2017 06 14

wangkuiyi

• https://github.com/PaddlePaddle/Paddle/pull/2432

• https://github.com/PaddlePaddle/Paddle/wiki/TensorFlow-Kernels

• https://github.com/PaddlePaddle/Paddle/wiki/TensorFlow-Ops

• https://github.com/PaddlePaddle/Paddle/wiki/Caffe2::Tensor

• https://github.com/PaddlePaddle/Paddle/wiki/Caffe2::Operator

helinwang

• Cloud: The master server now can dispatch training RecordIO task efficiently to Python library running on local machine.
  • Implement master server client: https://github.com/PaddlePaddle/Paddle/pull/2429
  • Modify pserver client C API, create better test: https://github.com/PaddlePaddle/Paddle/pull/2433
  • Implement master client for reading training tasks: https://github.com/PaddlePaddle/Paddle/pull/2468
• Fix issues:
  • Fix issue that C buffer may be no longer valid when RecordIO dumps records: https://github.com/PaddlePaddle/recordio/pull/7
  • Fix build script, English link and dead link: https://github.com/PaddlePaddle/book/pull/331
• Code Review:
  • Optimizer:
    • https://github.com/PaddlePaddle/Paddle/pull/2386#pullrequestreview-43282991
    • https://github.com/PaddlePaddle/Paddle/pull/2386#pullrequestreview-43889009
    • https://github.com/PaddlePaddle/Paddle/pull/2386#pullrequestreview-44186486
  • New Parameterupdater: https://github.com/PaddlePaddle/Paddle/pull/2413#pullrequestreview-43530995
  • Convert reader to RecordIO: https://github.com/PaddlePaddle/Paddle/pull/2407#pullrequestreview-43001836
  • StringPiece: https://github.com/PaddlePaddle/Paddle/pull/2432#pullrequestreview-43280013

typhoonzero(wuyi)

• Paddle cloud:
  • Master, optimizer PR reviews
  • Public dataset encryption: https://github.com/PaddlePaddle/Paddle/pull/2447
  • Automatically send email when registering
  • GPU Cluster deployment, re-install CentOS7(WIP)
• MPI GPU+RDMA Cluster binary update.
  • TODO: port RDMA to GitHub, add RDMA support docker images
• Catch up with current surveys and look into expression templates.

Yancey1989(yanxu)

• PaddleCloud
  • Document about how to make demo from book running on cloud: https://github.com/PaddlePaddle/cloud/issues/147
  • Code review for master client: https://github.com/PaddlePaddle/Paddle/pull/2429
  • Code review for PFSServer docker file: https://github.com/PaddlePaddle/cloud/pull/148#pullrequestreview-43899850
  • Discuss the data encrypt on Cloud

gangliao

• fix stringpieces under Mac OS via add ostream header https://github.com/PaddlePaddle/Paddle/pull/2446
• Prepare and present a talk named Memory Management under different deep learning systems, for instance, PaddlePaddle, Majel, Caffe2 and MXnet. [Incomplete] Add Tensorflow as complementary.
• Why don‘t you guys adopt the GPU implementation of Eigen's Tensor'? https://github.com/caffe2/caffe2/issues/798

livc(Zhao Li)

• Start using PaddlePaddle Cloud to train gesture data because the newly acquired data is very large.
• Research image caption and video caption of PaddlePaddle implementation, which can be added to the models repository. And we can design some interesting demos based it, then publish them on the WeChat official account.

kuke (Yibing Liu)

• DS2: CTC Beam Search Decoder

  • PR: https://github.com/PaddlePaddle/models/pull/59
  • Efficiency improvement:
    • line profiling
    • multiprocessesing implementation
    • pruning (doing)
  • Parameters tuning:
    • decrease WER (Word Error Rate) from 22% to 17%
  • Test: numerical stability and unit test (doing）

• PR Review:

  • https://github.com/PaddlePaddle/models/pull/82

Xinghai Sun

• Refactor DeepSpeech2 data pre-processer (https://github.com/PaddlePaddle/models/pull/91)
• PR Reviews:
  • https://github.com/PaddlePaddle/models/pull/74#pullrequestreview-43344709
  • https://github.com/PaddlePaddle/models/pull/82#pullrequestreview-43317811
• Discuss on potential reasons for the current DS2 performance gap.

wanghaoshuang

• Image
  • Implement a crop layer in paddle for fcn model (finish coding) https://github.com/PaddlePaddle/Paddle/issues/2470
  • Make image classification demo run on IDL GPU cluster.

yangyaming

• DS2
  • Add error rate calculation script (wer and cer)
    https://github.com/PaddlePaddle/models/pull/82
  • Add unittest for ci
    https://github.com/PaddlePaddle/models/pull/88
• SSD:
  • Refine DetectionUtil
    https://github.com/PaddlePaddle/Paddle/pull/2357
  • Add DetectionMAPEvaluator
    https://github.com/PaddlePaddle/Paddle/pull/2467
  • Bug fix for ADU
• Code Review:
  • https://github.com/PaddlePaddle/models/pull/78

Shaoyong Xu

• Deep Speech2:
  • Audio data augmentation
    https://github.com/PaddlePaddle/models/pull/99
  • Review: https://github.com/PaddlePaddle/models/pull/91#pullrequestreview-43972033

Cao Ying

• bug fix of PaddleBook after Paddle updates:
  • https://github.com/PaddlePaddle/book/pull/330
• Paddle/Models
  • review and merge a new model: neural language model, including the text generation process:
    • https://github.com/PaddlePaddle/models/pull/60#pullrequestreview-4283389
  • bug fix after Paddle updates:
    • https://github.com/PaddlePaddle/models/pull/85
  • We have merged several models, their learning performances are not validated, I am working on it.
  • Yongfeng gives some advice on making the documents more friendly to beginners and making the way users use these models easier, I am checking and formalize the merged examples one by one
    • rewrite text classification codes and doc
      • https://github.com/PaddlePaddle/models/pull/83
    • refactor NER:
      • https://github.com/PaddlePaddle/models/compare/develop...lcy-seso:refine_ner

hedaoyuan

Refactor convolution code
This week's work is three convolution-related code refactoring. The convolution function based on Gemm has been completed. BlockExpandLayer refactoring based on the ImageExpand function is being written.There is a Issue that describes the calculation of the size of the convolution output in the paddle.

• https://github.com/PaddlePaddle/Paddle/pull/2282 -- done
• https://github.com/PaddlePaddle/Paddle/pull/2449 -- doing
• https://github.com/PaddlePaddle/Paddle/issues/2460 -- todo

Code Review

• https://github.com/PaddlePaddle/Paddle/pull/2373
• https://github.com/PaddlePaddle/Paddle/pull/2354
• https://github.com/PaddlePaddle/Paddle/pull/2445

luotao

• Financial market forecasting on S&P500 dataset
  • study a paper using SVR(support vector regression) method, the classification accuracy > 60%
  • survey several deep learning method , but the best accuracy < 60%
  • The gap comes from data preprocessing.
• update rnn related doc from v1-api to v2-api: https://github.com/PaddlePaddle/Paddle/pull/2464

Liu Yiqun

• Learn the implementation of float16 and int8 in caffe2 and majel

Dang Qingqing

• DS2.

  • Update and Merge PR
    • https://github.com/PaddlePaddle/models/pull/74
    • https://github.com/PaddlePaddle/Paddle/pull/2373
  • Do experiments to verify the convergence for batch shuffle and variable-input.

• PaddlePaddle Refactor

  • Survey other framework, mainly caffe2, about how to do multi-gpu and how to manage the variables.
  • Discuss with YuYang. Review PR

• Code View about Image

  • https://github.com/PaddlePaddle/models/pull/52
  • https://github.com/PaddlePaddle/models/pull/58

yuyang

• Give A PR to show what's my opinion about neural network refactoring. #2445
  • Talks to qingqing and longfei, thanks them help me found some weakness of this PR, and let them help me give some evidence.
• Investigating Sparse Representation for Neural Network Ops.

fengjiayi

• Learn implements of Array and View in Majel, and try to port Array to Paddle.
• Survey on expression template and prepare for a presentation.

qiaolongfei

• PaddleCloud
  • new parameterupdater use paddle pserver cclient of go
  • https://github.com/PaddlePaddle/Paddle/pull/2413
• review:
  • Modify pserver client C API, create better test
  • https://github.com/PaddlePaddle/Paddle/pull/2433
• survey:
  • Mxnet Graph构建分析
  • https://github.com/PaddlePaddle/Paddle/wiki/Mxnet-Graph%E6%9E%84%E5%BB%BA%E5%88%86%E6%9E%90

Xingzhaolong

• test a model on cifar10 dataset on static pruning of paddle.
• Investigation on mobilenet.
• auto pruning on paddle

qijun

• survey the relationship between Tensor library and deep learning System, and have made a report
• make some new commits on porting allocation from majel to paddle

guosheng

• Image
  • Complete the tool for model conversion from Caffe to PaddlePaddle with usage document. Validate the tool on VGG16, ResNet50 and ResNet101. The PR have got merged.https://github.com/PaddlePaddle/models/pull/58

Zhihong Dong

• Develop
  • polish code with review, ready for merge https://github.com/PaddlePaddle/Paddle/pull/2386
  • merge go part, update, optimizer together.
  • add fault-tolerance etcd - doing
• PR Reviews:
  • master client https://github.com/PaddlePaddle/Paddle/pull/2468
  • https://github.com/PaddlePaddle/Paddle/pull/2413
• Refactoration Survey:
  • nnvm Graph

Yan Chunwei

• Refactoration Survey
  • mxnet::engine survey
  • DOING wiki about graph executor engine survey of different platforms