-
Notifications
You must be signed in to change notification settings - Fork 5.6k
2018 01 03
Tao Luo edited this page Dec 9, 2019
·
1 revision
- Distributed Training
- detection_output op（for SSD, done
- norm op done
- run tf ssd demo
- code review
- Send/Recv SelectedRowsm, https://github.com/PaddlePaddle/Paddle/pull/7161
- Return std::future in ThreadPool:Run interface, https://github.com/PaddlePaddle/Paddle/pull/7099
- Implement broadcast scale layer
- (add broadcast operation and group aggregate operation in BaseMatrix)
- https://github.com/PaddlePaddle/Paddle/pull/7188
- Models test
- Add render script to convert template to html report.
- Add script to send report mail in html format
- Add entry script for test job.
Multi-device:
- add helper function to get appropriate DeviceContext
Memory optimization transpiler:
- [WIP] design doc on memory optimization transpiler
Fix and enhance:
- fix compile
- fix bug in TransDataLayout
- Fix bug in SetAttrDescVisitor
- MKL(code review):
- RNN: enable MKL Packed Recurrent Layer: https://github.com/PaddlePaddle/Paddle/pull/6719
- CNN: double check the mklml and mkldnn benchmark, openblas is WIP. https://github.com/PaddlePaddle/Paddle/pull/7104, https://github.com/PaddlePaddle/Paddle/pull/7115
- Compiler:
- add paddle version of pip install and docker: https://github.com/PaddlePaddle/Paddle/pull/7110
- fix library not found for -lrt on MAC: https://github.com/PaddlePaddle/Paddle/pull/7119
- refine pybind when *_op.cc contains several operators: https://github.com/PaddlePaddle/Paddle/pull/7148
- code review:
- update md5 of flowers dataset: https://github.com/PaddlePaddle/Paddle/pull/7098
- Update README.md for image_classification: https://github.com/PaddlePaddle/models/pull/563
- fix build link rt: https://github.com/PaddlePaddle/Paddle/pull/7053/files
- Update edit distance operator
- [WIP] Add sequence erase operator
- Discuss on attention-based ASR and NMT
- Code Review on profiling tool:
- Complete Gradient Check of RNN
- [WIP] Documentation about implementation of Dynamic RNN
- Use DynamicRNN to understand sentiment
- [WIP] Performance tunning about understand sentiment by using DynamicRNN
Enhancements:
- https://github.com/PaddlePaddle/Paddle/pull/7160
- https://github.com/PaddlePaddle/Paddle/pull/7131
- Profiling tool:
- [Update and Merged] Add the event to record the timeline. https://github.com/PaddlePaddle/Paddle/pull/6701
- [Update and Merged] Add design documentation for profiling tool. https://github.com/PaddlePaddle/Paddle/pull/6700
- Image
- Survey the object detection framework of TensorFlow.
- Discussion and arrangement about the cooperation with Visual Technology Department.
- Mobile
- Analyze and optimize the predictive performance of OCR models.
- https://github.com/PaddlePaddle/Paddle/pull/7159
- Refine the activation type in the GRU operator related(Merged)
- Enhance reorder_lod_tensor_by_rank_op to support reorderring Tensor(WIP)
- PR review
- Add API documentation for methods in fluid
- Working on changing the storing format for inference framework of fluid
- Working on NMT using fluid
-
Complete backward doc and add comments
-
Refine backward no_grad_var handling
-
[WIP] Error clipping
-
Reviews:
PR
- Refine cos-sim-op
- Refine conv2d_transpose layer doc
Double caching survey
- tensorflow, caffe2
Review
- fix layout transform
- Feature/transform
- Update the ResNet model for TensorFlow.
- async sendrecv op
-
Muiti Device
- Cpu gpu transform
- [demo] crf run in mixed device env
- [WIP]cpu gpu transform function
-
Review
- Add Fluid Compiler design doc
- feature/print op
- Fluid Heterogeneous device support
- Multi kernel priority selection, Python op_test support multi-kernel
- add Layout Transform, NCHW <=> NHWC
- ContextDevice interface Get(Place) => Get(Place, Library)
- Layout Transform with functional bind
- fix test_LayerGrad
- cudnn operator change to cud kernel
- Fluid benchmark
- The first phase of benchmark
- benchmark models (4/5) done, Fluid paddle books (7/8) done.
- The second phase of benchmark
- collected lego, image, speech profiler and models needs.
- The first phase of benchmark
- scatter functors: https://github.com/PaddlePaddle/Paddle/pull/7045
- update distributed design doc: https://github.com/PaddlePaddle/Paddle/pull/5776
- project https://github.com/PaddlePaddle/Paddle/projects/30?
- Reviews of distributed training related PRs.
- VisualDL
- Graph edges:
- Design plan with FE and Product team
- Code review:
-
SSD Operator port:
-
Reading:
-
Code review:
- Enhance shrink_rnn_memory operator (WIP)
https://github.com/PaddlePaddle/Paddle/pull/7176 - Polish document for lstm_unit wrapper
https://github.com/PaddlePaddle/Paddle/pull/7157 - Add document for max_sequence_len
https://github.com/PaddlePaddle/Paddle/pull/7132 - Add document for lod_rank_table
https://github.com/PaddlePaddle/Paddle/pull/7128 - Code Review
https://github.com/PaddlePaddle/Paddle/pull/7135#pullrequestreview-86122211
-
performance and speed validation of Book.07 label_semantic_roles
- Fix doc:
- Read codes of Fluid, related papers, and initialize the NMT project. https://github.com/PaddlePaddle/Paddle/projects/37
- Fluid
- Add a simple C++ inference example for fluid
- Make a plan to develop the inference library of fluid
- VisualDL
- about to release.
- https://github.com/PaddlePaddle/VisualDL/pull/13
- https://github.com/PaddlePaddle/VisualDL/pull/47
- https://github.com/PaddlePaddle/VisualDL/pull/50
- https://github.com/PaddlePaddle/VisualDL/pull/54
- https://github.com/PaddlePaddle/VisualDL/pull/55
- https://github.com/PaddlePaddle/VisualDL/pull/52
-
Book.05 recommendation systems speed performance comparison with V2.
-
[WIP] Implement adaptive softmax.
- Discussion of the planning of inference library
- https://github.com/PaddlePaddle/Paddle/issues/7221
-
Documentation
- Adding API docs for ones and zeros methods https://github.com/PaddlePaddle/Paddle/pull/7150
-
PR Review
- https://github.com/PaddlePaddle/Paddle/pull/7151#pullrequestreview-86275401
- https://github.com/PaddlePaddle/Paddle/pull/6963#pullrequestreview-85468533
- https://github.com/PaddlePaddle/Paddle/pull/6862#pullrequestreview-86264489
- https://github.com/PaddlePaddle/Paddle/pull/6859#pullrequestreview-86264384
- https://github.com/PaddlePaddle/Paddle/pull/7206#pullrequestreview-86731680
-
Working on NMT using fluid
PR:
- https://github.com/PaddlePaddle/Paddle/pull/6862
- https://github.com/PaddlePaddle/Paddle/pull/7151 review:
- https://github.com/PaddlePaddle/Paddle/pull/7157
- Fixed some compilation errors and unit test failure (https://github.com/PaddlePaddle/VisualDL/pull/51)
- Add VisualDL as a library in paddleboard. Integrate into python wheel build (https://github.com/PaddlePaddle/board/pull/25)
- Add VisualDL SDK example to paddleboard server (https://github.com/PaddlePaddle/board/pull/30)
- add script to build VisualDL for local development (https://github.com/PaddlePaddle/board/pull/35)
- Visualization research and prototyping
- PaddleBoard project initiation preparation.