2018 03 07

wangkuiyi

• Documentation

  • https://github.com/PaddlePaddle/Paddle/pull/8689#pullrequestreview-100549834

• Performance tuning

  • https://github.com/PaddlePaddle/Paddle/pull/8673#pullrequestreview-100614721
  • https://github.com/PaddlePaddle/Paddle/pull/8690#pullrequestreview-100617886
  • https://github.com/PaddlePaddle/Paddle/issues/8638#issuecomment-371375683
  • https://github.com/PaddlePaddle/Paddle/issues/8729#issuecomment-371374110

abhinavarora

• CSP
  • Adding more unit tests for ChannelHolder class https://github.com/PaddlePaddle/Paddle/pull/8668
  • Redesign channel implementation for Select Op https://github.com/PaddlePaddle/Paddle/pull/8814
  • [WIP] Implement non blocking CanSend and CanReceive in Channels for Select OP https://github.com/PaddlePaddle/Paddle/issues/8815
  • Issues Opened to Support Select
    • Fluid Channel should match Go Channel in Semantics https://github.com/PaddlePaddle/Paddle/issues/8813
    • 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
• Update Authors.md https://github.com/PaddlePaddle/Paddle/pull/8692
• PR Reviews
  • https://github.com/PaddlePaddle/Paddle/pull/8666#pullrequestreview-100283573
  • https://github.com/PaddlePaddle/Paddle/pull/8848#pullrequestreview-102045995
  • https://github.com/PaddlePaddle/Paddle/pull/8844#pullrequestreview-102043939
  • https://github.com/PaddlePaddle/Paddle/pull/8801#pullrequestreview-101624463
  • https://github.com/PaddlePaddle/Paddle/pull/8760#pullrequestreview-101309898
  • https://github.com/PaddlePaddle/Paddle/pull/8800#pullrequestreview-101625047

cs2be(thuan)

• Paddle
  • Fix Documentation and API links on paddlepaddle.org (https://github.com/PaddlePaddle/Paddle/pull/8760)
  • CSP
    • Select Operator, Select Operator Unit tests (https://github.com/PaddlePaddle/Paddle/pull/8867)

nickyfantasy

• VisualDL
  • Fix error when scalar data empty https://github.com/PaddlePaddle/VisualDL/pull/298
  • Fix issue where download data link points to empty run https://github.com/PaddlePaddle/VisualDL/pull/294
  • Reorganize css and fix font style and size https://github.com/PaddlePaddle/VisualDL/pull/287

gongweibao

• Improve distribution GPU fluid
  • ISSUE: https://github.com/PaddlePaddle/Paddle/issues/8638
  • Locate where the problem: https://github.com/PaddlePaddle/Paddle/pull/8762
  • Communication with BRPC: https://github.com/brpc/brpc/issues/255
  • Unit test of package request: https://github.com/typhoonzero/grpc_zerocopy_async_example/pull/1
• PaddleCloud
  • ISSUE: https://github.com/PaddlePaddle/cloud/issues/637
• Document:
  • https://github.com/PaddlePaddle/Paddle/pull/8848
• Review:
  • https://github.com/PaddlePaddle/cloud/pull/633
  • https://github.com/PaddlePaddle/cloud/pull/625
  • https://github.com/PaddlePaddle/Paddle/pull/8839

kexinzhao

• Inference Engine
  • To-do lists for float16 support: https://github.com/PaddlePaddle/Paddle/issues/8693
  • Add float16 GEMM GPU function in math_function: https://github.com/PaddlePaddle/Paddle/pull/8695
  • Add float16 support for Mul Op: https://github.com/PaddlePaddle/Paddle/pull/8817
  • Fix data_type_transform failure: https://github.com/PaddlePaddle/Paddle/pull/8850
• Review:
  • https://github.com/PaddlePaddle/Paddle/pull/8497#issuecomment-370011993

qijun

Memory optimization on Fluid:

Model	no optimize	reuse memory	release memory	forward memory
Resnet	170590208	92995584(reduce 45.5%)	78004224(reduce 54.3%)	77488128

• https://github.com/PaddlePaddle/Paddle/pull/8690
• Try to combine scheduler and memory optimization together

Fix and Enhance:

• https://github.com/PaddlePaddle/Paddle/pull/8755
• https://github.com/PaddlePaddle/Paddle/pull/8856

Docs:

• https://github.com/PaddlePaddle/Paddle/pull/8833

caoying

• PR review:
  • Inception-V4 for Fluid： https://github.com/PaddlePaddle/models/pull/672#pullrequestreview-100702951
  • squeeze/unsqueeze opeartor: https://github.com/PaddlePaddle/Paddle/pull/8541#pullrequestreview-100686151
• PR:
  • reorganized NMT directory: https://github.com/PaddlePaddle/models/pull/690
  • [WIP, not create a PR yet]: add multi-card/multi-thread implementation for Transformer
    • need to enhance reshape operator first: https://github.com/PaddlePaddle/Paddle/issues/8781
    • finish basic implementation, need to test and clean codes.

wanghaoshuang

• OCR CTC
  • Refine training net, final sequence error is 0.286(fluid) VS 0.278(V1) without 'model average'
    • https://github.com/PaddlePaddle/models/pull/596
  • Add model average option for fluid API [WIP]
• Add guide for howto documentation:
  • https://github.com/PaddlePaddle/Paddle/pull/8849

fengjiayi

• Remove Accuracy from Evaluator:
  • https://github.com/PaddlePaddle/Paddle/pull/8643
• A new design of model save/load:
  • save/load layer: https://github.com/PaddlePaddle/Paddle/pull/8711
  • related issue: https://github.com/PaddlePaddle/Paddle/issues/8521
• [WIP] Double buffer for C++ reader:
  • https://github.com/PaddlePaddle/Paddle/pull/8841
• Doc update:
  • https://github.com/PaddlePaddle/Paddle/pull/8765

luotao

• compile:
  • fix only shared variables could be declared as static in the device code on cuda 7.5: https://github.com/PaddlePaddle/Paddle/pull/8716
  • fix warning: statement is unreachable: https://github.com/PaddlePaddle/Paddle/pull/8740
  • refine operators/math/CMakeLists.txt: https://github.com/PaddlePaddle/Paddle/pull/8682
• inference:
  • compile and install the static library of fluid inference: https://github.com/PaddlePaddle/Paddle/pull/7827
• doc:
  • remove doc_theme: https://github.com/PaddlePaddle/Paddle/pull/8752
  • Create new structure for V2 and fluid doc: https://github.com/PaddlePaddle/Paddle/pull/8761
  • fix document deployment: https://github.com/PaddlePaddle/Paddle/pull/8772
• code review:
  • Fix building error of missing end-group for Android.: https://github.com/PaddlePaddle/Paddle/pull/8680
  • Fix fluid distribute build: https://github.com/PaddlePaddle/Paddle/pull/8704
  • [Intel] MKLDNN conv2d kernels added: https://github.com/PaddlePaddle/Paddle/pull/8451
  • 5 doc guides： https://github.com/PaddlePaddle/Paddle/issues?q=is%3Aissue+label%3Adocumentation+author%3Ashanyi15+is%3Aopen

wuyi

• distribute training performance boost:
  • https://github.com/PaddlePaddle/Paddle/pull/8839
  • https://github.com/typhoonzero/grpc_zerocopy_async_example
• EDL reviews:
  • https://github.com/PaddlePaddle/cloud/pull/634

Dang Qingqing

• SSD on Fluid:
  • Verify the backward of SSD loss.
    • https://github.com/PaddlePaddle/Paddle/issues/8626#issuecomment-369518496
  • Verify the correctness of detection_output API in SSD.
    • https://github.com/PaddlePaddle/Paddle/issues/8687
  • Implement detection mAP evaluator wrapper and unify label format between SSD loss and mAP evaluator.
    • https://github.com/PaddlePaddle/Paddle/pull/8736
  • Enable device automatically switching in mine_hard_examples_op.
    • https://github.com/PaddlePaddle/Paddle/pull/8706
  • Refine the doc in detection_output API.
    • https://github.com/PaddlePaddle/Paddle/pull/8689
  • Fix bug in detection mAP evaluator.
    • https://github.com/PaddlePaddle/Paddle/pull/8778
  • Refine MobileNet SSD model and add mAP evaluator.
    • https://github.com/PaddlePaddle/models/pull/694
  • Fix the test reader for SSD.
    • https://github.com/PaddlePaddle/models/pull/695
  • Converted a MobileNet-SSD model of Tensorflow based on COCO dataset for pre-training and train MobileNet-SSD.

Yan Xu

• [WIP] Fix sparse update bug, https://github.com/PaddlePaddle/Paddle/issues/8678
• document fix: https://github.com/PaddlePaddle/Paddle/pull/8853
• fix nccl version in manylinux Dockerfile, https://github.com/PaddlePaddle/Paddle/pull/8708

qiaolongfei

fluid

• Profile

  • se_resnet_152 multi-gpu profile https://github.com/PaddlePaddle/Paddle/issues/8661

  • [Speed]speed up python executor in fluid https://github.com/PaddlePaddle/Paddle/issues/8729

  • add se resnet 152 profile script https://github.com/dzhwinter/benchmark/pull/84

  • Add program cache for executor.py https://github.com/PaddlePaddle/Paddle/pull/8744

  • add image resnet profile https://github.com/dzhwinter/benchmark/pull/83

  • fluid vs pytorch profile https://github.com/PaddlePaddle/Paddle/issues/8677

    `time_per_batch` | memory
    ---- | ---
    1.95/1.15 = 1.696 | 18341 / 13359.0 = 1.373
    

  • DeepASR profile https://github.com/PaddlePaddle/Paddle/issues/8750

  • add timeline profile howto https://github.com/PaddlePaddle/Paddle/pull/8844

  • DataReader profile https://github.com/PaddlePaddle/Paddle/issues/8857

• bugfix and code optimize

  • fix snappy build https://github.com/PaddlePaddle/Paddle/pull/8804
  • add print_log to memory_optimize https://github.com/PaddlePaddle/Paddle/pull/8831

• Review

  • Add CPU time and MemCopy to the timeline. https://github.com/PaddlePaddle/Paddle/pull/8679
  • Enable device switching automatically for some operators in detection. https://github.com/PaddlePaddle/Paddle/pull/8684
  • Enable device automatically switching in mine_hard_examples_op. https://github.com/PaddlePaddle/Paddle/pull/8706
  • add inplace to reshape https://github.com/PaddlePaddle/Paddle/pull/8747
  • [Speed]Avoid init_nccl for every steps. https://github.com/PaddlePaddle/Paddle/pull/8758
  • Improve the timeline profiler https://github.com/PaddlePaddle/Paddle/pull/8775
  • [Speed]Refine elementwise_mul_op gradient functor https://github.com/PaddlePaddle/Paddle/pull/8810
  • [Speed] Refine elementwise sub,div,min,max gradient functor https://github.com/PaddlePaddle/Paddle/pull/8820
  • fix mac build error https://github.com/PaddlePaddle/Paddle/pull/8856

ranqiu

• Refine doc

  https://github.com/PaddlePaddle/Paddle/pull/8801

• Survey and Dissusion about API document standard

  https://github.com/PaddlePaddle/Paddle/issues/8834

  https://github.com/PaddlePaddle/Paddle/issues/8866

zhangchao

• Fix the error in Readme.md of mt_with_external_memory.
  • https://github.com/PaddlePaddle/models/pull/685
• Replace paddle.v2.fluid by paddle.fluid in benchmark.
  • https://github.com/dzhwinter/benchmark/pull/85
• [WIP]Operators profiling in Transformer model.
  • https://github.com/PaddlePaddle/models/issues/697
• [WIP] Implement rnn_search model with multi-gpu.
  • https://github.com/peterzhang2029/models/tree/add_rnn_search/fluid/neural_machine_translation/rnn_search

Yang Yang (tonyyang-svail)

• ParallelDo Profiling: https://github.com/PaddlePaddle/Paddle/issues/8719
• VGG net: https://github.com/PaddlePaddle/Paddle/issues/8718
• NCCL: https://github.com/PaddlePaddle/Paddle/pull/8575
• Parallel Executor: https://github.com/helinwang/Paddle/commit/07709351157f30260bff2da12c72a565b021bb70

tangwei

• Document: https://github.com/PaddlePaddle/Paddle/pull/8656
• ISSUE: https://github.com/PaddlePaddle/cloud/issues/638
• EDL Deploy EDL on Sys Kubernetes Cluster

guosheng

• NMT:
  • [WIP] Add inference and external beam search implementation for Transformer.
  • Tune Transformer model achieving comparative training cost with Pytorch.

yangyaming

• Refine document
  https://github.com/PaddlePaddle/Paddle/pull/8737
• Refine api design for beam search
  https://github.com/PaddlePaddle/models/pull/675
• Enhance data reader for DeepASR [WIP]

Liu Yiqun

• Inference Framework
  • [Merged] Add profiling information for inference example
    • https://github.com/PaddlePaddle/Paddle/pull/8748
  • [Reviewing] Add test for nested RecordEvent
    • https://github.com/PaddlePaddle/Paddle/pull/8773
  • Review
    • Fix nullptr when doing nested profiling: https://github.com/PaddlePaddle/Paddle/pull/8782
    • refine operators/math/CMakeLists.txt: https://github.com/PaddlePaddle/Paddle/pull/8682
    • Enable is_test attr of batch_norm and drop out op for test program: https://github.com/PaddlePaddle/Paddle/pull/8642
    • compile and install the static library of fluid inference: https://github.com/PaddlePaddle/Paddle/pull/7827
• Mobile
  • [Merged] Fix building error of missing end-group for Android
    • https://github.com/PaddlePaddle/Paddle/pull/8680
  • [Merged] Try to add build_android task back to travis
    • https://github.com/PaddlePaddle/Paddle/pull/8699
  • [Doing] Enable the building of Fluid for Android
    • https://github.com/PaddlePaddle/Paddle/pull/8709

Yibing Liu

• Add guideline for the doc of cmd parameters
  • https://github.com/PaddlePaddle/Paddle/pull/8732
• Setup DeepASR training on k8s cluster (thanks to @wuyi @weibao @yanxu), 2x speedup (4 GPUs, P40 vs. K40).
• Model training, find the bad convergence of momentum optimizer
  • https://github.com/PaddlePaddle/models/issues/696
  • https://github.com/PaddlePaddle/models/pull/692
• Some small fix
  • https://github.com/PaddlePaddle/models/pull/677
• Code Review:
  • https://github.com/PaddlePaddle/Paddle/pull/8848
  • https://github.com/PaddlePaddle/Paddle/pull/8773

Yu Yang

• [Reviewing] RecordIO Reader
  • https://github.com/PaddlePaddle/Paddle/pull/8830
• [Merged] RecordIO File Utilities
  • https://github.com/PaddlePaddle/Paddle/pull/8780
• Profiling GPU/CPU code
  • Several internal web page linkes.
• Several enhancements
  • https://github.com/PaddlePaddle/Paddle/pull/8657
  • https://github.com/PaddlePaddle/Paddle/pull/8791

zhaochengduo

• Performance optimization of single card single card
  • Refine elementwise_mul_op gradient functor
    • https://github.com/PaddlePaddle/Paddle/pull/8810
  • Refine concat_op
    • https://github.com/PaddlePaddle/Paddle/pull/8669
  • Refine elementwise sub,div,min,max gradient functor
    • https://github.com/PaddlePaddle/Paddle/pull/8820
• Enhancement
  • Add log before op Run
    • https://github.com/PaddlePaddle/Paddle/pull/8859
• Debug
  • For test_machine_translation.py, with regularization, the program will be hung out. https://github.com/PaddlePaddle/Paddle/issues/8697
• Review
  • https://github.com/PaddlePaddle/Paddle/pull/8845
  • https://github.com/PaddlePaddle/models/pull/694
  • https://github.com/PaddlePaddle/Paddle/pull/8689

dongzhihong

• Fluid
  • add develop standard
    • https://github.com/PaddlePaddle/Paddle/pull/8803
  • add feature/recordio io format
    • https://github.com/PaddlePaddle/Paddle/issues/8763
    • https://github.com/PaddlePaddle/Paddle/pull/8759
  • refine gpu common
    • https://github.com/PaddlePaddle/Paddle/pull/8742

Xin Pan

• paddle performance profiling
  • https://github.com/PaddlePaddle/Paddle/pull/8782
  • https://github.com/PaddlePaddle/Paddle/pull/8775
  • https://github.com/PaddlePaddle/Paddle/pull/8679
• paddle performance optimization
  • https://github.com/PaddlePaddle/Paddle/pull/8758
• paddle multi-threaded scheduling
  • https://github.com/PaddlePaddle/Paddle/issues/8818

sidgoyal78

Inference:

• TensorRT results: https://github.com/PaddlePaddle/Paddle/issues/8790 (Code has been shared with Kexin and Xreki in another repo)
• Discussion with Nvidia folks for TensorRT integration

PR review:

• Profiling info for inference example: https://github.com/PaddlePaddle/Paddle/pull/8748
• float16 add context wait: https://github.com/PaddlePaddle/Paddle/pull/8850
• float16 data type tranform: https://github.com/PaddlePaddle/Paddle/pull/8619

helinwang

• Performance tuning:
  • https://github.com/PaddlePaddle/Paddle/pull/8720
  • https://github.com/PaddlePaddle/Paddle/pull/7814#pullrequestreview-101288703
  • https://github.com/PaddlePaddle/Paddle/pull/7814#discussion_r172305161
  • https://github.com/PaddlePaddle/Paddle/issues/8724#issuecomment-370894576
  • https://github.com/PaddlePaddle/Paddle/pull/8674#pullrequestreview-101689269
  • https://github.com/PaddlePaddle/Paddle/issues/8638#issuecomment-370907258
• Move EDL from PaddlePaddle/cloud to PaddlePaddle/edl with Yi: https://github.com/PaddlePaddle/edl/pull/5

jetfuel

• Integrate PyBind documentation into Sphinx documentation generation: https://github.com/PaddlePaddle/VisualDL/pull/292
• Update Travis-CI setting: https://github.com/PaddlePaddle/VisualDL/pull/295, https://github.com/PaddlePaddle/VisualDL/pull/297
• Fix incorrect documentation link: https://github.com/PaddlePaddle/VisualDL/pull/288
• Add English examples of how to use VisualDL with PyTorch, MXNet, Keras: https://github.com/PaddlePaddle/VisualDL/pull/299
• Publish the new examples to the websites: https://github.com/PaddlePaddle/VisualDL/pull/300

varunarora

• WIP of CSP Select op and Channels redesign: https://github.com/PaddlePaddle/Paddle/pull/8867
• Collaborate with Jeff on VDL issues