2018 06 14

2018-06-14

Yan Chunwei

• OPEN bugfix/add_inference_lib_to_release
  • https://github.com/PaddlePaddle/Paddle/pull/11455
• OPEN Feature/pass manager
  • https://github.com/PaddlePaddle/Paddle/pull/11440
• OPEN bugfix/trt op with kernel
  • https://github.com/PaddlePaddle/Paddle/pull/11408
• OPEN doc/inference api
  • https://github.com/PaddlePaddle/Paddle/pull/11332
• MERGED feature/anakin ci
  • https://github.com/PaddlePaddle/Paddle/pull/11330
• MERGED loose threshold of TRT for CI in different model
  • https://github.com/PaddlePaddle/Paddle/pull/11305

Tangwei

• [WIP] Checkpoint For lookup table
  • https://github.com/PaddlePaddle/Paddle/issues/11410
• documentation about checkpoint
  • https://github.com/PaddlePaddle/Paddle/issues/11266

Qingsheng Li

• Seq2seq Attention Model:
  • Convergence issue only exists on GPU.
  • Removed the GPU kernel of sequence_expand_op.
  • Debugging sequence_expand_op GPU kernel.

luotao

• mkldnn:
  • add FLAGS_use_mkldnn to global control use_mkldnn: https://github.com/PaddlePaddle/Paddle/pull/11319
  • align the mklml v2 benchmark between v0.11.0 and latest images on 6148 CPU. Test the mklml fluid benchmark using ParallelDo: http://agroup.baidu.com/paddlepaddle/md/article/964808
  • OCR (Doing)：fluid MKL version will affect the accuracy of Caffe MKL version on the same machine
• refine api docs:
  • refine docs of elementwise_op etc: https://github.com/PaddlePaddle/Paddle/pull/11369
  • add url of cuda9.0_cudnn7_avx_mkl：https://github.com/PaddlePaddle/Paddle/pull/11417
• code review:
  • feature/anakin ci: https://github.com/PaddlePaddle/Paddle/pull/11330
  • doc/inference api：https://github.com/PaddlePaddle/Paddle/pull/11332
  • bugfix/add_inference_lib_to_release：https://github.com/PaddlePaddle/Paddle/pull/11455
  • bugfix/trt op with kernel： https://github.com/PaddlePaddle/Paddle/pull/11408

Dang Qingqing

• Face Detection:
  • Train and debug model.
  • Make the normalization operator more general and fix bug in l2_normalize.
    • https://github.com/PaddlePaddle/Paddle/pull/11348
  • Refine the VGG-SSD network. https://github.com/PaddlePaddle/models/pull/977
  • Implement a bilinear initializer for transposed convolution to do upsampling.
    • https://github.com/PaddlePaddle/Paddle/pull/11404
  • Change the transposed conv2d initializer.
    • https://github.com/PaddlePaddle/models/pull/984
  • maxout Python API: https://github.com/PaddlePaddle/Paddle/pull/11278
• Others:
  • Debug nan in SSD model with @buxingyuan.
    • https://github.com/PaddlePaddle/Paddle/issues/11445
  • Some work for image-classification release: how to download ImageNet dataset and Realse the trained models.
• Code review:
  • SSD doc: https://github.com/PaddlePaddle/models/pull/968#pullrequestreview-127443167
  • Image classification: https://github.com/PaddlePaddle/models/pull/979#pullrequestreview-128214829
  • Achor generator: https://github.com/PaddlePaddle/Paddle/pull/11218#pullrequestreview-127454877
  • Argsort op: https://github.com/PaddlePaddle/Paddle/pull/11174#pullrequestreview-127448662

Yu Yang

• Documentation
  • https://github.com/PaddlePaddle/Paddle/pull/11308
  • https://github.com/PaddlePaddle/Paddle/pull/11287
  • https://github.com/PaddlePaddle/Paddle/pull/11284
• BugFix
  • https://github.com/PaddlePaddle/Paddle/pull/11346
• Single GPU Performance Profile

fengjiayi

• Preprocessing ops:
  • image_center_crop: https://github.com/PaddlePaddle/Paddle/pull/11245
  • crop_op updating: https://github.com/PaddlePaddle/Paddle/pull/11293
  • cast_op updating: https://github.com/PaddlePaddle/Paddle/pull/11442
• API reference:
  • https://github.com/PaddlePaddle/Paddle/pull/11457

Zeng Jinle

• [WIP] Profiling: single-GPU performance
• Pull Requests
  • Finish argmin and argmax operators (both CPU and GPU kernels)
    • https://github.com/PaddlePaddle/Paddle/pull/11313
  • Fix identifier errors in benchmark model vgg.py
    • https://github.com/PaddlePaddle/Paddle/pull/11350
  • Refine ZeroGradFunctor implementation in activation_op.h
    • https://github.com/PaddlePaddle/Paddle/pull/11380
• Issues
  • Identifier error bug in benchmark model vgg.py and stacked_dynamic_lstm.py
    • https://github.com/PaddlePaddle/Paddle/issues/11353
    • https://github.com/PaddlePaddle/Paddle/issues/11349
  • Issues on paddle installation and benchmark model running
    • https://github.com/PaddlePaddle/Paddle/issues/11394
    • https://github.com/PaddlePaddle/Paddle/issues/11360
• Code Review
  • Review argsort op and make suggestions on CPU kernel implementation
    • https://github.com/PaddlePaddle/Paddle/pull/11174

wanghaoshuang

• Metric learning

  • triplet loss op[WIP]
    • https://github.com/PaddlePaddle/Paddle/pull/11461
  • Enhance Print op to print tensors on specified CUDA devices.
    • https://github.com/PaddlePaddle/Paddle/pull/11397

• Others

  • Fix and merge mean iou op
    • https://github.com/PaddlePaddle/Paddle/pull/10519
  • Review and validate ICNet
    • https://github.com/PaddlePaddle/models/pull/975

• Review

  • https://github.com/PaddlePaddle/Paddle/pull/11348

tensor-tang

• PR

  • [Merged] Infer multi-threads API Demo and UT https://github.com/PaddlePaddle/Paddle/pull/11247
  • [Merged] add initial cpu memory flag in MB for infer https://github.com/PaddlePaddle/Paddle/pull/11392
  • [WIP] fix unknown use_mkldnn flag https://github.com/PaddlePaddle/Paddle/pull/11395
  • [WIP] scope thread safe https://github.com/PaddlePaddle/Paddle/pull/11258

• code review

  • [Merged] MKLDNN layout: Support for convolution operator https://github.com/PaddlePaddle/Paddle/pull/11099
  • [Merged] MKLDNN layout: Support for pool operator https://github.com/PaddlePaddle/Paddle/pull/11101
  • [Merged] MKLDNN layout: Support for batch norm operator https://github.com/PaddlePaddle/Paddle/pull/11098
  • [Merged] Add an interface to set the number of threads for math function, and set the default value to 1 for inference.https://github.com/PaddlePaddle/Paddle/pull/10789

• issue

  • [fixed] cpu memory issue https://github.com/PaddlePaddle/Paddle/issues/11272
  • [WIP] test fail with fluid only, unknown use_mkldnn https://github.com/PaddlePaddle/Paddle/issues/11393

• mklml

  • preproduce performance with v2 API，but fluid ParallelDo can not match the performance of v2 API.

qiaolongfei

• paddle fluid framework

  • fix build on mac
    • https://github.com/PaddlePaddle/Paddle/pull/11413

• paddle document

  • update split_lod_tensor, create_array and array_length doc
    • https://github.com/PaddlePaddle/Paddle/pull/11383

• distributed lookup table

  • outvar must be create in local scope for prefetch
    • https://github.com/PaddlePaddle/Paddle/pull/11450
  • Refine prefetch
    • https://github.com/PaddlePaddle/Paddle/pull/11366
  • Add merge_ids_op
    • https://github.com/PaddlePaddle/Paddle/pull/11342
  • fix distribute_transpiler
    • https://github.com/PaddlePaddle/Paddle/pull/11326

guosheng

• NMT:
  • Refine the beam_search PR and add unittests and docs.
    • https://github.com/PaddlePaddle/Paddle/pull/11238
  • Experiments on WMT16 en-de dataset (BPE data and weight sharing).
    • Compare with Tensor2Tensor and tune the model with new features (BPE data and weight sharing)
      • test data: newstest2013, BLEU paper reported: 25.8
      • multi_bleu: Fluid: 24.68; T2T: 23.5
      • t2t-bleu (uncased, cased): Fluid: (27.41, 26.93); T2T: (26.40, 25,89)
• Review:
  • https://github.com/PaddlePaddle/Paddle/pull/11391

Yibing Liu

• DeepASR:

  1. Acoustic model training;
  2. The enchancement & integration of decoder
  • https://github.com/PaddlePaddle/models/pull/967

• Transformer:

  1. Catch up with the training strategies in T2T;
  2. Model training of en-de translation model

• Dectection: Add GPU implementation and enhance CPU computation in Argsort Op for Faster RCNN https://github.com/PaddlePaddle/Paddle/pull/11174

baiyifan

• face detection:
  • refine infer to eval the model https://github.com/PaddlePaddle/models/pull/986
  • [WIP] check different shape input error

wuyi:

• reader support test: https://github.com/PaddlePaddle/Paddle/pull/11390
• Documentation:
  • docstring checker: https://github.com/PaddlePaddle/Paddle/pull/11351
  • https://github.com/PaddlePaddle/Paddle/pull/11314
• Fixes:
  • https://github.com/PaddlePaddle/Paddle/pull/11423
  • https://github.com/PaddlePaddle/Paddle/pull/11411
  • https://github.com/PaddlePaddle/Paddle/pull/11356

minqiyang

• Distribution:
  • Add robust cases support into fluid benchmark and add robust case into ce-latest-kpi
    • Issue
      • https://github.com/PaddlePaddle/Paddle/issues/11206
    • PR
      • https://github.com/PaddlePaddle/Paddle/pull/11207
  • Fix the bug that ParallelExecutor will hang up when GPU's memory is NOT enough
    • Issue
      • https://github.com/PaddlePaddle/Paddle/issues/11375
    • PR
      • https://github.com/PaddlePaddle/Paddle/pull/11377
  • Update the document of transpiler's parameter split strategy
    • Issue
      • https://github.com/PaddlePaddle/Paddle/issues/11250
    • PR
      • https://github.com/PaddlePaddle/Paddle/pull/11283
  • Fix the bug that Parameter Server does not add in the control block
    • Issue
      • https://github.com/PaddlePaddle/Paddle/issues/11429
• EDL:
  • Fix CRD's bugs in Dockerfile and edl_controller.yaml, which will cause the build failure and start failure
    • PR
      • https://github.com/PaddlePaddle/edl/pull/30

Xin Pan

• cuptiFinalize remove
  • https://github.com/PaddlePaddle/Paddle/pull/11454
• Fix activation op doc
  • https://github.com/PaddlePaddle/Paddle/pull/11295
• make status thread-safe
  • https://github.com/PaddlePaddle/Paddle/pull/11370
• Merge release 0.12.0 and 0.13.0 to master
  • https://github.com/PaddlePaddle/Paddle/pull/11358
• fix sparse var in dist train
  • https://github.com/PaddlePaddle/Paddle/pull/11248

dongzhihong

• memory optimize transpiler(add inlace)
  • https://github.com/PaddlePaddle/Paddle/pull/11386
  • https://github.com/PaddlePaddle/Paddle/pull/11385
• mixed precision training fp16 training
  • https://github.com/PaddlePaddle/Paddle/pull/11459
• eigen memory aligment
  • https://github.com/PaddlePaddle/Paddle/pull/11458
• eigen change default device to threadpool
  • https://github.com/PaddlePaddle/Paddle/pull/11407

zhaochengduo

• PR

  • Enable CPU on Parallel executor
    • https://github.com/PaddlePaddle/Paddle/pull/11306
  • Check SSA Graph
    • https://github.com/PaddlePaddle/Paddle/pull/11277
  • Fuse AllReduce Operator
    • https://github.com/PaddlePaddle/Paddle/pull/11141
  • Refine multi thread cpu parallel exe
    • https://github.com/PaddlePaddle/Paddle/pull/11406
  • SE-ResNeXt-152 multi card acceleration ratio tuning process
    • https://github.com/PaddlePaddle/Paddle/pull/11261
  • Add conv3d/pool3d/conv3d_trans Python API
    • https://github.com/PaddlePaddle/Paddle/pull/11437

• Review

  • overlap rpc op memcpy in distributed training
    • https://github.com/PaddlePaddle/Paddle/pull/11221
  • Fix NCCLBcast hang up bug in Parallel Executor
    • https://github.com/PaddlePaddle/Paddle/pull/11377
  • Add lock to record_event
    • https://github.com/PaddlePaddle/Paddle/pull/11346

weixing

• doc fix & fix generate doc error
  • https://github.com/PaddlePaddle/Paddle/pull/11421
  • https://github.com/PaddlePaddle/Paddle/pull/11419
  • https://github.com/PaddlePaddle/Paddle/pull/11438
  • https://github.com/PaddlePaddle/Paddle/pull/11427
  • https://github.com/PaddlePaddle/Paddle/pull/11424
• book v2 fix: machine_translation
  • https://github.com/PaddlePaddle/book/pull/538
• validate models to be merged
  • Distill：https://github.com/PaddlePaddle/models/pull/957
  • FCN：https://github.com/PaddlePaddle/models/pull/960

guochaorong

• CE duty shift mechanism
  • https://github.com/PaddlePaddle/continuous_evaluation/wiki/On-duty-manual
  • https://github.com/PaddlePaddle/continuous_evaluation/wiki/Duty-Records
• CE model reconstruction scheme
  • Celery distributed framework research:
    • https://github.com/PaddlePaddle/continuous_evaluation/issues/67
  • CE model reconstruction:
    • https://github.com/PaddlePaddle/continuous_evaluation/issues/66
    • https://github.com/PaddlePaddle/paddle-ce-latest-kpis/pull/37
• paddle bad code fix
  • https://github.com/PaddlePaddle/Paddle/pull/11317
  • https://github.com/PaddlePaddle/Paddle/pull/11374

qiuxuezhong

• NMT transformer distribute training run locally, but on paddlecloud core dump.
  • PR
    • fix bug: get var return null, https://github.com/PaddlePaddle/Paddle/pull/11431
    • distribute training script, https://github.com/PaddlePaddle/models/pull/982

daming-lu

• Re-write book chapters (Eng & Chi), demo programs (train.py) and Jupyter Notebook in Fluid High Level API syntax:

  • https://github.com/PaddlePaddle/book/pull/545
  • https://github.com/PaddlePaddle/book/pull/541
  • https://github.com/PaddlePaddle/book/pull/530
  • https://github.com/PaddlePaddle/book/commit/928d178c3a1b206eab54ab44047f15e70dd8cf0b
  • https://github.com/PaddlePaddle/book/commit/1b3bb17befaf08310d3f0cb1b1d5dc64a06acd84
  • https://github.com/PaddlePaddle/book/commit/48f26463b2ec46d0e59eafa8712a6b60589750f3
  • https://github.com/PaddlePaddle/book/commit/7fbabdd62394c217c05904f6b76a5c5c30838244

• The new PaddlePaddle official website

  • discussed with Beijing team and clarified the requirements

• Review:

  • https://github.com/PaddlePaddle/book/pull/542
  • https://github.com/PaddlePaddle/book/pull/539

nicky

• Add plot support in jupyter for book chapter 5 https://github.com/PaddlePaddle/book/pull/546
• Rewrite document and code in jupyter for book chapter 6 sentimental analysis https://github.com/PaddlePaddle/book/pull/539

Review:

• https://github.com/PaddlePaddle/book/pull/545
• https://github.com/PaddlePaddle/book/pull/544

jetfuel (Jeff Wang)

• PR

  • Re-write book chapter image classification: https://github.com/PaddlePaddle/book/pull/542
  • Fluid High-level-api can't infer with drop layer: https://github.com/PaddlePaddle/Paddle/pull/11323
  • Fix the issue where the paddlepaddlep.org docker tool not working: https://github.com/PaddlePaddle/PaddlePaddle.org/pull/485

• Review and Issues

  • https://github.com/PaddlePaddle/Paddle/issues/11299
  • https://github.com/PaddlePaddle/book/pull/541#pullrequestreview-126998051
  • https://github.com/PaddlePaddle/book/pull/544#pullrequestreview-127776258
  • https://github.com/PaddlePaddle/book/pull/539#pullrequestreview-128587195
  • https://github.com/PaddlePaddle/PaddlePaddle.org/issues/484#issuecomment-395910304
  • https://github.com/PaddlePaddle/PaddlePaddle.org/issues/484#issuecomment-395914398

varunarora

• Build offline real-time search for all documentation - in WIP PR: https://github.com/PaddlePaddle/PaddlePaddle.org/pull/481
• Work with Synopsys, Intel folks on ONNX issues

kexinzhao

• [Merged] continue to Modify Pybind LoDTensor API according to length-based LoD:
  • https://github.com/PaddlePaddle/Paddle/pull/11106
• [Merged] continue to Modify lod tensor doc based on new LoDTensor Python API:
  • https://github.com/PaddlePaddle/Paddle/pull/11253
• Review
  • https://github.com/PaddlePaddle/Paddle/pull/11453#pullrequestreview-128611743
  • https://github.com/PaddlePaddle/Paddle/pull/11451#pullrequestreview-128604980
  • https://github.com/PaddlePaddle/Paddle/pull/11323#pullrequestreview-127321922

Lei Wang

• Fix the incremental build issue
• https://github.com/PaddlePaddle/Paddle/pull/11378
• Read tensorflow design doc and papers