issues/8941

• Abacus on Fluid

  • Add distributed lookup table design https://github.com/PaddlePaddle/Paddle/pull/9075

• Review

  • Extract Prepare from Executor. https://github.com/PaddlePaddle/Paddle/pull/9000
  • Enhance regularizer.py https://github.com/PaddlePaddle/Paddle/pull/8934
  • Enhance look_up_table op https://github.com/PaddlePaddle/Paddle/pull/8932
  • remove unnecessary build graph logic https://github.com/PaddlePaddle/Paddle/pull/8896

Yan Chunwei

• Model Continuous Evaluation

  • https://github.com/PaddlePaddle/ModelCE
  • discussion, issue
  • init design
    • https://github.com/PaddlePaddle/Paddle/pull/9018
  • WIP, initial implementation
    • works for one ResNet30 model, tracks the train cost and time consume.
    • truning the details

• Conv sequence to sequence model

  • train, reviewing

• reviews

  • https://github.com/PaddlePaddle/VisualDL/pull/304/files#r173348961
  • https://github.com/PaddlePaddle/Paddle/pull/9025#pullrequestreview-103373541

cs2be(thuan)

• PaddlePaddle.org
  • Fix Paddle Doc Generator (https://github.com/PaddlePaddle/PaddlePaddle.org/pull/439)
• Paddle
  • Implement Select OP (https://github.com/PaddlePaddle/Paddle/pull/9088)
    • Create Select OP unit tests
    • Create channel_utils, move shared helper methods for channel here
    • Update Channel class to support select op

wangkuiyi

Performance

• https://github.com/PaddlePaddle/Paddle/issues/8873#issuecomment-371686819

EDL

• https://github.com/PaddlePaddle/edl/issues/1

Distributed embedding operator design

• https://github.com/jacquesqiao/Paddle/pull/4

Continuous evluation

• https://github.com/PaddlePaddle/Paddle/pull/9002#pullrequestreview-103280369

Customier service

• https://github.com/PaddlePaddle/Paddle/issues/8999#issuecomment-372385999
• https://github.com/PaddlePaddle/Paddle/issues/8998#issuecomment-372388468
• https://github.com/PaddlePaddle/Paddle/issues/8949#issuecomment-372050985
• https://github.com/PaddlePaddle/Paddle/issues/8948#issuecomment-372389473
• https://github.com/PaddlePaddle/Paddle/issues/8942#issuecomment-372051198
• https://github.com/PaddlePaddle/Paddle/issues/8931#issuecomment-372395072
• https://github.com/PaddlePaddle/Paddle/issues/8892#issuecomment-372396478
• https://github.com/PaddlePaddle/Paddle/issues/8888#issuecomment-371678245
• https://github.com/PaddlePaddle/Paddle/issues/8885#issuecomment-371678718
• https://github.com/PaddlePaddle/Paddle/issues/8884#issuecomment-371679065
• https://github.com/PaddlePaddle/Paddle/issues/8883#issuecomment-371679505
• https://github.com/PaddlePaddle/Paddle/issues/8881#issuecomment-371680666
• https://github.com/PaddlePaddle/Paddle/issues/8881#issuecomment-373258541
• https://github.com/PaddlePaddle/Paddle/issues/8880#issuecomment-371674598
• https://github.com/PaddlePaddle/Paddle/issues/8878#issuecomment-371682403
• https://github.com/PaddlePaddle/Paddle/issues/8877#issuecomment-371685194
• https://github.com/PaddlePaddle/Paddle/issues/8876#issuecomment-371684707
• https://github.com/PaddlePaddle/Paddle/issues/8875#issuecomment-371685425
• https://github.com/PaddlePaddle/Paddle/issues/8860#issuecomment-371687425

jetfuel

PRs

• Deprecated san: https://github.com/PaddlePaddle/VisualDL/pull/302
• Only write the modified tablets to file system.: https://github.com/PaddlePaddle/VisualDL/pull/304
• Update the NODE_ENV variable to fix Vue in Production: https://github.com/PaddlePaddle/VisualDL/pull/310
• gcc can't properly parse the raw string literal with #define: https://github.com/PaddlePaddle/VisualDL/pull/311
• Add python 3 build support. : https://github.com/PaddlePaddle/VisualDL/pull/314
• Update Documentation for the release: Fixed the image not shown issue: https://github.com/PaddlePaddle/VisualDL/pull/317/files
• Update protobuf version to 3.4: https://github.com/PaddlePaddle/Paddle/pull/9091

Issue:

• import paddle.v2.fluid has TypeError: init() got an unexpected keyword argument ‘file'https://github.com/PaddlePaddle/Paddle/issues/9090

Support:

• https://github.com/PaddlePaddle/VisualDL/issues/320
• https://github.com/PaddlePaddle/VisualDL/issues/319
• https://github.com/PaddlePaddle/VisualDL/issues/315
• https://github.com/PaddlePaddle/VisualDL/issues/307
• https://github.com/PaddlePaddle/VisualDL/issues/305

nickyfantasy

• Issues & performance:
  • Fix smoothing value and tooltip issues https://github.com/PaddlePaddle/VisualDL/pull/323
  • Improve smoothing sliding performance by throttling https://github.com/PaddlePaddle/VisualDL/pull/312
• Design and UI Polish:
  • Polish scalar and histogram chart UI, reorganize chart tools https://github.com/PaddlePaddle/VisualDL/pull/303
  • Redesign chart toolbar icons and improve UX for download chart JSON data https://github.com/PaddlePaddle/VisualDL/pull/309
  • Improve chart color sequence https://github.com/PaddlePaddle/VisualDL/pull/318
  • Optimize expanded chart UI https://github.com/PaddlePaddle/VisualDL/pull/324

sidgoyal78

Inference:

• Bechmarking recognize_digits example using TensorRT:
  • Results: https://github.com/PaddlePaddle/Paddle/issues/8790
  • Code: https://github.com/sidgoyal78/shrp
• Discussion with people from Nvidia TensorRT regarding:
  • paddle to UFF converter (UFF is the representation format from TensorRT)
  • Issues with the Batch norm layer when using a Tensorflow model in Python with TensorRT
• Float16: Discussion with Kexin on modifying load-store ops for float16

kexinzhao

Inference:

• Add float16 support Mul Op: https://github.com/PaddlePaddle/Paddle/pull/9017
• Bind numpy float16 to paddle float16: https://github.com/PaddlePaddle/Paddle/pull/9017
• Add GPU compute compatibility check: https://github.com/PaddlePaddle/Paddle/pull/8946
• [WIP] add float16 support for cudnn conv op: https://github.com/PaddlePaddle/Paddle/pull/9098

varunarora

• Implementation and refinement of Python API, and Python tests for Select Op and cases

  • https://github.com/PaddlePaddle/Paddle/pull/9088

• Support miscellaneous issues and PRs in VDL: https://github.com/PaddlePaddle/VisualDL/issues/307, https://github.com/PaddlePaddle/VisualDL/pull/310, https://github.com/PaddlePaddle/VisualDL/pull/310, https://github.com/PaddlePaddle/VisualDL/issues/291 etc.