• Reviews mainly for Paddle Cloud:
  • https://github.com/PaddlePaddle/cloud/pull/321#pullrequestreview-56765613
  • https://github.com/PaddlePaddle/cloud/issues/285#issuecomment-321715555
  • https://github.com/PaddlePaddle/cloud/pull/291#discussion_r132756739
  • https://github.com/PaddlePaddle/Paddle/pull/3459#pullrequestreview-56217889
  • https://github.com/PaddlePaddle/recordio/pull/11/files#r133054348
• Fixes:
  • https://github.com/PaddlePaddle/Paddle/pull/3434
  • RecordIO reader: https://github.com/PaddlePaddle/Paddle/pull/3479
• Add more demo to Paddle Cloud:
  • https://github.com/PaddlePaddle/cloud/pull/300
  • https://github.com/PaddlePaddle/cloud/pull/307
• Improve article《基于容器的AI系统开发》，『程序员』magazine will post to the front page of http://www.csdn.net/ .
• Read Caffe2 cpp code: Tensor, Net, Workspace.

• Enable Code Coverage for Paddle refactor code https://github.com/PaddlePaddle/Paddle/pull/3489
• FIX: Release CPU/GPU memory at the end of the Program https://github.com/PaddlePaddle/Paddle/pull/3534
• Demangle exception call stack for PADDLE_ENFORCE https://github.com/PaddlePaddle/Paddle/pull/3463
• Fix macro !defined(PADDLE_ONLY_CPU) https://github.com/PaddlePaddle/Paddle/pull/3450

Review:

• Fix bug of enforce when dladdr not found #3527
• Implement EigenCudaStreamDevice #3497

• Implement the scatter function
• Use scatter function and gather function to implement gather_op operation
• Get familiar with Docker and everything.

Operator --> OpKernel --> Tensor/DeviceContext --> Eigen

• Implement EigenCudaStreamDevice
  • https://github.com/PaddlePaddle/Paddle/pull/3497
• [WIP]Implement device context manager
  • https://github.com/PaddlePaddle/Paddle/pull/3530
• reduce unittest gpu memory
  • https://github.com/PaddlePaddle/Paddle/pull/3448

reviews:

• https://github.com/PaddlePaddle/Paddle/pull/3461#issuecomment-322363848
• https://github.com/PaddlePaddle/Paddle/pull/3505#discussion_r133364219

• Bug Fix: always restart pod on PaddleCloud https://github.com/PaddlePaddle/cloud/pull/320
• Bug Fix: invalied paddle executalbe file with pip install,https://github.com/PaddlePaddle/Paddle/pull/3421
• Bug Fix: invalid path with hostPath mode,https://github.com/PaddlePaddle/cloud/pull/304
• Bug Fix: invalid trainer_count with GPU job,https://github.com/PaddlePaddle/cloud/pull/306
• Bug Fix: save large parameters failed,https://github.com/PaddlePaddle/Paddle/pull/3397
• Help with JPass team to deploy PaddleCloud on BaiduCloud
• Lean about sparse parameter updater(Go PServer supports sparse)
• Comand-line interface design doc:https://github.com/PaddlePaddle/cloud/pull/291

• Compiler related:
  • build documentation don't need install Paddle in docker before: https://github.com/PaddlePaddle/Paddle/pull/3488
  • add shared warpctc lib in whl: https://github.com/PaddlePaddle/Paddle/pull/3503
  • deep speech2 can directly use warpctc instead by export LD_LIBRARY_PATH: https://github.com/PaddlePaddle/models/
  • auto update the requirements in .travis.yml with python/setup.py.in: https://github.com/PaddlePaddle/Paddle/pull/3419pull/208
• Intel MKLDNN:
  • code review:
    • add shared mkl lib in whl: https://github.com/PaddlePaddle/Paddle/pull/3461
    • fix crash when disable WITH_SWIG_PY: https://github.com/PaddlePaddle/Paddle/pull/3495
  • discuss mkldnn parameters, with @hedaoyuan:
    • remove use_mkldnn_wgt flag: https://github.com/PaddlePaddle/Paddle/issues/3409
    • use param header to save mkldnn format info: https://github.com/PaddlePaddle/Paddle/pull/3516
• wechat PaddlePaddle: 389 fans -> 400 fans

• DS2
  • Refine submitting scripts for PaddleCloud:
    • https://github.com/PaddlePaddle/models/pull/199
    • https://github.com/PaddlePaddle/models/pull/202
  • Rebuild docker image for PaddleCloud
• Image
  • Read dilated conv code of caffe.
• Fixes:
  • pnpair evaluator: https://github.com/PaddlePaddle/Paddle/pull/3473
  • SGD trainer doc: https://github.com/PaddlePaddle/Paddle/pull/3526
• Reviews:
  • https://github.com/PaddlePaddle/models/pull/201#pullrequestreview-56069309
  • https://github.com/PaddlePaddle/models/pull/205#pullrequestreview-56308637
  • https://github.com/PaddlePaddle/cloud/pull/291#pullrequestreview-55432712

• Fix proto msg in refactoring code.
  • https://github.com/PaddlePaddle/Paddle/pull/3322
  • co-work with qingqing & wangyi
• Remove FC OP https://github.com/PaddlePaddle/Paddle/pull/3417
• Using constructor to create op https://github.com/PaddlePaddle/Paddle/pull/3444
• Survey on Caffe2's python side code
  • https://github.com/PaddlePaddle/Paddle/pull/3502
• [WIP] Removing shared_ptr of OP
  • [Merged] https://github.com/PaddlePaddle/Paddle/pull/3522
  • [Reviewing] https://github.com/PaddlePaddle/Paddle/pull/3538
  • [Reviewing] https://github.com/PaddlePaddle/Paddle/pull/3546
• [Reviewing] Implement Fill Op
  • https://github.com/PaddlePaddle/Paddle/pull/3505
• Tiny fixes
  • https://github.com/PaddlePaddle/Paddle/pull/3467
  • https://github.com/PaddlePaddle/Paddle/pull/3527

• delete useless codes in softmax backward. https://github.com/PaddlePaddle/Paddle/pull/3411
• enable self-defined index data in testLayerGrad: https://github.com/PaddlePaddle/Paddle/pull/3449
• globally normalized cross entropy over beam (not finished yet, the whole implementations will be finished in this week):
  • https://github.com/PaddlePaddle/Paddle/compare/develop...lcy-seso:cross_entropy_over_beam

• DS2
  • Train LM on baidu zhidao corpus
  • Prepare Unicom ASR dataset
  • Fix bug
    https://github.com/PaddlePaddle/models/pull/211
• SSD
  • Merge v2 demo
    • https://github.com/PaddlePaddle/models/pull/196
    • https://github.com/PaddlePaddle/models/pull/137
• Code Review
  • https://github.com/PaddlePaddle/models/pull/208#pullrequestreview-56812960

• PaddlePaddle Image Enhancement
  • Faster R-CNN: RCNNLossLayer and RCNNDetectionLayer with PR: https://github.com/PaddlePaddle/Paddle/pull/3496.
• Cooperation with Video Analysis Team in IDL:
  • Debug and train the Attention Net.
  • Differences were found between the models implemented by PaddlePaddle and Pytorch. And the layer to do shifting is needed and in developing.

DS2:

• PaddleCloud adaptation & experiments.
  • https://github.com/PaddlePaddle/models/pull/201 (merged)
  • https://github.com/PaddlePaddle/models/pull/205 (merged)
  • https://github.com/PaddlePaddle/models/pull/144#pullrequestreview-55731734
  • https://github.com/PaddlePaddle/models/pull/199#pullrequestreview-56057176
  • Experiments for performance verification.
• Unicom3k data preparation (with Yaming).

• PaddlePaddle Refactoring
  • make unit test of backward_test pass after refactoring framework proto:
    • https://github.com/wangkuiyi/Paddle/pull/12
  • Compare the gradient consistency between GPU and CPU calculations:
    • https://github.com/PaddlePaddle/Paddle/pull/3476
  • Implement GPU kernel for cross entropy operator:
    • https://github.com/PaddlePaddle/Paddle/pull/3501
  • code review:
    • Refactorize framework/*.proto: https://github.com/PaddlePaddle/Paddle/pull/3322#pullrequestreview-55991047
    • Caffe2 Survey for Python Topology: https://github.com/PaddlePaddle/Paddle/pull/3502#pullrequestreview-56518876
• Other: make PPT.

• Add two c-api interface
  • Create a gradient machine with merged model, https://github.com/PaddlePaddle/Paddle/pull/3400
  • Get output of a specified layer, https://github.com/PaddlePaddle/Paddle/pull/3528
• Help genome to use Paddle instead of caffe in there album app, and create a benchmark of Paddle and Caffe on Android, test the runtime.
• issues:
  • #3498
  • #3477
  • #3506

• Mobile
  • Work with genome and input method, try to deploy paddle into the mobile end.
  • Optimize
    • https://github.com/PaddlePaddle/Paddle/pull/3392
  • Review
    • https://github.com/PaddlePaddle/Paddle/pull/3400
    • https://github.com/PaddlePaddle/Mobile/pull/2
    • https://github.com/PaddlePaddle/Paddle/pull/3528
• Bug fix & Review
  • https://github.com/PaddlePaddle/Paddle/pull/3348

• Train DS2 model based on mfcc feature

  • https://github.com/PaddlePaddle/models/issues/209

• Learn pyCUDA and accelerate the mfcc computation

  • https://github.com/kuke/pycuda_mfcc_featurizer

• Medical train flow:
  • http://gitlab.baidu.com/gongweibao/medical
• recordio:
  • pip install on mac and linux
  • fix read binary data bugs:https://github.com/PaddlePaddle/recordio/pull/11
• PaddleCloud:
  • Upload or download large files timeout: https://github.com/PaddlePaddle/cloud/pull/321
• MPI cluster:
  • https://github.com/PaddlePaddle/cloud/issues/322

• Add 3D-Pooling
  • CPU/GPU implementation
  • Add 3D-Pooling unit test.
  • https://github.com/PaddlePaddle/Paddle/pull/3504
  • https://github.com/PaddlePaddle/Paddle/issues/438
• Conv3DLayer, DeConv3DLayer
  • https://github.com/PaddlePaddle/Paddle/pull/3348
• Adapting to the BatchNorm structure to support 3D data

• paddle refactor
  • LODTensor related
    • remove detail from LODTensor
    • add lod tensor interface design
    • tensor return const&
    • discussion, LODTensor related Variable type-deduce with inherience
    • discussion, Caffe2 sparse tensor and sequence, compared with LODTensor's design
    • add lod tensor interface design
  • framework related
    • add some doc to Backward
    • add version check of clang-format into pre-commit
  • rnn related
    • rnnop forward result test in python
    • rnnop make stepnet a member
    • add rnnop backward creation
    • open, rnnop backward unittest in python
  • reviews
    • https://github.com/PaddlePaddle/Paddle/pull/3538#discussion_r133593065
  • discussion
    • Caffe2 Survey for Python Topology
• others
  • Community construction
    • unify all the site's themes with @yongfeng
    • internal wiki
    • blog
    • models

• learning v1 sparse update
• large model prefetch update bug
  • https://github.com/PaddlePaddle/Paddle/pull/3518
  • https://github.com/PaddlePaddle/Paddle/pull/3445
• Some go math lib survey

• Got onboard.
• Trainings, got gears ready, learn and config Emacs.

• Finish Refactorizing of OpRegistry

  • https://github.com/PaddlePaddle/Paddle/pull/3373
  • https://github.com/PaddlePaddle/Paddle/pull/3436
  • https://github.com/PaddlePaddle/Paddle/pull/3481
  • https://github.com/PaddlePaddle/Paddle/pull/3512
  • https://github.com/PaddlePaddle/Paddle/pull/3542

• Read caffe2 python API design

• code review

  • https://github.com/PaddlePaddle/Paddle/pull/3428
  • https://github.com/PaddlePaddle/Paddle/pull/3430
  • https://github.com/PaddlePaddle/Paddle/pull/3433

Refactoring

• add auto gradient check design doc
  • https://github.com/PaddlePaddle/Paddle/pull/3410
• change operator public to protected
  • https://github.com/PaddlePaddle/Paddle/pull/3460
• caffe2 model design

Review:

• Feature/remove shared ptr
  • https://github.com/PaddlePaddle/Paddle/pull/3538
• Implement GPU kernel for cross entropy operator.
  • https://github.com/PaddlePaddle/Paddle/pull/3501
• Compare the gradient consistency between GPU and CPU calculations
  • https://github.com/PaddlePaddle/Paddle/pull/3476