-
Notifications
You must be signed in to change notification settings - Fork 5.6k
2018 06 20
-
brpc
-
others:
-
Inference
- 子图 for TensorRT 初步跑通,基于图的 analysis 优化框架基本成型
- OPEN feature/analysis to support subgraph for TRT
- OPEN feature/trt add softmax converter
- MERGED bugfix/trt-op
- MERGED feature/pass manager
- 高层 API output buffer 统一管理
- MERGED inference/unify output buffer management
- 图像3个demo 完成一个,其他基本类似
- OPEN WIP demo/mobilenet inference
- bugfix
- MERGED bugfix/add_inference_lib_to_release
- MERGED bugfix/anakin-ci
- 子图 for TensorRT 初步跑通,基于图的 analysis 优化框架基本成型
-
文档
-
reviews
https://github.com/PaddlePaddle/continuous_evaluation/pull/70 https://github.com/PaddlePaddle/Paddle/pull/11466#pullrequestreview-128691509 https://github.com/PaddlePaddle/continuous_evaluation/pull/52#pullrequestreview-129019309 https://github.com/PaddlePaddle/continuous_evaluation/pull/69#pullrequestreview-129045735 https://github.com/PaddlePaddle/Paddle/pull/11525/files#r195896649 https://github.com/PaddlePaddle/Paddle/pull/11537#pullrequestreview-129570511
- Train Pyramidbox model
- Refine infer to use the pyramid method
- Evaluate Pyramidbox model
-
text classification benchmark
-
checkpoint m2: pserver checkpoint about lookup table
-
Optimized user API about checkpoint
- Refine preprocessor:
- Support transform dynamic shape images to RecordIO: https://github.com/PaddlePaddle/Paddle/pull/11570
- Support uint8 bilinear interp resizing: https://github.com/PaddlePaddle/Paddle/pull/11576
- API reference:
- reader support test: https://github.com/PaddlePaddle/Paddle/pull/11390
- docs:
- ut fixes: https://github.com/PaddlePaddle/Paddle/pull/11470
-
PR
- [Merged] fix mkldnn compile issue https://github.com/PaddlePaddle/Paddle/pull/11571
- [Merged] refine the initial cpu memory flag for mkldnn https://github.com/PaddlePaddle/Paddle/pull/11525
- [Merged] initial with only 1 mkl/openblas threads for each pthreads https://github.com/PaddlePaddle/Paddle/pull/11464
- [Merged] doc https://github.com/PaddlePaddle/Paddle/pull/11497
- [Merged]fix unknown use_mkldnn flag https://github.com/PaddlePaddle/Paddle/pull/11395
- [Merged] scope thread safe https://github.com/PaddlePaddle/Paddle/pull/11258
-
code Review
- MKLDNN: SoftmaxGrad Op https://github.com/PaddlePaddle/Paddle/pull/11519
- MKLDNN layout: Support for activation operator https://github.com/PaddlePaddle/Paddle/pull/11124
- MKLDNN layout: Support for sum operator https://github.com/PaddlePaddle/Paddle/pull/11102
- bnorm+relu fuse for mkldnn (inference) https://github.com/PaddlePaddle/Paddle/pull/11434
- fix compiler error in contrib/tape https://github.com/PaddlePaddle/Paddle/pull/11466
- mkldnn need more memory https://github.com/PaddlePaddle/Paddle/pull/11481
-
issue
- 解决mkldnn在gcc4.8.2编译问题,编译paddlepaddle出错 https://github.com/PaddlePaddle/Paddle/issues/11565
- mixed precision training (fp16 training)
- fix memory optimize/polish document
- debug memory optimize/memory leak in paddle cloud
- Document enhance
- add 20 API reference
- 20+ non-layer module reference
- Training Benchmark of Fluid CPU multi-thread (with @zhaochengduo),similar with V2: https://github.com/PaddlePaddle/Paddle/issues/11620
- OCR: batch inference on OpenBlas version is similar to the online version.
- doc:
- add doc of sequence_softmax and parallelDo: https://github.com/PaddlePaddle/Paddle/pull/11488
- add doc for inference_transpiler: https://github.com/PaddlePaddle/Paddle/pull/11521
- add url of cuda9.0_cudnn7_avx_mkl library:https://github.com/PaddlePaddle/Paddle/pull/11595
- fix compiler error in contrib/tape: https://github.com/PaddlePaddle/Paddle/pull/11466
- code review:
- bugfix/add_inference_lib_to_release: https://github.com/PaddlePaddle/Paddle/pull/11482
- Add a ctest hung tool: https://github.com/PaddlePaddle/Paddle/pull/11471
- MKLDNN:
- [Merge] remove mkldnn flag from gtest strdup for cpu: https://github.com/PaddlePaddle/Paddle/pull/11395
- [merge] MKLDNN layout: Support for activation operator: https://github.com/PaddlePaddle/Paddle/pull/11124
- [merge] MKLDNN layout: Gaussian random layout: https://github.com/PaddlePaddle/Paddle/pull/11523
- [merge] MKLDNN layout: Support for sum operator: https://github.com/PaddlePaddle/Paddle/pull/11102
- [review] bnorm+relu fuse for mkldnn (inference) : https://github.com/PaddlePaddle/Paddle/pull/11434
- Metric learning[WIP]
- OCR
- Document
- Review
- NMT:
- Add weight sharing into Transformer and refine code to adapt to WMT14' en-de settings.
- Add README for Transformer (WIP).
- Experiments on WMT14 en-fr dataset
- Review:
- Face Detection:
- Fix concat_op CUDA kernel bug: https://github.com/PaddlePaddle/Paddle/pull/11541
- Expose crop op into Python API: https://github.com/PaddlePaddle/Paddle/pull/11546
- Pyramidbox inference: https://github.com/PaddlePaddle/models/pull/992
- Train and evalute model
- [WIP] speed training:
-
- use memory optimization to increase batch size: https://github.com/PaddlePaddle/Paddle/issues/11605
-
- more profiling
-
- Batch Norm fix:
- Doc:
- Code revew:
- fake_quantize_op: https://github.com/PaddlePaddle/Paddle/pull/11359#pullrequestreview-130230036
- ICNet: https://github.com/PaddlePaddle/models/pull/975#pullrequestreview-128714172
- SSD Doc: https://github.com/PaddlePaddle/models/pull/987
- pyramidbox eval: https://github.com/PaddlePaddle/models/pull/986
- performance
- [merged] overlap memcpy in multiple rpc ops, improve 15% performance on vgg16 model,https://github.com/PaddlePaddle/Paddle/pull/11221
- [WIP] independent cuda stream in rpc op
- stability and code polish
- ocr distributed model, https://github.com/Yancey1989/paddle_fluid_benchmark/tree/master/ocr_recognition
- bug fix and code polish
- dist train on ce valid diff value, https://github.com/PaddlePaddle/paddle-ce-latest-kpis/pull/48
- new distribute op details folder opertors/distributed, https://github.com/PaddlePaddle/Paddle/pull/11603
- hidden bcast_params while dist train,https://github.com/PaddlePaddle/Paddle/pull/11575
- fix pserver sub-blocks bug, https://github.com/PaddlePaddle/Paddle/pull/11585
- Polish Documentation
- Add script to check API changes
- Design Python data feeding method
- [WIP]Implement Python data feeding in both C++ and Python side, and self-testing is about to be carried out.
- Find a bug in
tensor_py.h
, which may cause "multiple definition" error when compiling if it is included in two or more compilation unit. It is caused by implementing fully specialized template functions in header filetensor_py.h
. Addinginline
before these functions would fix it. Ready to create a PR.
- Seq2seq Attention Model:
- Fixing sequence_expand_op https://github.com/PaddlePaddle/Paddle/pull/11618
- Help Chan with book chapter
-
paddle fluid framework
- add keep_kids flag for executor https://github.com/PaddlePaddle/Paddle/pull/11524
-
paddle document
- update
split_lod_tensor
,create_array
andarray_length
doc - Add none layers api doc https://github.com/PaddlePaddle/Paddle/pull/11532
- update
-
distributed lookup table
- Feed Ctr模型配置完成,基于线上数据的data reader完成,训练跑通。
- 用小数据跑训练任务,能正常收敛,还需要在大数据上验证。
- concat_op support data layer as input https://github.com/PaddlePaddle/Paddle/pull/11527 https://github.com/PaddlePaddle/Paddle/pull/11579
- support Ftrl optimizer https://github.com/PaddlePaddle/Paddle/pull/11547
- optimize auc calculate https://github.com/PaddlePaddle/Paddle/pull/11574
-
MERGE: Fix hang up bug when the 2nd GPU's memory is full
-
MERGE: Fix the bug that sub blocks in lr decay op
-
OPEN: unit test about gracefully shutdown in paddlepaddle/paddle:latest-dev takes no effect
-
OPEN: Update Doc: Parameter Split Strategy in Distribute Transpiler
-
OPEN: Update Design Doc: Distributed Block Design
- detection_map文档修复
- 其他API文档注释
- Fix thread-safety
- Compatibility
- debug transformer 多机
- Parallel Executor
- Enhance Parallel Executor stable
- Doc
- DataReader, fluid.nets, Parallel Executor, ParameAttr
- Small fix
- Fix concat compile warning
- Enable more type for SplitOp and ConcatOp
- Add FAQ(libmkldnn.so is not found when importing paddle.fluid)
- Review
- small thread-safety fix and doc improvemen
- Make the CUDA kernel of concat correct and fix unit tests
- concat support data as input
-
CE duty problem following
-
CE platform migration
-
python api monitor script
-
paddle machine management
-
Fix some problems in 10+ api reference
-
Add python api for argsort operator
-
Transformer: Setup model training of en-fr translation model
-
DeepASR:
- Update open sourced training data
- Support multi-thread parallel decoding
Code Review:
- NMT:
- fix convergence bug for distributed transformer:need broadcast parameters among GPUs and learning rate operater bug
- performance problem with distributed transformer, multi-machine is lower than single machine.
- issue: https://github.com/PaddlePaddle/Paddle/issues/11568
- CTR:
- discussion on distributed training of ctr model
Communicating with ShanYi and Wang Guan to prepare PaddlePaddle.org new release. Targeting release next Friday.
-
PRs:
- Rearrange the feature layout: https://github.com/PaddlePaddle/PaddlePaddle.org/pull/498
- Add a new "Community" links on the top navigation bar : https://github.com/PaddlePaddle/PaddlePaddle.org/pull/499
-
Reviews:
-
Project and Issues Management - PaddlePaddle.org
- https://github.com/PaddlePaddle/PaddlePaddle.org/projects/5
- Update the homepage background section: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/495: 4 Nicky
- Update the homepage with a "Feature” section: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/496: 4 Jeff
- Update the homepage with a "Model" section: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/491: 8 Daming
- Update the homepage with a "Resource" section: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/492: 8 Jeff
- Update the homepage with a "First Deep Learning Experience": https://github.com/PaddlePaddle/PaddlePaddle.org/issues/490: 16 Daming
- Update the homepage with an new installation guide: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/489: 4 Jeff
- Add "About" on to the top navigation: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/486: 8 Nicky
- Add "Community" on to the top navigation: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/487: 2 Jeff
- Add "Blog" to the top navigation: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/488: 24
- Consolidate Documentations: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/493: 24 Varun
- Version select: https://github.com/PaddlePaddle/PaddlePaddle.org/issues/494: Varun
- Add lazy execution feature in tape: https://github.com/PaddlePaddle/tape/pull/18
- Add data feeder in tape: https://github.com/PaddlePaddle/tape/pull/22
- Review:
- New version of deployment of docs on PRs on other repos: https://github.com/PaddlePaddle/PaddlePaddle.org/pull/481
- Working on designing / feedback on new version of PaddlePaddle.org
- Dynamic Graph: https://github.com/PaddlePaddle/tape
- Update chapter 6 https://github.com/PaddlePaddle/book/pull/539
- Rewrite book chapter 8 machine translation https://github.com/PaddlePaddle/book/pull/552
- Review:
- Finished 3 chapters for the new book version (Fluid)
- Started working on PaddlePaddle.org new version
- Review:
- Learn imitation learning
- read paper
- setup race game training environment for imitation learning
- think and discuss how to apply it on our boat
- need a switch between self drive and guided drive.
- if current/wind is big, need to make that part of the state.
- need accurate sensor, perhaps GPS is a good one.
- Reading papers Yi referred for imitation learning
- Reading book chapters to get basic knowledge about machine learning
- Tape
- Add dropout, batchnorm, and pool2d op: https://github.com/PaddlePaddle/tape/pull/34
- Add adam optimizer: https://github.com/PaddlePaddle/tape/pull/36
- Add accuracy function and test on mnist: https://github.com/PaddlePaddle/tape/pull/40
- Update paddle submodule to a fork: https://github.com/PaddlePaddle/tape/pull/25
- Review:
- https://github.com/PaddlePaddle/tape/pull/39#pullrequestreview-132935747
- https://github.com/PaddlePaddle/tape/pull/38#pullrequestreview-132640078
- https://github.com/PaddlePaddle/tape/pull/31#pullrequestreview-132226068
- https://github.com/PaddlePaddle/tape/pull/28#pullrequestreview-131386567
- https://github.com/PaddlePaddle/tape/pull/26#pullrequestreview-131366295