2017 05 24

wangkuiyi

• math refactorization

  • https://github.com/PaddlePaddle/Paddle/pull/2202#pullrequestreview-39084916
  • https://github.com/PaddlePaddle/Paddle/pull/2195#pullrequestreview-39095263
  • https://github.com/PaddlePaddle/Paddle/pull/2217#pullrequestreview-39257488
  • https://github.com/PaddlePaddle/Paddle/pull/2217#pullrequestreview-39896460
  • https://github.com/PaddlePaddle/Paddle/pull/2251#pullrequestreview-40166854

• dataset

  • https://github.com/PaddlePaddle/Paddle/pull/2242#pullrequestreview-39804245

• DS2

  • https://github.com/PaddlePaddle/Paddle/pull/2255#pullrequestreview-40162866

• v2 API implementation improvement

  • https://github.com/PaddlePaddle/Paddle/pull/2249#pullrequestreview-40176383

• Paddle Cloud

  • https://github.com/wangkuiyi/recordio/pull/3#pullrequestreview-40174138

Xinghai Sun

• DeepSpeech2
  • Project & tasks issues creation: https://github.com/PaddlePaddle/Paddle/projects/17
  • PR for two issues (data provider and simplified model): https://github.com/PaddlePaddle/models/pull/55
  • PR for design doc: https://github.com/PaddlePaddle/Paddle/pull/2255
• Models
  • Reviews:
    • https://github.com/PaddlePaddle/models/pull/36#pullrequestreview-39331875
    • https://github.com/PaddlePaddle/models/pull/43#pullrequestreview-39355307

helinwang

• Paddle Could
  • Master server task handling: https://github.com/PaddlePaddle/Paddle/pull/2245
  • Implement parameter server RPC: https://github.com/PaddlePaddle/Paddle/pull/2243
• RecordIO format improvement and fix: https://github.com/wangkuiyi/recordio/pull/3

Yancey1989(yanxu)

• Paddle Cloud
  • Support HDFS-VFS as backend storage: https://github.com/PaddlePaddle/cloud/pull/60
  • Support CephFS as backend storage: https://github.com/PaddlePaddle/cloud/pull/49
  • Support secret registry: https://github.com/PaddlePaddle/cloud/pull/67
  • Optimize the performance of training in internal k8s cluster(support at least 200 trainer node)
  • Deploy PaddlePaddleCloud on the internal k8s cluster, use mulan HDFS server as the backend storage.
• MPI
  • Find the reason of job running faild.(most is FPE and port is allocated)

GongWeibao

• pfsclient & pfsserver:
  The basic functionality of the pfsclient and pfsserver has been completed.I'm testing it now.It supports ls、rm、cp with regular expression arguments.
  https://github.com/gongweibao/cloud/tree/file_manager_simple/go

typhoonzero(Wuyi)

• PaddleCloud
  • Version 0.1.0-alpha for internal training jobs. 169 nodes CPU cluster now production ready.
  • Usage doc: https://github.com/PaddlePaddle/cloud/blob/develop/doc/usage_cn.md
  • Client release: https://github.com/PaddlePaddle/cloud/releases
  • Cluster monitor: http://nmg01-jpaas-paddle00-wrk99.nmg01.baidu.com:8103/dashboard/db/cluster
  • Running baidu-feed word embedding job with 10billion data with 200 trainer pods tested.
  • 国家工程实验室 cluster timelines and solution discussions. Will release the GPU cluster next week(5 nodes)
• MPI issues:
  • yangquan smart 2.0 cluster issue solved.
  • MPI hang problem always due to FPE or pserver port conflicts.

CaoYing

• code review
  • https://github.com/PaddlePaddle/models/pull/36#pullrequestreview-39372042
  • https://github.com/PaddlePaddle/models/pull/43#pullrequestreview-39377338
  • https://github.com/PaddlePaddle/models/pull/46#pullrequestreview-39636653
  • https://github.com/PaddlePaddle/models/pull/30#pullrequestreview-39913767
  • https://github.com/PaddlePaddle/models/pull/30#pullrequestreview-39954982
  • https://github.com/PaddlePaddle/models/pull/22#pullrequestreview-39919482
  • https://github.com/PaddlePaddle/models/pull/34#pullrequestreview-39937082
• refine some examples in Paddle/modles repo
  • https://github.com/PaddlePaddle/models/pull/49
  • https://github.com/PaddlePaddle/models/pull/53
  • https://github.com/PaddlePaddle/models/pull/54
• Up to now, five examples have been merged into the Paddle/model repo.
• I am working on refactoring the way V2 API parses the network topology after we discussed last week, this work is in the process, but not finished yet.

wanghaoshuang

• image
  • Add flowers dataset reader: https://github.com/PaddlePaddle/Paddle/pull/2242
  • Test image classification model on mpi cluster
• models
  • Create a space and put some documents on baidu wiki to introduce models to users from baidu
  • http://wiki.baidu.com/pages/viewpage.action?pageId=327596395

Liu Yiqun

• Learn the architecture feature and compiler options of NVIDIA DRIVE PX2
• Build the CPU and GPU version of Paddle on NVIDIA DRIVE PX2
• Add the support of using CUDA 8.0 to build warp-ctc: https://github.com/gangliao/warp-ctc/pull/1/files

Yu Yang

• Design doc for 'Parameters class in C++', PR 2249

  • It contains a step by step approach of how to implement Parameters concept in C++.

• Discuess for fixing V2 API Layer Configuration.

  • Design Doc PR 2104
  • Demo implementation PR 2096
  • Thanks Cao Ying for helping me to implement this in Paddle.

• Implementation for Registering paddle::Function.

  • It is a step in refactoring paddle topology.
  • Code PR 2087
    • it is basically done.
    • Need Fix unittests and complete the design doc.
    • It could be ready for reviewing in this week.

Xing Zhaolong

• stay investigate on pruning's paper
• learn the deatils of paddle implementation and improve the pruning module for paddle, this is in the process. https://github.com/PaddlePaddle/Paddle/issues/2025#Extend the pruning

Dang Qingqing

• PR
  • Support variable-dimension input feature for 2D convolution operation.
    • https://github.com/PaddlePaddle/Paddle/pull/2197
    • https://github.com/PaddlePaddle/Paddle/pull/2215
  • others
    • https://github.com/PaddlePaddle/Paddle/pull/2193
• PR View
  • https://github.com/PaddlePaddle/Paddle/pull/2247
  • https://github.com/PaddlePaddle/Paddle/pull/2242
  • https://github.com/PaddlePaddle/Paddle/pull/2216
  • https://github.com/PaddlePaddle/Paddle/pull/2053
• Issue
  • DS2: https://github.com/PaddlePaddle/Paddle/issues/2228
• Image Segmentation
  • Make it clear that we need to develop cropping operation and an elment-wise softmax operation. Not open the new issues yet.
• Help to solve video classification problems for IDL-GNOME when using PaddlePaddle to train. This task is urgent for them and they send an email to ask for help.

luotao

• fix some issue:
  • add doc for AggregateLevel and ExpandLevel: #2200
  • rename AggregateLevel and ExpandLevel: #2218
• code review on models
• finish gitchat discussion, and Wechat PaddlePaddle: 150 fans -> 179 fans
• learn deep speech2 related papers

fengjiayi

• Paddle refactorization
  • add DDim to majel in Paddle: https://github.com/PaddlePaddle/Paddle/pull/2202 (merged)
  • refine majel folder structure: https://github.com/PaddlePaddle/Paddle/pull/2251
• models
  • mpi cluster using doc: http://wiki.baidu.com/pages/viewpage.action?pageId=327596461

hedaoyuan

• Paddle Mobile & Convolution Layer Reconstruction
  https://github.com/PaddlePaddle/Paddle/issues/2196
  https://github.com/hedaoyuan/Function

• Code Review & Bug fix
  https://github.com/PaddlePaddle/Paddle/pull/2214
  https://github.com/PaddlePaddle/Paddle/pull/2197#pullrequestreview-38868698

qijun

• Paddle refactorization
  • port allocation to paddle: https://github.com/PaddlePaddle/Paddle/pull/2217
  • on going: port array to paddle

GuoSheng

• Complete the configuration and README of NER, merged to models: https://github.com/PaddlePaddle/models/pull/30
• Design and basic functions implementations for tools for convert model from Caffe to PaddlePaddle

kuke(Yibing Liu)

• Refine and merge NMT without attention:
• Enable the setting of global threshold for gradient clipping (merged):
  • https://github.com/PaddlePaddle/Paddle/pull/2216
• Figure out how to stabilize the training of models in v2 APIs
  • https://github.com/PaddlePaddle/Paddle/issues/2262
  • https://github.com/PaddlePaddle/Paddle/pull/2254
• Prepare to implement CTC-LM-beam-search decoder
  • (c.f. https://github.com/PaddlePaddle/Paddle/issues/2230)

yangyaming

• Bugs
  • Find and fix bugs in unitest of CrossChannelNormLayer
    https://github.com/PaddlePaddle/Paddle/issues/2240
    https://github.com/PaddlePaddle/Paddle/pull/2247
• SSD Model
  • Refactor code
    MultiBoxLossLayer 70%
    DetectionOutputLayer 30%

dongzhihong

• Paddle cloud
  • joint debug with helin’s pserver service :https://github.com/PaddlePaddle/Paddle/pull/2190
• Paddle debug parameter server hanging probleml https://github.com/PaddlePaddle/Paddle/issues/2224
• discussion
  • optimizer C_Api discussion s https://github.com/PaddlePaddle/Paddle/issues/21681

Shaoyong Xu

• DeepSpeech2
  • Learn papers and documentation about deepspeech2
  • Project & tasks issues creation: https://github.com/PaddlePaddle/Paddle/projects/17
  • https://github.com/PaddlePaddle/Paddle/issues/2227