2017 05 24 - PaddlePaddle/Paddle GitHub Wiki
-
mathrefactorization- 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 -
DS2
-
v2 API implementation improvement
-
Paddle Cloud
- 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
- 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
- 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)
- pfsclient & pfsserver:
The basic functionality of the pfsclient and pfsserver has been completed.I'm testing it now.It supportsls、rm、cpwith regular expression arguments.
https://github.com/gongweibao/cloud/tree/file_manager_simple/go
- 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.
- 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
- 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.
- 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
- 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
-
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.
-
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.
- 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
- PR
- Support variable-dimension input feature for 2D convolution operation.
- others
- PR View
- Issue
- 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.
- fix some issue:
- code review on models
- finish gitchat discussion, and Wechat PaddlePaddle: 150 fans -> 179 fans
- learn deep speech2 related papers
- 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
-
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
- Paddle refactorization
- port allocation to paddle: https://github.com/PaddlePaddle/Paddle/pull/2217
- on going: port array to paddle
- 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
- Refine and merge NMT without attention:
- Enable the setting of global threshold for gradient clipping (merged):
- Figure out how to stabilize the training of models in v2 APIs
- Prepare to implement CTC-LM-beam-search decoder
- Bugs
- Find and fix bugs in unitest of CrossChannelNormLayer
https://github.com/PaddlePaddle/Paddle/issues/2240
https://github.com/PaddlePaddle/Paddle/pull/2247
- Find and fix bugs in unitest of CrossChannelNormLayer
- SSD Model
- Refactor code
MultiBoxLossLayer 70%
DetectionOutputLayer 30%
- Refactor code
- 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
- 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