-
Notifications
You must be signed in to change notification settings - Fork 5.9k
2017 07 26
Tao Luo edited this page Dec 9, 2019
·
1 revision
- paddel::memory
- versioning
- infer shape
- operators
- tensor
- net
- Paddle Cloud:
- Master and Pserver graceful shutdown: https://github.com/PaddlePaddle/Paddle/pull/3062
- Fix client discover pserver: https://github.com/PaddlePaddle/Paddle/pull/2999
- Improve logging: https://github.com/PaddlePaddle/Paddle/pull/2997
- https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-51537224
- https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-51897010
- https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-52489318
- https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-52503132
- Add CopyFrom for CPU/GPU in Tensor and its Unit Test https://github.com/PaddlePaddle/Paddle/pull/3056
- In memory, for clarity, change DLOG to VLOG https://github.com/PaddlePaddle/Paddle/pull/3057
- Replace pool allocator to buddy allocator in Paddle https://github.com/PaddlePaddle/Paddle/pull/3030
- Qianmo Docker for internal users
Review:
- Make PADDLE_ENFORCE and PADDLE_THROW catchable https://github.com/PaddlePaddle/Paddle/pull/3055
- Fix memory deps on device context https://github.com/PaddlePaddle/Paddle/pull/3018
- Move Copy out from memory.h into memcpy.h https://github.com/PaddlePaddle/Paddle/pull/3016
- Fault Tolerant
- Paddle version API: https://github.com/PaddlePaddle/Paddle/pull/2985
- MPI
- Paddle on MPI P40 cluster, done with tcp mode.
- PaddleCloud
- Release 0.1.1-beta.3
- update doc: https://github.com/PaddlePaddle/cloud/pull/248
- bug fix: https://github.com/PaddlePaddle/cloud/pull/234
-
Working in process a major PR
Backwardwith Feng jiayi, Dong Zhihong.- framework::
Backwardcan generate Backward Operators of a Fast-Forward network.- It can ignore any gradient variable in forward operator.
- It can handle shared variable's gradient
- https://github.com/PaddlePaddle/Paddle/pull/3068
- framework::
-
Remove ScopePtr, OpPtr,
-
Fix a bug in SequenceActivation.
-
Make framework::Net -> framework::NetOp, remove framework::PlainNet
- Also add common using header in operators module.
- https://github.com/PaddlePaddle/Paddle/pull/3067
-
Make PADDLE_ENFORCE throw same type exception
EnforceNotMet -
Fix a import error when WITH_GOLANG=OFF
-
Some bug fix:
-
About Paddle Models:
- fix some diff with previous v1 version.
- after the bug of sequence_softmax has been fixed, the training process works well. I trained the model again. It now can generate interesting Chinese poetry correctly. I will add the documentation and update it to Paddle models later.
-
About GNR:
- Read some paper on learning to search.
- make clear every computation details of GNR model.
- I will add four new layers: subsequence_select, sequence_slice, cross_entroy_over_beam (including decoding)
- subsequence_select is in progress
- work with Andrew to first make some simplified model running. https://github.com/lcy-seso/paddle_experiment
-
paddle refactor
- RecurrentOp implementation with @qingqing01 @luotao @longfei
- https://github.com/PaddlePaddle/Paddle/pull/2890
- python unittest with forward completed
- ready to merge the first version
- design of RecurrentOp with varient length sequence support with @luotao @longfei @qingqing01
- RecurrentOp implementation with @qingqing01 @luotao @longfei
-
others
- visualDL, add design of data structures
- Community construction
- paddle blog with @helin ready.
- RecurrentOp:
- clean the code
- use the latest add_op and mul_op to test the forward
- remove ScopePtr/OperatorPtr
- deep speech2:
- Make setup.sh more robust: https://github.com/PaddlePaddle/models/pull/177
- Follow-up Intel Optimization on DeepSpeech2 with @liuyibing
- code review:
Operator --> OpKernel --> Tensor/DeviceContext --> Eigen
Implement some basic OpKernels and add python unittest
[WIP]enable operator gpu unittest
[WIP] Elementwise Operator Functor
fix bug in OpKernel register macro
- use operator context and infer context
- 讨论RNN变长输入设计
- 讨论ViaulDL一些事情
- fix bug in OpKernel register macro
- Add type_alias to import framework into ops
- Make network op
- Make CreateOp in Plain C++ params
- Python Generate OpCreation Methods by OpProto
- PaddlePaddle Mobile
- Face Model
Multithreading performance optimization for recognition model prediction(doning). Study of the face detection model deployed based on Paddle, and paddle is missing slice layer(todo). - https://github.com/PaddlePaddle/Paddle/issues/3034
- https://github.com/PaddlePaddle/Paddle/issues/3037
- https://github.com/PaddlePaddle/Paddle/pull/2449 -- merge
- Face Model
- Bug fix
- PaddleCloud
- PR:
- Integration testing pfs and pcloud :https://github.com/PaddlePaddle/cloud/pull/148
- Fix bugs:
- PR:
- PaddlePaddle:
- Embed etcd into Go for easy unit test: https://github.com/PaddlePaddle/Paddle/issues/2504
- Code review:
- Refactor:
- Add Backward Operator registry : https://github.com/PaddlePaddle/Paddle/pull/2949
- Add Random Operator https://github.com/PaddlePaddle/Paddle/pull/3060
- Feature/Backward https://github.com/PaddlePaddle/Paddle/pull/3068
- bug fix:
- fix gen data shape https://github.com/PaddlePaddle/Paddle/pull/3063
- reviews
- polish gradient registry https://github.com/PaddlePaddle/Paddle/pull/3036
- implement distribute save model https://github.com/PaddlePaddle/Paddle/pull/2973
- backward
- Update gradient op registry mechanism. With @dongzhihong
- [WIP] Refine depends of op_registry to make it support general DAG net. With @yuyang and @dongzhihong
- review:
- Learn the design philosophy and overall architecture of paddle
- Learn the design patterns of Op in paddle
- Contrast the differences in Op design between paddle and caffe2
- Try to update the design document for Op
- PaddlePaddle Refactoring
- DS2
- WarpCTC debug with Yaming: https://github.com/PaddlePaddle/Paddle/issues/3076
- Bug when Multi-GPU infering: https://github.com/PaddlePaddle/Paddle/issues/3073
- Two days off
- FCN
- Debug and tune FCN model demo[WIP]
- Fix bugs[merged]
- DeepSpeech2
- Profiling of DS2 on paddle cloud
- Refine DS2 paddle cloud submit scripts
DS2:
- Train the model based on mfcc audio feature
- Profile the whole training pipline (with@lutao)
- Compare the beam search decoder with that in speech-dl (with@yaming)
- Refine the decoder for deployment: add best_path_decoder, unify the interface
- DS2:
- Debug WarpCTCLayer with @qingqing (WIP)
https://github.com/PaddlePaddle/Paddle/issues/3076 - Wrap python interface for decoder of dl_speech (Done)
- Process 3k hours mandarin data (WIP)
- Help to validate performance for latest PaddlePaddle version (Done)
- Debug WarpCTCLayer with @qingqing (WIP)
- Code Review:
-
pr
-
https://github.com/PaddlePaddle/Paddle/pull/3009 add param_attr for
img_conv_groups(...)in /python/trainer_config_helpers/networks.py
-
https://github.com/PaddlePaddle/Paddle/pull/3009 add param_attr for
-
investigate
- ncnn (a light forward framework)
- survey on supporting Genome Group to use paddle with Guo Sheng.
- bug fix https://github.com/PaddlePaddle/Paddle/pull/2992
PaddleCloud
- Sync master client between passes and fix recordio split https://github.com/PaddlePaddle/Paddle/pull/2948
- paddlecloud domain bug fixes
- dlnel notebook release
- dlnel 10*P40 machines deploy WIP
-
PaddlePaddle Image Enhancement
- Faster R-CNN: fix bug for the before PR, including ROIPoolLayer and ProposalTargetLayer; Complete the RCNNLossLayer:
-
Cooperation with Video Analysis Team in IDL:
- Communication about the model for ActivityNet2017, and some enhancements are needed to satisfy the model. Layers for L2-normalization and clipping will be added. RowL2NormLayer is in developing.
- DS2: task list, experiments, detail verification, and bug fixing.