2017 05 10 - PaddlePaddle/Paddle GitHub Wiki

• Errands

  1. Switch the base image of Docker files to Ubuntu 16.04: https://github.com/PaddlePaddle/Paddle/pull/2040
  2. Add Majel's concept place to Paddle: https://github.com/PaddlePaddle/Paddle/pull/2091

• Code Review:

  1. Design Doc: The Client Library of Parameter Server

     • https://github.com/PaddlePaddle/Paddle/pull/2075#pullrequestreview-37193920

  2. Design Of Refactor Topology

     • https://github.com/PaddlePaddle/Paddle/pull/1665#pullrequestreview-29350963
     • https://github.com/PaddlePaddle/Paddle/pull/1665#pullrequestreview-29351169
     • https://github.com/PaddlePaddle/Paddle/pull/1665#pullrequestreview-37317850
     • https://github.com/PaddlePaddle/Paddle/pull/1665#pullrequestreview-37321577

  3. ParameterServer Design Doc

     • https://github.com/PaddlePaddle/Paddle/pull/1964#pullrequestreview-37123287

  4. add attr() interface to layer

     • https://github.com/PaddlePaddle/Paddle/pull/2073#pullrequestreview-37054676

  5. Design doc: master server

     • https://github.com/PaddlePaddle/Paddle/pull/1953#pullrequestreview-35687724
     • https://github.com/PaddlePaddle/Paddle/pull/1953#pullrequestreview-36636147

  6. Use KernelType instead of FunctionBase in Layer

     • https://github.com/PaddlePaddle/Paddle/pull/2049#pullrequestreview-36834558

• ISSUE

  • About FileServer's interface

• PR

  • FileManager
  • Modify path format and file format

• Review

  • PServer Design Doc
  • Trainer Inerface

• Paddle cloud:

  • now the paddle cloud website is able to submit a cluster training job by running a notebook.
  • Download user’s certs on the website

• [WIP]Kubernetes high-performance networking researchs https://thenewstack.io/hackers-guide-kubernetes-networking/

• [WIP]Paddle v2 API distributed trianing https://github.com/PaddlePaddle/Paddle/issues/1942 https://github.com/PaddlePaddle/Paddle/pull/2072

• PR reviews:

  • upload-download: https://github.com/PaddlePaddle/Paddle/pull/2013 by @巩伟宝
  • submit job: https://github.com/PaddlePaddle/Paddle/pull/1770 by @闫旭
  • submit job discuss: https://github.com/PaddlePaddle/Paddle/issues/2047
  • pserver:https://github.com/PaddlePaddle/Paddle/pull/1964 by @董志宏
  • master: https://github.com/PaddlePaddle/Paddle/pull/1953 by @helin。

• Issues fixing for submitting MPI paddle job to new cluster at yangquan(baidu feed team). They are willing to try to use our new kubernetes cluster later.

Goals: Code Clean Up for Paddle Layer Configuration and Registration.

• Thinking how to clean up Paddle Code, give a overview documentaion #1665
  • After discuss, maybe user stories are not fit for design. I will change it into following subsections.
    1. What are the problems that this PR will solve? What are the problems this PR paved the way to solve?
    2. How do we solve the problems?
    3. How does other framework solve the similar problems?
• Decoupled the BaseFunction and Layer, prepare for code clean up job #2049
• Give a implementation of Code Clean Up of Function package #2087

• Paddle Cloud

  • Using Paddle Job to submit a distributed job with CPU resource.

• Code Review

  • Design Doc: FileManager
  • Design Doc: ParameterServer
  • Design Doc: Master Server

• PR, issue:

  • bug fix: paddle init failed with env https://github.com/PaddlePaddle/Paddle/pull/2086
  • design doc: submit a distributed job https://github.com/PaddlePaddle/Paddle/pull/1770
  • discuss: how specify resource for distributed training, https://github.com/PaddlePaddle/Paddle/issues/2019

• Paddle V2 api
  • PaddlePaddle v2 api design(https://github.com/PaddlePaddle/Paddle/issues)
  • add attr() interface to layer https://github.com/PaddlePaddle/Paddle/pull/2073

• Build TeamCity Continuous Integration System to Support GPU Unit Test and it works now.
• TeamCity CI will push Paddle GPU/CPU docker images to docker hub automatically.
• Release Paddle v0.10.0 and its docker images
• Read and learn how to define Tensor from Majel source code.

• Paddle v2 demo
  • finish training and predicting language model by ngrm
  • finish training language model by rnn
• Paddle Image
  • discuss tasks for image and prepare to implement converter from caffe to paddle

• Finish two configurations for the seq2seq demo

  • NMT without attention. https://github.com/PaddlePaddle/models/pull/22
  • Grid LSTM. https://github.com/PaddlePaddle/models/pull/38

• Report a bug of v2 api in nested recurrent group
  https://github.com/PaddlePaddle/Paddle/issues/2071

• code review on all pr to the Paddle/models repo
• Support application from FEED to fix bugs (finally, to find it is not Paddle's bug) in training the models. Currently, the training process can stably run. After a discussion, we have made a plan on how to do increamental learning, and Wangsijiang is working on it now.

• Paddle Models:
  • PR: configurations for NMT with external memory (https://github.com/PaddlePaddle/models/pull/36).
  • Issue: discussion for topological parsing problem/bug (https://github.com/PaddlePaddle/Paddle/issues/2061).

• PR：
  • Extend dataset imikolov to support generation of sequence data (merged)
  • Add usage document for hsigmoid layer (https://github.com/PaddlePaddle/models/pull/19)
• Complete network configuration for ntm_addressing_mechanism
• Make a survey of SSD Object Detection model

To promote application in image field: (https://github.com/PaddlePaddle/Paddle/issues/2070)

• Optimize performance of image classification trainer in local mode
• Prepare dataset and reader for image classification training in cluster mode - Make a survey of SSD Object Detection model

• Learn the usage of docker, write a dockfile to cross-compile Paddle for Android, and test the building on Linux server and my Windows 7 laptop.
• Survey the mobile-related demand of face and discuss the work plan with daoyuan (http://agroup.baidu.com/paddlepaddle/md/article/311873).

• finish the gitchat paper: http://agroup.baidu.com/paddlepaddle/md/article/306294
• code review on repo models
• add missing api doc of network and evaluator: #2012
• Wechat PaddlePaddle: 67 fans -> 126 fans, and will group a prize activity.

• Image

  • New issues about some work:

    • https://github.com/PaddlePaddle/Paddle/issues/2070
    • https://github.com/PaddlePaddle/models/issues/28
    • https://github.com/PaddlePaddle/models/issues/27

  • PR about a simple image classification demo for v2 API

    • https://github.com/PaddlePaddle/models/pull/39
    • https://github.com/PaddlePaddle/Paddle/pull/2084

• Code Review:

  • https://github.com/PaddlePaddle/models/pull/31

• Paddle cluster training:
  • create pserver design doc: https://github.com/PaddlePaddle/Paddle/pull/2075
  • discussion:
    • pserver design doc: https://github.com/PaddlePaddle/Paddle/pull/1964
    • master server design doc: https://github.com/PaddlePaddle/Paddle/pull/1953
    • upload and download: https://github.com/PaddlePaddle/Paddle/pull/2013, https://github.com/PaddlePaddle/Paddle/pull/2013
    • submit dist train job: https://github.com/PaddlePaddle/Paddle/pull/1770
• Learn majel tensor library.

• Work on Embedded Has been talk about with Face, and converting their models to Paddle for inference. https://github.com/PaddlePaddle/Paddle/issues/2025

• Other https://github.com/PaddlePaddle/Paddle/pull/2049#pullrequestreview-36759233

• Paddle models: scheduled sampling. PR: https://github.com/PaddlePaddle/models/pull/29
• Paddle models: image classification model. Using v2 API to implement VGG, AlexNet, ResNet and GoogLeNet

• Paddle cloud:
  • review ParameterServer, Trainer DesignDoc and dicussion with helin
  • pserver:https://github.com/PaddlePaddle/Paddle/pull/1964
• Paddle demo -ranknet demo, https://github.com/PaddlePaddle/Paddle/pull/2053, -lambarank demo https://github.com/PaddlePaddle/models/pull/31

• update text classification demo according to review results
• finish demo document -https://github.com/PaddlePaddle/models/pull/24
• help wuyi test v2 api on mpi cluster

• PR: https://github.com/PaddlePaddle/book/pull/313
• Review: https://github.com/PaddlePaddle/book/pull/310#event-1073995644
• Constructed a model to recognize English letters from gesture data.

• Paddle Models: configurations for NER. PR: https://github.com/PaddlePaddle/models/pull/30

• update regression demo according to review results
• finish demo document -https://github.com/PaddlePaddle/models/pull/34