-
Notifications
You must be signed in to change notification settings - Fork 5.6k
2018 05 08
- Code cleanup
- Matrix multiplication: https://github.com/PaddlePaddle/Paddle/pull/10449#pullrequestreview-118172871
- pre-commit hook: https://github.com/PaddlePaddle/Paddle/issues/10559
- copyright hook: https://github.com/PaddlePaddle/Paddle/pull/10560
- float16 summarization: https://github.com/PaddlePaddle/Paddle/issues/10315
- build system cleanup: https://github.com/PaddlePaddle/Paddle/issues/10558
- Plan to tuning ParallelExecutor and readers
- tuning ParallelExecutor
- Combine auto-batching and kernel fusion will speed up MultiGPUs computation
- discuessed with Zhao Chengduo & Feng Jiayi. There is no complete design right now.
- tuning readers
- discuessed CustomReader with Feng Jiayi
- tuning ParallelExecutor
- Python Trainer.train by Executor
- [WIP] Code Cleanup
-
Correct several misuse of
TensorCopy
: -
Implementation of
trainer.test()
of new Python API: -
Profile of Python data feeding:
-
[WIP] data preprocessing of C++ data feeding (Implementation of
CustomReader
)
- Travis CI: Enable ccache for docker build in Travis
- Doc: fix broken links in document
- Teamcity CI: fix uncatched test failures.
- Fix check style error:
-
fluid high level api
- Add inferencer infer https://github.com/PaddlePaddle/Paddle/pull/10445
-
distributed lookup table
- prepare envionment and run some test.
-
Paddle support Abacus
- 方案讨论,初步确定了一个大家能接受的迁移方案。
- Abacus系统分成多个模块,逐步替换成Paddle。
- 第一步替换模型配置和forward-backward的部分(Abacus对现有系统进行改造和模型调试,Paddle提供核心框架技术支持)。
- 第二步替换pserver的部分(和ps-lib对接)。
- 后续其他模块逐步切换。
- Abacus系统分成多个模块,逐步替换成Paddle。
- 方案讨论,初步确定了一个大家能接受的迁移方案。
-
Bugfix & Code clean
- fix fluid Metric https://github.com/PaddlePaddle/Paddle/pull/10439
- aws tool generating csv file to save metric data
- https://github.com/PaddlePaddle/Paddle/pull/10542#event-1618830622
- going to keep working on aws CE integration
- tried to fix manyliux build
- Tested seq2seq training and inference with batch_size=1. The performance is poor when batch_size > 1. Still looking into it.
- Found an issue caused by empty tensor:
- [WIP] Adding an op to see if a tensor holds data
- inference:
- implement the framework of convert tensorrt op, and convert between fluid lod_tensor and tensorrt itensor.
- implement convert tensorrt relu op, and its unit-test.
- fix tensorrt_engine compiler error:https://github.com/PaddlePaddle/Paddle/pull/10545
- code review:
- feature/inference analysis dot:https://github.com/PaddlePaddle/Paddle/pull/10494
- Feature/engine refactor:https://github.com/PaddlePaddle/Paddle/pull/10497
- Bump MKL-DNN version from 0.11 to 0.14:https://github.com/PaddlePaddle/Paddle/pull/10501
- fix deadline in machine_translation/README.cn.md:https://github.com/PaddlePaddle/book/pull/519
- Fixed-point quantization
- Survey TensorFlow and https://arxiv.org/abs/1712.05877
- Desing doc: https://github.com/PaddlePaddle/Paddle/pull/10553
- Task issue: https://github.com/PaddlePaddle/Paddle/issues/10551
- Support excluded and included mode for average pooling
- Work with Guoyi to support: 对长内用户的一键式安装whl包 + demo
- memory leak:
- performance of some framework:
- Develop GPU kernel for multiclass_nms_op
- The acceleration effect is not ideal,need to be further optimized
- [WIP]Optimize multiclass_nms_op GPU kernel
- [WIP] MPI Enabled
- [WIP] PServer checkpoint
- float16 demo code and documentation
- Add use_cudnn option in FC function
- Fix bug in save op
- Remove unnecessary tensor copy in load op
- Fix sign unsigned comparison warning
- Add float16 support to conv3d op
- Add float16 support to pool3d op
- Add float16 support to load op
- Review:
- Remove Chinese dataset of CRNN-CTC
- Add mean-IOU op.
- Help VIS with im2sequence supporting the batch inference.
- code cleanup and bug fix
- unit test for distribute transpiler, https://github.com/PaddlePaddle/Paddle/pull/10485
- [WIP]overlap send op and backward ops, https://github.com/PaddlePaddle/Paddle/pull/10550
- Debug decoder for RNN Search (with QingSheng)
- [WIP] Add SequencePadOp
https://github.com/PaddlePaddle/Paddle/pull/10533 - Code review for code cleaning
https://github.com/PaddlePaddle/models/pull/894#pullrequestreview-118592003
- NCCL2 dist train, testing perf with RDMA: https://github.com/PaddlePaddle/Paddle/pull/10349
- serde code polish: https://github.com/PaddlePaddle/Paddle/pull/10531
- testing training with LARS: https://github.com/PaddlePaddle/Paddle/pull/10374
- Add documents about training with RDMA hardware.
- Refine quotes' format and comments & add new operators
- Fix the fetch var bug when the arg is renamed
- Adapt the convertor to tensorrt backend
- [WIP] Add Inception_v4 model config in Fluid API
- NMT:
- Transformer code clean, configuration and data utility.
- Beamsearch of Transformer(WIP)
- Inference
- feature/convert tensorrt io
- refector inference engine
- feature/analysis node representation
- init Inference top APIs
- discuss with Anakin team
https://github.com/PaddlePaddle/Paddle/pull/10144#pullrequestreview-117504134 https://github.com/PaddlePaddle/Paddle/pull/10414#pullrequestreview-117565473 https://github.com/Superjomn/paddle-ce-latest-kpis/pull/7#pullrequestreview-117902306 https://github.com/PaddlePaddle/Paddle/pull/10448#pullrequestreview-117916086 https://github.com/PaddlePaddle/Paddle/pull/10461#pullrequestreview-118194601 https://github.com/Superjomn/paddle-ce-latest-kpis/pull/8#pullrequestreview-118245105 https://github.com/PaddlePaddle/Paddle/pull/10495#pullrequestreview-118594253
PR:
Bug fix:
-
Add classification_cost to v2 documentation
-
Fix some image deadlinks in fluid documentation
-
V2 deadlinks
issue:
- Two deadlinks in fluid documentation
[WIP]: Documents for Hsigmoid
- precision alignment in image classification
- refine batch norm op
- https://github.com/PaddlePaddle/Paddle/pull/10502
- se-renext float64 support, for debugging the precision alignment
- https://github.com/PaddlePaddle/Paddle/pull/10407
- https://github.com/PaddlePaddle/Paddle/pull/10346
- https://github.com/PaddlePaddle/Paddle/pull/10450
- [WIP] 累计误差会影响精度,op 高精度对齐op_test support, 对齐se-renext中op GPU
PR
- Fix CI(Merged)
- Feature/update sparse parameter(Merged)
- Updated different gradients on different cards
- Clean Parallel exe
- Fix fetch_op_handle(Merged)
- Wrap shfl_x_sync(Merged)
Review:
- Rewrite Matmul, make code cleaner
- Add float16 support to conv3d op
- Add float16 support to pool3d
- Correct tensor copy misuse
- Inference Framework on Server
- [Merged] Add the call of DropKids at the end of executor.Run to delete the local scopes created in operators
- [Reviewing] Fix a bug related to dispensable inputs and refile the inference unittest
- Profile the performance of OCR's recognization model
- Remove the unused fetch_targets (ctc_align op is slow)
- Update the Paddle Fluid library to newest version
- Use the low level api, such as CreateVariables, Prepare and RunPreparedContext
- The performance of recognization part is almost the same with v1 version, but the linking of Paddle Fluid library slows down the detection part.
- monitor for models repo
PaddlePaddle
- Simplified API: Made fit_a_line mixed version ( trainer.train() ) work. Working on word2vect
- Doc: Steps of building latest PaddlePaddle (with local change) from scratch
VisualDL
- [WIP] Deploy playground server.
- PRs reviewed:
Review and feedback for website navigation Handle Dnspod account transfer. Currently working with Parker to resolve the domain record issue.
-
PRs
- Convert all executable file to lowercase: https://github.com/PaddlePaddle/VisualDL/pull/449
- Fix the toolbox text overflow issue: https://github.com/PaddlePaddle/VisualDL/pull/445
- Provide zoom level control on Embedding: https://github.com/PaddlePaddle/VisualDL/pull/448
- Add save and load params: https://github.com/PaddlePaddle/Paddle/pull/10386
- Move all new API tests to high-level-api folder: https://github.com/PaddlePaddle/Paddle/pull/10534
- [Test-driven] Label Semantic: start testing label_semantic: https://github.com/PaddlePaddle/Paddle/pull/10540
- [Test-driven] Image Classification: https://github.com/PaddlePaddle/Paddle/pull/10537
- [Test-driven] Word2Vect: https://github.com/PaddlePaddle/Paddle/pull/10538
- [Test-deiven] Understand Sentiment: https://github.com/PaddlePaddle/Paddle/pull/10539
- [Test-driven] Recognize Digit: https://github.com/PaddlePaddle/Paddle/pull/10507
-
Feedbacks
- https://github.com/PaddlePaddle/Paddle/pull/10535#pullrequestreview-118891565
- https://github.com/PaddlePaddle/PaddlePaddle.org/issues/475#issuecomment-387567143
- https://github.com/PaddlePaddle/PaddlePaddle.org/issues/475#issuecomment-387188858
- https://github.com/PaddlePaddle/VisualDL/issues/447#issuecomment-387899713
- Qualcomm Android Demo
- Completed SNPE Image Classification Demo with Alexnet trained by Caffe
- Get AlexNet Paddle Model from PaddleCloud team
- Analyze DLC converter error and follow up with Qualcomm team for unsupported ops
- VisualDL
-
Code clean up:
- PR https://github.com/PaddlePaddle/Paddle/pull/10394
- review: https://github.com/PaddlePaddle/Paddle/pull/10390
- review: https://github.com/PaddlePaddle/Paddle/pull/10392
- review: https://github.com/PaddlePaddle/Paddle/pull/10386
- Issue: non deterministic outputs: https://github.com/PaddlePaddle/Paddle/issues/10405
- review: float64 fix: https://github.com/PaddlePaddle/Paddle/pull/10407
-
ONNX:
-
New api:
- PR, semantic roles: https://github.com/PaddlePaddle/Paddle/pull/10368
- PR, recommendation system: https://github.com/PaddlePaddle/Paddle/pull/10535
- Review, fit a line: https://github.com/PaddlePaddle/Paddle/pull/10301
- Review: reorg of new scripts: https://github.com/PaddlePaddle/Paddle/pull/10534
-
FP16:
- PR for save_combine_op: https://github.com/PaddlePaddle/Paddle/pull/10471
- PR LoD enhancement for FP16 test: https://github.com/PaddlePaddle/Paddle/pull/10508
- Review: FP16 in load op: https://github.com/PaddlePaddle/Paddle/pull/10541
- Work on new navigation for documentation (primarily tutorials) after needs assessments and studying other ML frameworks: https://github.com/PaddlePaddle/Paddle/issues/9860#issuecomment-387863602
- Working with Jeff to resolve DNS problems associated with paddlepaddle.org
- Working with Nicky/Qualcomm to prepare ONNX-DLC model for AI day demo
- [WIP] Helping Daming to deploy VDL playground
- Working on new brand assets for PaddlePaddle and VDL
- Fixing documentation (https://github.com/PaddlePaddle/paddle-onnx/pull/35) and feedback on TensorRT backend support update
- [WIP] Survey: control flow in automatical differentiation.
- Speed up CI: https://github.com/PaddlePaddle/Paddle/pull/10506