2018 05 08

wangkuiyi

• Code cleanup
  • Matrix multiplication: https://github.com/PaddlePaddle/Paddle/pull/10449#pullrequestreview-118172871
  • pre-commit hook: https://github.com/PaddlePaddle/Paddle/issues/10559
  • copyright hook: https://github.com/PaddlePaddle/Paddle/pull/10560
• float16 summarization: https://github.com/PaddlePaddle/Paddle/issues/10315
• build system cleanup: https://github.com/PaddlePaddle/Paddle/issues/10558

Yu Yang

• Plan to tuning ParallelExecutor and readers
  • tuning ParallelExecutor
    • Combine auto-batching and kernel fusion will speed up MultiGPUs computation
    • discuessed with Zhao Chengduo & Feng Jiayi. There is no complete design right now.
  • tuning readers
    • discuessed CustomReader with Feng Jiayi
• Python Trainer.train by Executor
  • https://github.com/PaddlePaddle/Paddle/pull/10343
• [WIP] Code Cleanup
  • Reduce 1000+ loc
  • https://github.com/PaddlePaddle/Paddle/pull/10371
  • https://github.com/PaddlePaddle/Paddle/pull/10486
  • https://github.com/PaddlePaddle/Paddle/pull/10458
  • https://github.com/PaddlePaddle/Paddle/pull/10449

fengjiayi

• Correct several misuse of TensorCopy:

  • https://github.com/PaddlePaddle/Paddle/pull/10412
  • https://github.com/PaddlePaddle/Paddle/pull/10548

• Implementation of trainer.test() of new Python API:

  • https://github.com/PaddlePaddle/Paddle/pull/10453

• Profile of Python data feeding:

  • https://docs.google.com/spreadsheets/d/1T0-BbPDwTj2epmDEQ8i9xJCqzGkMRESj-NgGdu1hyaw/edit#gid=0

• [WIP] data preprocessing of C++ data feeding (Implementation of CustomReader)

Lei Wang

• Travis CI: Enable ccache for docker build in Travis
  • https://github.com/PaddlePaddle/Paddle/pull/10395
• Doc: fix broken links in document
  • https://github.com/PaddlePaddle/Paddle/pull/10463
• Teamcity CI: fix uncatched test failures.
  • https://github.com/PaddlePaddle/Paddle/pull/10427
• Fix check style error:
  • https://github.com/PaddlePaddle/Paddle/pull/10422

qiaolongfei

• fluid high level api

  • Add inferencer infer https://github.com/PaddlePaddle/Paddle/pull/10445

• distributed lookup table

  • prepare envionment and run some test.

• Paddle support Abacus

  • 方案讨论，初步确定了一个大家能接受的迁移方案。
    • Abacus系统分成多个模块，逐步替换成Paddle。
      • 第一步替换模型配置和forward-backward的部分(Abacus对现有系统进行改造和模型调试，Paddle提供核心框架技术支持)。
      • 第二步替换pserver的部分(和ps-lib对接)。
      • 后续其他模块逐步切换。

• Bugfix & Code clean

  • fix fluid Metric https://github.com/PaddlePaddle/Paddle/pull/10439

Chenxi

• aws tool generating csv file to save metric data
  • https://github.com/PaddlePaddle/Paddle/pull/10542#event-1618830622
  • going to keep working on aws CE integration
• tried to fix manyliux build
  • going to stop supporting cuda7.5. https://github.com/PaddlePaddle/Paddle/issues/10469
  • https://github.com/PaddlePaddle/Paddle/issues/10466
  • https://github.com/PaddlePaddle/Paddle/issues/10419
  • https://github.com/PaddlePaddle/Paddle/issues/10418

Qingsheng Li

• Tested seq2seq training and inference with batch_size=1. The performance is poor when batch_size > 1. Still looking into it.
• Found an issue caused by empty tensor:
  • https://github.com/PaddlePaddle/Paddle/issues/10512
• [WIP] Adding an op to see if a tensor holds data
  • https://github.com/PaddlePaddle/Paddle/pull/10513

luotao

• inference:
  • implement the framework of convert tensorrt op, and convert between fluid lod_tensor and tensorrt itensor.
  • implement convert tensorrt relu op, and its unit-test.
    • [merge] https://github.com/PaddlePaddle/Paddle/pull/10144
    • [merge] https://github.com/PaddlePaddle/Paddle/pull/10461
    • https://github.com/PaddlePaddle/Paddle/pull/10495
  • fix tensorrt_engine compiler error：https://github.com/PaddlePaddle/Paddle/pull/10545
• code review:
  • feature/inference analysis dot：https://github.com/PaddlePaddle/Paddle/pull/10494
  • Feature/engine refactor：https://github.com/PaddlePaddle/Paddle/pull/10497
  • Bump MKL-DNN version from 0.11 to 0.14：https://github.com/PaddlePaddle/Paddle/pull/10501
  • fix deadline in machine_translation/README.cn.md：https://github.com/PaddlePaddle/book/pull/519

Dang Qingqing

• Fixed-point quantization
  • Survey TensorFlow and https://arxiv.org/abs/1712.05877
  • Desing doc: https://github.com/PaddlePaddle/Paddle/pull/10553
  • Task issue: https://github.com/PaddlePaddle/Paddle/issues/10551
• Support excluded and included mode for average pooling
  • https://github.com/PaddlePaddle/Paddle/pull/10235
• Work with Guoyi to support: 对长内用户的一键式安装whl包 + demo

gongweibao

• memory leak:
  • https://github.com/PaddlePaddle/Paddle/issues/10492
• performance of some framework:
  • https://github.com/PaddlePaddle/Paddle/pull/10120
  • https://github.com/gongweibao/tests/tree/develop/grpc_test

BaiYifan

• Develop GPU kernel for multiclass_nms_op
  • The acceleration effect is not ideal，need to be further optimized
• [WIP]Optimize multiclass_nms_op GPU kernel

tangwei

• [WIP] MPI Enabled
  • https://github.com/PaddlePaddle/Paddle/issues/9405
• [WIP] PServer checkpoint
  • https://github.com/PaddlePaddle/Paddle/pull/10532

kexinzhao

• float16 demo code and documentation
  • https://github.com/PaddlePaddle/Paddle/pull/10331
  • https://github.com/PaddlePaddle/Paddle/pull/10465
  • https://github.com/PaddlePaddle/Paddle/pull/10473
• Add use_cudnn option in FC function
  • https://github.com/PaddlePaddle/Paddle/pull/10387
• Fix bug in save op
  • https://github.com/PaddlePaddle/Paddle/pull/10392
• Remove unnecessary tensor copy in load op
  • https://github.com/PaddlePaddle/Paddle/pull/10402
• Fix sign unsigned comparison warning
  • https://github.com/PaddlePaddle/Paddle/pull/10424
• Add float16 support to conv3d op
  • https://github.com/PaddlePaddle/Paddle/pull/10425
• Add float16 support to pool3d op
  • https://github.com/PaddlePaddle/Paddle/pull/10433
• Add float16 support to load op
  • https://github.com/PaddlePaddle/Paddle/pull/10541
• Review:
  • https://github.com/PaddlePaddle/Paddle/pull/10508#pullrequestreview-118833140
  • https://github.com/PaddlePaddle/Paddle/pull/10471#pullrequestreview-118529788

wanghaoshuang

• Remove Chinese dataset of CRNN-CTC
  • https://github.com/PaddlePaddle/models/pull/898
• Add mean-IOU op.
  • https://github.com/PaddlePaddle/Paddle/pull/10519
• Help VIS with im2sequence supporting the batch inference.

Yan Xu

• code cleanup and bug fix
  • https://github.com/PaddlePaddle/Paddle/pull/10410
  • https://github.com/PaddlePaddle/Paddle/pull/10377
  • https://github.com/PaddlePaddle/Paddle/pull/10415
• unit test for distribute transpiler, https://github.com/PaddlePaddle/Paddle/pull/10485
• [WIP]overlap send op and backward ops, https://github.com/PaddlePaddle/Paddle/pull/10550

yangyaming

• Debug decoder for RNN Search (with QingSheng)
• [WIP] Add SequencePadOp
  https://github.com/PaddlePaddle/Paddle/pull/10533
• Code review for code cleaning
  https://github.com/PaddlePaddle/models/pull/894#pullrequestreview-118592003

wuyi

• NCCL2 dist train, testing perf with RDMA: https://github.com/PaddlePaddle/Paddle/pull/10349
• serde code polish: https://github.com/PaddlePaddle/Paddle/pull/10531
• testing training with LARS: https://github.com/PaddlePaddle/Paddle/pull/10374
• Add documents about training with RDMA hardware.

Yibing Liu

• Refine quotes' format and comments & add new operators
  • https://github.com/PaddlePaddle/paddle-onnx/pull/40
• Fix the fetch var bug when the arg is renamed
  • https://github.com/PaddlePaddle/paddle-onnx/pull/45
• Adapt the convertor to tensorrt backend
  • https://github.com/PaddlePaddle/paddle-onnx/pull/47
• [WIP] Add Inception_v4 model config in Fluid API
  • https://github.com/PaddlePaddle/models/pull/900

guosheng

• NMT:
  • Transformer code clean, configuration and data utility.
    • https://github.com/PaddlePaddle/models/pull/894
  • Beamsearch of Transformer(WIP)

Yan Chunwei

• Inference
  • feature/convert tensorrt io
  • refector inference engine
  • feature/analysis node representation
  • init Inference top APIs
  • discuss with Anakin team

https://github.com/PaddlePaddle/Paddle/pull/10144#pullrequestreview-117504134 https://github.com/PaddlePaddle/Paddle/pull/10414#pullrequestreview-117565473 https://github.com/Superjomn/paddle-ce-latest-kpis/pull/7#pullrequestreview-117902306 https://github.com/PaddlePaddle/Paddle/pull/10448#pullrequestreview-117916086 https://github.com/PaddlePaddle/Paddle/pull/10461#pullrequestreview-118194601 https://github.com/Superjomn/paddle-ce-latest-kpis/pull/8#pullrequestreview-118245105 https://github.com/PaddlePaddle/Paddle/pull/10495#pullrequestreview-118594253

weixing

PR:

Bug fix：

• Add classification_cost to v2 documentation

  • https://github.com/PaddlePaddle/Paddle/pull/10478

• Fix some image deadlinks in fluid documentation

  • https://github.com/PaddlePaddle/Paddle/pull/10483

• V2 deadlinks

  • https://github.com/PaddlePaddle/Paddle/pull/10518

issue:

• Two deadlinks in fluid documentation
  • https://github.com/PaddlePaddle/Paddle/issues/10484

[WIP]: Documents for Hsigmoid

dongzhihong

• precision alignment in image classification
  • refine batch norm op
  • https://github.com/PaddlePaddle/Paddle/pull/10502
  • se-renext float64 support, for debugging the precision alignment
  • https://github.com/PaddlePaddle/Paddle/pull/10407
  • https://github.com/PaddlePaddle/Paddle/pull/10346
  • https://github.com/PaddlePaddle/Paddle/pull/10450
  • [WIP] 累计误差会影响精度，op 高精度对齐op_test support, 对齐se-renext中op GPU

zhaochengduo

PR

• Fix CI(Merged)
  • https://github.com/PaddlePaddle/Paddle/pull/10480
• Feature/update sparse parameter(Merged)
  • https://github.com/PaddlePaddle/Paddle/pull/10351
• Updated different gradients on different cards
  • https://github.com/PaddlePaddle/Paddle/pull/10546
• Clean Parallel exe
  • https://github.com/PaddlePaddle/Paddle/pull/10476
• Fix fetch_op_handle(Merged)
  • https://github.com/PaddlePaddle/Paddle/pull/10454
• Wrap shfl_x_sync(Merged)
  • https://github.com/PaddlePaddle/Paddle/pull/10414

Review:

• Rewrite Matmul, make code cleaner
• Add float16 support to conv3d op
• Add float16 support to pool3d
• Correct tensor copy misuse

Liu Yiqun

• Inference Framework on Server
  • [Merged] Add the call of DropKids at the end of executor.Run to delete the local scopes created in operators
    • https://github.com/PaddlePaddle/Paddle/pull/10403
  • [Reviewing] Fix a bug related to dispensable inputs and refile the inference unittest
    • https://github.com/PaddlePaddle/Paddle/pull/10527
  • Profile the performance of OCR's recognization model
    • Remove the unused fetch_targets (ctc_align op is slow)
    • Update the Paddle Fluid library to newest version
    • Use the low level api, such as CreateVariables, Prepare and RunPreparedContext
    • The performance of recognization part is almost the same with v1 version, but the linking of Paddle Fluid library slows down the detection part.

guochaorong

• monitor for models repo
  • [merge] https://github.com/PaddlePaddle/continuous_evaluation/pull/36
  • [issue] https://github.com/PaddlePaddle/continuous_evaluation/issues/28

daming-lu

PaddlePaddle

• Simplified API: Made fit_a_line mixed version ( trainer.train() ) work. Working on word2vect
  • https://github.com/PaddlePaddle/Paddle/pull/10301
  • https://github.com/PaddlePaddle/Paddle/pull/10538
• Doc: Steps of building latest PaddlePaddle (with local change) from scratch
  • https://github.com/PaddlePaddle/Paddle/wiki/Build-PaddlePaddle-from-source-with-latest-local-changes

VisualDL

• [WIP] Deploy playground server.
• PRs reviewed:
  • https://github.com/PaddlePaddle/VisualDL/pull/445
  • https://github.com/PaddlePaddle/VisualDL/pull/449

jetfuel

Review and feedback for website navigation Handle Dnspod account transfer. Currently working with Parker to resolve the domain record issue.

• PRs

  • Convert all executable file to lowercase: https://github.com/PaddlePaddle/VisualDL/pull/449
  • Fix the toolbox text overflow issue: https://github.com/PaddlePaddle/VisualDL/pull/445
  • Provide zoom level control on Embedding: https://github.com/PaddlePaddle/VisualDL/pull/448
  • Add save and load params: https://github.com/PaddlePaddle/Paddle/pull/10386
  • Move all new API tests to high-level-api folder: https://github.com/PaddlePaddle/Paddle/pull/10534
  • [Test-driven] Label Semantic: start testing label_semantic: https://github.com/PaddlePaddle/Paddle/pull/10540
  • [Test-driven] Image Classification: https://github.com/PaddlePaddle/Paddle/pull/10537
  • [Test-driven] Word2Vect: https://github.com/PaddlePaddle/Paddle/pull/10538
  • [Test-deiven] Understand Sentiment: https://github.com/PaddlePaddle/Paddle/pull/10539
  • [Test-driven] Recognize Digit: https://github.com/PaddlePaddle/Paddle/pull/10507

• Feedbacks

  • https://github.com/PaddlePaddle/Paddle/pull/10535#pullrequestreview-118891565
  • https://github.com/PaddlePaddle/PaddlePaddle.org/issues/475#issuecomment-387567143
  • https://github.com/PaddlePaddle/PaddlePaddle.org/issues/475#issuecomment-387188858
  • https://github.com/PaddlePaddle/VisualDL/issues/447#issuecomment-387899713

nicky

• Qualcomm Android Demo
  • Completed SNPE Image Classification Demo with Alexnet trained by Caffe
  • Get AlexNet Paddle Model from PaddleCloud team
  • Analyze DLC converter error and follow up with Qualcomm team for unsupported ops
• VisualDL
  • Review:
    • https://github.com/PaddlePaddle/VisualDL/pull/449
    • https://github.com/PaddlePaddle/VisualDL/pull/445
    • https://github.com/PaddlePaddle/VisualDL/pull/448

sidgoyal78

• Code clean up:

  • PR https://github.com/PaddlePaddle/Paddle/pull/10394
  • review: https://github.com/PaddlePaddle/Paddle/pull/10390
  • review: https://github.com/PaddlePaddle/Paddle/pull/10392
  • review: https://github.com/PaddlePaddle/Paddle/pull/10386
  • Issue: non deterministic outputs: https://github.com/PaddlePaddle/Paddle/issues/10405
  • review: float64 fix: https://github.com/PaddlePaddle/Paddle/pull/10407

• ONNX:

  • PR https://github.com/PaddlePaddle/paddle-onnx/pull/43
  • Review: https://github.com/PaddlePaddle/paddle-onnx/pull/40

• New api:

  • PR, semantic roles: https://github.com/PaddlePaddle/Paddle/pull/10368
  • PR, recommendation system: https://github.com/PaddlePaddle/Paddle/pull/10535
  • Review, fit a line: https://github.com/PaddlePaddle/Paddle/pull/10301
  • Review: reorg of new scripts: https://github.com/PaddlePaddle/Paddle/pull/10534

• FP16:

  • PR for save_combine_op: https://github.com/PaddlePaddle/Paddle/pull/10471
  • PR LoD enhancement for FP16 test: https://github.com/PaddlePaddle/Paddle/pull/10508
  • Review: FP16 in load op: https://github.com/PaddlePaddle/Paddle/pull/10541

varunarora

• Work on new navigation for documentation (primarily tutorials) after needs assessments and studying other ML frameworks: https://github.com/PaddlePaddle/Paddle/issues/9860#issuecomment-387863602
• Working with Jeff to resolve DNS problems associated with paddlepaddle.org
• Working with Nicky/Qualcomm to prepare ONNX-DLC model for AI day demo
• [WIP] Helping Daming to deploy VDL playground
• Working on new brand assets for PaddlePaddle and VDL
• Fixing documentation (https://github.com/PaddlePaddle/paddle-onnx/pull/35) and feedback on TensorRT backend support update

tonyyang-svail

• [WIP] Survey: control flow in automatical differentiation.
• Speed up CI: https://github.com/PaddlePaddle/Paddle/pull/10506