2018 06 06

Lei Wang

• Fix dependencies for test_paddle_inference_api_impl and build error when WITH_TESTING is OFF
  • https://github.com/PaddlePaddle/Paddle/pull/11064
• Fix CI build for paddlepaddle/PARL repo
  • https://github.com/PaddlePaddle/PARL/pull/14
• Fix documents:
  • https://github.com/PaddlePaddle/Paddle/pull/11170
  • https://github.com/PaddlePaddle/Paddle/pull/11169
  • https://github.com/PaddlePaddle/Paddle/pull/11130
• Point go package installation path into build directory
  • https://github.com/PaddlePaddle/Paddle/pull/11166
• Fix teamcity build to skip CI when only changing documents files

Bai Yifan

• face detection
  • Add object_coverage constrain and fix label bug
    • https://github.com/PaddlePaddle/models/pull/962
  • [WIP] Verification of face detection accuracy
• code review
  • Add infer scripts: https://github.com/PaddlePaddle/models/pull/966
  • Add head bbox for Pyramid-Box: https://github.com/PaddlePaddle/models/pull/963

luotao

• inference engine:
  • add build and install document of fluid inference library: https://github.com/PaddlePaddle/Paddle/pull/11090
  • rewrite unittest of trt_activation_op: https://github.com/PaddlePaddle/Paddle/pull/11222
• mkldnn:
  • add ParallelDo CPU multi-thread training example for benchmark/fluid, fix test and flower dataset error and refine the codes:
    • https://github.com/PaddlePaddle/Paddle/pull/11158
    • https://github.com/PaddlePaddle/Paddle/pull/11184
    • https://github.com/PaddlePaddle/Paddle/pull/11199
    • https://github.com/PaddlePaddle/Paddle/pull/11204
  • OCR CPU Inference:
    • Newest MKLML library hang with the old static MKL library: Intel @[email protected] reproduce the issue, and report to MKL interenal team.
    • provide cuda8+cudnn5+MKL_static libpaddle_fluid.so
• code review:
  • MKLDNN:
    • Mkldnn layout：https://github.com/PaddlePaddle/Paddle/pull/11040
    • rename Mkldnn to MKLDNN: https://github.com/PaddlePaddle/Paddle/pull/11147

tensor-tang

• PR

  • [Merged] text_classification infer performance test https://github.com/PaddlePaddle/Paddle/pull/11080
  • [Merged] mkldnn name https://github.com/PaddlePaddle/Paddle/pull/11147
  • [Merged] enable infer api with multi-threads https://github.com/PaddlePaddle/Paddle/pull/11162
  • [Merged] fix abort multi-threads https://github.com/PaddlePaddle/Paddle/pull/11233
  • [WIP] Infer multi-threads API Demo and UT https://github.com/PaddlePaddle/Paddle/pull/11247

• review code

  • mkldnn layout https://github.com/PaddlePaddle/Paddle/pull/11040
  • scope clean up https://github.com/PaddlePaddle/Paddle/pull/11243
  • add python-opencv in latest https://github.com/PaddlePaddle/Paddle/pull/11242

• issue

  • [fixed] large memory when infer issue https://github.com/PaddlePaddle/Paddle/issues/11185
  • [WIP] abort in multi-threads inference on CPU https://github.com/PaddlePaddle/Paddle/issues/11231

• nlp performance report against online text_classification http://agroup.baidu.com/paddlepaddle/md/article/917068. We can get about 1.35X boost.

• mkldnn feedback http://agroup.baidu.com/paddlepaddle/md/edit/942676.

  • we are still trying to find a way use mkl_sequential
  • They do not have any plan of Sparse Matrix yet.
  • RNN is fairly done with gemm.

Tangwei

• Fix and Optimized Checkpoint

  • https://github.com/PaddlePaddle/Paddle/pull/10878

• Checkpoint On PaddleCloud

• the init parameter "optimizer" of Trainer() should be a function

  • https://github.com/PaddlePaddle/Paddle/issues/11157

• Definition of next steps

  • https://github.com/seiriosPlus/Paddle/wiki/Definition-of-next-steps

• Code Review:

  • https://github.com/PaddlePaddle/Paddle/pull/11155

guosheng

• NMT:
  • Fix and enhance beam_search_op and beam_search_decode_op.
    • https://github.com/PaddlePaddle/Paddle/pull/11238
  • Experiments on WMT14 en-de dataset.
    • Compare with Tensor2Tensor and tune the model with new features (BPE data and weight sharing)

qiaolongfei

• paddle fluid framework

  • fix protobuf memory leak https://github.com/PaddlePaddle/Paddle/pull/11177
  • add host_memory_profiling_cn.md https://github.com/PaddlePaddle/Paddle/pull/11191 https://github.com/PaddlePaddle/Paddle/pull/11208 https://github.com/PaddlePaddle/Paddle/pull/11212
  • "change eigen mirror" https://github.com/PaddlePaddle/Paddle/pull/11240
  • fix build error on mac https://github.com/PaddlePaddle/Paddle/pull/11134
  • fix transpiler package https://github.com/PaddlePaddle/Paddle/pull/11087
  • Fix compile error on mac caused by std move https://github.com/PaddlePaddle/Paddle/pull/11034

• distributed trianing

  • get data from wangsijiang@feed, start to implement deep & wide model.

• AbacusToPaddle

  • fix all compile problem, now two system can work together.

• fengchao reinforcement learning with Paddle

  • Discuss how to use VDL to debug their rl model
  • discuss the usage of reshape op

Chenxi

• aws integration with CE merged
• NCCL2 support in progress
• team city restore

fengjiayi

• Refine RecordIO reader with Yuyang
  • https://github.com/PaddlePaddle/Paddle/pull/11151
• Data preprocessing related job:
  • image resize API: https://github.com/PaddlePaddle/Paddle/pull/11198
  • reverse op: https://github.com/PaddlePaddle/Paddle/pull/11223
  • [WIP] center crop: https://github.com/PaddlePaddle/Paddle/pull/11245
• bug fix:
  • https://github.com/PaddlePaddle/Paddle/pull/11083

Yan Xu

• performance
  • overlap memcpy in rpc op, experiment on vgg model imrprove 15% performance, https://github.com/PaddlePaddle/Paddle/pull/11221
• stablity
  • dist mnist unit test, https://github.com/PaddlePaddle/Paddle/pull/11189
  • sppedup test_listen_and_serv, https://github.com/PaddlePaddle/Paddle/pull/11126
• PR review
  • checkpoint API, https://github.com/PaddlePaddle/Paddle/pull/10878#pullrequestreview-125431510
  • support record in becnhmark, https://github.com/PaddlePaddle/Paddle/pull/11121#pullrequestreview-125452035

Dang Qingqing

• Face Detection:
  • Add head bbox for Pyramid-Box model. https://github.com/PaddlePaddle/models/pull/963
  • Add infer scripts. https://github.com/PaddlePaddle/models/pull/966
  • Debug the training.
• Others:
  • API: https://github.com/PaddlePaddle/Paddle/issues/11246
  • code review:
    • box_coder_op:
      • https://github.com/PaddlePaddle/Paddle/pull/11164
      • https://github.com/PaddlePaddle/Paddle/pull/11129
    • slice_op: https://github.com/PaddlePaddle/Paddle/pull/11052#pullrequestreview-125821114
    • Prune dims supported by reduce op: https://github.com/PaddlePaddle/Paddle/pull/11113

gongweibao

• Clean up codes:
  • Add rpc_client interface: https://github.com/PaddlePaddle/Paddle/pull/11154
  • Add comment of grpc.tar.xz: https://github.com/PaddlePaddle/Paddle/pull/11153
  • Fix benmark/readme bug: https://github.com/PaddlePaddle/Paddle/pull/10960
  • Move sync_mode device ctx from grpc server: https://github.com/PaddlePaddle/Paddle/pull/10881
• Add brpc surpport
  • https://github.com/PaddlePaddle/Paddle/pull/11263

kexinzhao

• Modify Pybind LoDTensor API according to length-based LoD:
  • https://github.com/PaddlePaddle/Paddle/pull/11106
• Modify lod tensor doc based on new LoDTensor Python API:
  • https://github.com/PaddlePaddle/Paddle/pull/11253
• Fix LodTensor API in memory opt machine translation example:
  • https://github.com/PaddlePaddle/Paddle/pull/11171
• Review
  • https://github.com/PaddlePaddle/Paddle/pull/11170#pullrequestreview-125777639
  • https://github.com/PaddlePaddle/Paddle/pull/11169#pullrequestreview-125759055
  • https://github.com/PaddlePaddle/Paddle/pull/11167#pullrequestreview-125717589
  • https://github.com/PaddlePaddle/Paddle/pull/11166#pullrequestreview-125749672
  • https://github.com/PaddlePaddle/Paddle/pull/11130#pullrequestreview-125335601

tonyyang-svail

• Tape Prototype: https://github.com/PaddlePaddle/Paddle/pull/11019

Yan Chunwei

• Inference

  • high level API

    • merged, simplify inference api
      • https://github.com/PaddlePaddle/Paddle/pull/11104
    • merged, inference API little fix
      • https://github.com/PaddlePaddle/Paddle/pull/11069
    • merged, feature/simple inference demo
      • https://github.com/PaddlePaddle/Paddle/pull/11105
    • merged, Feature/anakin embed
      • https://github.com/PaddlePaddle/Paddle/pull/11135

  • sub-graph related

    • open, feature/trt engine op test
      • https://github.com/PaddlePaddle/Paddle/pull/11182
    • merged, Feature/fc converter
      • https://github.com/PaddlePaddle/Paddle/pull/11043
    • merged, feature/tensorrt engine op
      • https://github.com/PaddlePaddle/Paddle/pull/11001
    • add merged, dfg graphviz pass
      • https://github.com/PaddlePaddle/Paddle/pull/11211

  • enhancement && bugfix

    • WIP, add replace enforce with glog switch for debug
      • https://github.com/PaddlePaddle/Paddle/pull/11244
    • merged, fix manylinux compile error caused by inference lib
      • https://github.com/PaddlePaddle/Paddle/pull/10994
    • merged, clean docstring_checker.pyc
      • https://github.com/PaddlePaddle/Paddle/pull/11093
    • merged, fix compile error
      • https://github.com/PaddlePaddle/Paddle/pull/11119

https://github.com/PaddlePaddle/Paddle/pull/11090#pullrequestreview-125044331 https://github.com/PaddlePaddle/continuous_evaluation/pull/59#pullrequestreview-125044827 https://github.com/PaddlePaddle/continuous_evaluation/pull/64#pullrequestreview-126628433

wuyi

• Train ImageNet On 64 GPUs with LARS, data prepare (preprocess op?)
• fluid_benchmark update: https://github.com/PaddlePaddle/Paddle/pull/11121
• Trainer send complete signal: https://github.com/PaddlePaddle/Paddle/pull/11220
• Refine RPC client sync wait: https://github.com/PaddlePaddle/Paddle/pull/11132
• small fixes and reviews
• TODO: New EDL design doc

zhaochengduo

• PR

  • [WIP]SE-ResNeXt-152 multi card acceleration ratio tuning process
    • https://github.com/PaddlePaddle/Paddle/pull/11261
  • Fuse AllReduce Operator
    • https://github.com/PaddlePaddle/Paddle/pull/11141
  • [Feature] Add fuse vars op handle
    • https://github.com/PaddlePaddle/Paddle/pull/11237
  • Refine fluid_benchmark.py
    • https://github.com/PaddlePaddle/Paddle/pull/11118
  • Balance parameter opt
    • https://github.com/PaddlePaddle/Paddle/pull/11079
  • Drop the last batch, if the size of last batch is not equal to batch_size.
    • https://github.com/PaddlePaddle/Paddle/pull/11062
  • Add resnet 50
    • https://github.com/dzhwinter/benchmark/pull/105

• Review

  • Fluid benchmark support recordio reader
    • https://github.com/PaddlePaddle/Paddle/pull/11121
  • SSA Graph Builder Factory
    • https://github.com/PaddlePaddle/Paddle/pull/11234
  • Add image_resize_short and refine resize API
    • https://github.com/PaddlePaddle/Paddle/pull/11198

Zeng Jinle

• Pull requests
  • https://github.com/PaddlePaddle/Paddle/pull/11038
  • https://github.com/PaddlePaddle/Paddle/pull/11176
• issues
  • https://github.com/PaddlePaddle/Paddle/issues/11037
  • https://github.com/PaddlePaddle/Paddle/issues/11175
• [WIP]Add argmin and argmax ops

Yibing Liu

• DeepASR: 1) Acoustic model training; 2)Adapt decoder to the new net config

  • https://github.com/PaddlePaddle/models/pull/967

• Transformer: model training （with@guosheng）

• Dectection: Add Argsort Op

  • https://github.com/PaddlePaddle/Paddle/pull/11174

• ONNX convertor: Merge compare ops & several relu ops

  • https://github.com/PaddlePaddle/paddle-onnx/pull/51

Code Review:

• https://github.com/PaddlePaddle/Paddle/pull/11052

Yu Yang

• Speed up read RecordIO
  • https://github.com/PaddlePaddle/Paddle/pull/11116
• Refactor ParallelExecutor to factory
  • https://github.com/PaddlePaddle/Paddle/pull/11234
• Fuse AllReduceOp
  • https://github.com/PaddlePaddle/Paddle/pull/11141

Xin Pan

• Release 0.13.0
  • https://github.com/PaddlePaddle/Paddle/releases/tag/v0.13.0
• Debug memory leak
  • https://github.com/panyx0718/Paddle/commit/ca8913fe484e02ba92014d62c0ddf05515c65b1d
  • https://github.com/panyx0718/Paddle/commit/b05e173db64b032ac2e08133bb5f9048f6ab1a00
  • https://github.com/panyx0718/Paddle/commit/b87b4905c5eff5db8b11491f341f25a52dc0e6a5
• Debug distributed train hang
• fluid_benchmark.py
  • https://github.com/PaddlePaddle/Paddle/pull/11215
• Scope clean up
  • https://github.com/PaddlePaddle/Paddle/pull/11243

guochaorong

• CE frame

  • support run modified models of CE
    • https://github.com/PaddlePaddle/continuous_evaluation/pull/64
  • CE web problem
    • https://github.com/PaddlePaddle/continuous_evaluation/issues/63
  • CE document （CE onduty and Hi alarm）
    • https://github.com/PaddlePaddle/continuous_evaluation/wiki

• Teamcity CI server & db fault tolerance

  • https://github.com/PaddlePaddle/Paddle/issues/11254

• CE resnet add multi card， speedup testing on P40（doing）

  • https://github.com/PaddlePaddle/Paddle/issues/11225

• paddle code scan（c plus and python）

  • https://github.com/PaddlePaddle/Paddle/issues/11256
  • https://github.com/PaddlePaddle/Paddle/issues/11257

dongzhihong

• memory optimize
  • static ssa graph convert and optimize
• model non-determinstic/reproducible
  • check op in http://agroup.baidu.com/paddlepaddle/view/office/946408
  • compatibale with cudnn5
    • https://github.com/PaddlePaddle/Paddle/pull/11224
  • fix cudnn non-determinstic
    • https://github.com/PaddlePaddle/Paddle/pull/11205
  • try to fix non-determinstic issue with @Goyal
    • https://github.com/PaddlePaddle/Paddle/pull/11133
    • https://github.com/PaddlePaddle/Paddle/pull/11229
• Paddle AMD device support
  • accelerate amd device support
    • https://github.com/PaddlePaddle/Paddle/pull/11202
  • solve the save algorithm issue in conv/conv_grad
    • https://github.com/PaddlePaddle/Paddle/issues/11203

wanghaoshuang

• Adapt ModelAverage to latest high level api.
  • https://github.com/PaddlePaddle/Paddle/pull/11249
• Prune dims supported by reduce op.
  • https://github.com/PaddlePaddle/Paddle/pull/11113
• Rewrite OCR CTC model by latest high level api.[WIP]

qiuxuezhong

• nmt

  • running transformer on paddle cloud with multi-machines，with some problems:
    • stability: cored when trainer num is large，for example 8，16
    • precision: avg loss is 3.5+, the best is 2.5 when trained wite multi-cards on one machine
    • old version：only one card multi-machines
  • move newest version of transformer to paddle cloud
    • one bug to be fixed：gpu release verson of paddle won't assert when embbeding index out of range

• abacus2paddle

  • paddle cloud can afford 100+ machine for ctr trainning test

sidgoyal78

• High level API:

  • PR: Modify optimizer in new API: https://github.com/PaddlePaddle/Paddle/pull/11168
  • PR: Fix optimizer: https://github.com/PaddlePaddle/Paddle/pull/11172
  • PR: Label semantic roles book example: https://github.com/PaddlePaddle/book/pull/540
  • PR: sentiment analysis book example: https://github.com/PaddlePaddle/book/pull/539
  • Review: Recommendation system book example: https://github.com/PaddlePaddle/Paddle/pull/11252
  • Review: Compare results for data results: https://github.com/PaddlePaddle/book/pull/536
  • Review: Update MNIST book with new optimizer: https://github.com/PaddlePaddle/book/pull/535
  • Review: recommender system example: https://github.com/PaddlePaddle/book/pull/526
  • Review: Image classification book example: https://github.com/PaddlePaddle/book/pull/533
  • Review: LoDTensor API change: https://github.com/PaddlePaddle/Paddle/pull/11171
  • Review: Recognize digits example: https://github.com/PaddlePaddle/book/pull/529

• Fix non-determinism in Paddle CUDA kernels:

  • PR: Sparse sgd without atomicAdd: https://github.com/PaddlePaddle/Paddle/pull/11229
  • PR: Non-determinism in Sentiment analysis implementation: https://github.com/PaddlePaddle/Paddle/pull/11133

• Others:

  • PR: Fix signed-unsigned: https://github.com/PaddlePaddle/Paddle/pull/11167

daming-lu

• Finished 2 chapters in book following the new Fluid API

  • https://github.com/PaddlePaddle/book/pull/530
  • https://github.com/PaddlePaddle/book/pull/525
  • https://github.com/PaddlePaddle/book/pull/524

• Found a few issues while working on the chapter re-writing

  • https://github.com/PaddlePaddle/book/issues/534
  • https://github.com/PaddlePaddle/Paddle/issues/11057
  • https://github.com/PaddlePaddle/Paddle/issues/11210

• Reviewed PRs:

  • https://github.com/PaddlePaddle/book/pull/528
  • https://github.com/PaddlePaddle/Paddle/pull/11058/
  • https://github.com/PaddlePaddle/Paddle/pull/11020

jetfuel (Jeff)

• PR:
  • Recognize digit example updated with high level api draft: https://github.com/PaddlePaddle/book/pull/528
  • Recognize digit example train script updated and draft 2: https://github.com/PaddlePaddle/book/pull/529
  • Recognize digit Chinese Markdown update: https://github.com/PaddlePaddle/book/pull/531
  • Recognize digit Update MNIST to use optimizer_func: https://github.com/PaddlePaddle/book/pull/535
  • Image Classification train.py: https://github.com/PaddlePaddle/book/pull/533
• Issues: https://github.com/PaddlePaddle/book/issues/527 https://github.com/PaddlePaddle/VisualDL/issues/459

Nicky

• PR:
  • Recommendation System Book chapter 5 with high level api code and documentation: https://github.com/PaddlePaddle/book/pull/526
  • Second draft of Recommendation System Book https://github.com/PaddlePaddle/Paddle/pull/11252
  • Update High level API test of Recommendation System https://github.com/PaddlePaddle/Paddle/pull/11252
• Reviews:
  • https://github.com/PaddlePaddle/book/pull/528
  • https://github.com/PaddlePaddle/book/pull/525
  • https://github.com/PaddlePaddle/book/pull/524

varunarora

• Updates to handle language and version switching, fully working menu editor on PaddlePaddle.org: https://github.com/PaddlePaddle/PaddlePaddle.org/pull/481
• Testing paddle-onnx on TensorRT issues for nGraph-ing