-
Notifications
You must be signed in to change notification settings - Fork 5.6k
2017 07 19
Tao Luo edited this page Dec 9, 2019
·
1 revision
- Expose Tensor to Python
- Rewrite Tensor
- PADDLE_ENFORCE
- Python style
- Do the integration test of PaddleCloud and fix bugs:
- Code review:
- other:
- Repair kubernetes system on America machines
- Add CPU/GPU Memcpy in memory folder https://github.com/PaddlePaddle/Paddle/pull/2970
- FIX: cpplint code style https://github.com/PaddlePaddle/Paddle/pull/2888
- ENH: unify PADDLE_ENFORCE for cublas, cudnn, curand https://github.com/PaddlePaddle/Paddle/pull/2883
Review:
- cmake: fix problem that go_library is never rebuilt. #2954
- enable Intel MKLDNN library and Intel MKL small package #2940
- PR:
- Paddle Cloud:
- Save model: https://github.com/PaddlePaddle/Paddle/pull/2973
- Fix optimizer: https://github.com/PaddlePaddle/Paddle/pull/2855
- Make save model examples no longer compress model: https://github.com/PaddlePaddle/book/pull/342
- Go, cmake and CI:
- Turn on race detector in CI for Go: https://github.com/PaddlePaddle/Paddle/pull/2856
- Reduce cmake build time for Go: https://github.com/PaddlePaddle/Paddle/pull/2884
- Add Go linter pre commit hook: https://github.com/PaddlePaddle/Paddle/pull/2885
- Fix travis Go pre commit check: https://github.com/PaddlePaddle/Paddle/pull/2887
- cmake: fix problem that go_library is never rebuilt: https://github.com/PaddlePaddle/Paddle/pull/2954
- Paddle Cloud:
- Review:
- Paddle Cloud:
- https://github.com/PaddlePaddle/Paddle/pull/2948#pullrequestreview-51000329
- https://github.com/PaddlePaddle/Paddle/pull/2072#pullrequestreview-50478810
- https://github.com/PaddlePaddle/Paddle/pull/2849#pullrequestreview-50220786
- https://github.com/PaddlePaddle/Paddle/pull/2848#pullrequestreview-49864606
- https://github.com/PaddlePaddle/Paddle/pull/2869#pullrequestreview-50126894
- Refactor:
- Paddle Cloud:
- change net to operator
- Optimize ptr
- Add Init to OperatorBase
- Add sgd op
- add ADD_OP_CPU to enable add op with only cpu kernel
- check duplicate of ProtoAndCheckerMaker
- support multiple template parameter for REGISTER_OP_XPU
- [wip]add cross-entropy-op
- [wip]use var name to get variable instead of index
- Feature/add op test
- Skeleton Of fully connected operator
- Make CreateOp in Plain C++ params
- Python Generate OpCreation Methods by OpProto
- Make Python can create variable from a default scope.
- PR #2808
- Design of OperatorWithKernel class
- PR #2815
- Add a sample operator skeletons,
add_op
- PR #2827
- Make OpProto's VarProto support multiple and temporary.
- PR #2860
- Python OpCreation methods
- Polish operator cmake file, add
op_library
function- PR #2897
- Override
Net::DebugString
to make it better for reading.- PR #2912
- Add skeletons of
mul
,rowwise_add
,sigmoid
,softmax
ops- PR #2914
- Make Python can create, read, write
Tensor
- PR #2928
- Change Others PR to make them better
- Make CreateOp can be invoked by plain C++ params
- PR #2936
- Add fully connected operator.
- PR #2945
- FullyConnectedOperator is not a plain operator with kernel. It is a composed/grouped operator.
- Add a operator unittest framework in Python
-
paddle refactor
-
RNN Operator with @qingqing @luotao
- https://github.com/PaddlePaddle/Paddle/pull/2890
- C++ code finished
- writing unittest in python
- waiting for some other dependency operators to be ready.
-
Named getter for operator arguments
- https://github.com/PaddlePaddle/Paddle/pull/2901
- to make argument retrive more convenient
- to hide details of arguments with multiple inputs
-
PADDLE_ENFORCE add a switch for glog/CHECK
- https://github.com/PaddlePaddle/Paddle/pull/2850
- to better display the stack info when ENFORCE failed.
-
-
paddle models
-
DSSM
- merged
- an enhancement
- https://github.com/PaddlePaddle/models/pull/172
- add infer.py
- add more command line arguments support to help configure model structure
- fix bugs
-
CTR model enhancement
- https://github.com/PaddlePaddle/models/pull/138
- decouple data format and model structure
- add infer script
- support classification/regression
-
-
Others
-
http://models.paddlepaddle.org site
- check all html pages and fix some models markdown file format
- support math equation display
- add a navigation bar
-
visualDL with @longfei @yongfeng
- discussion with FE
- determine what features first version support.
-
Operator --> OpKernel --> Tensor/DeviceContext --> Eigen
Implement some basic OpKernels
Tensor to EigenTensor
- https://github.com/PaddlePaddle/Paddle/pull/2805
- https://github.com/PaddlePaddle/Paddle/pull/2953
- https://github.com/PaddlePaddle/Paddle/pull/2978
- PaddlePaddle Refactoring
- refactor recurrent group forward operator with @chunwei @luotao
- implement recurrent group backward operator with @chunwei @luotao
- Code Review
- Add Tensor::CopyFrom and Tensor::mutable_data(Place place)
- Op varient inputs
- Code Review
- FCN
- add crop layer
- Add pixel softmax layer for FCN model
- Add cross entropy layer for pixel classification:
- Add voc2012 dataset for image segment
- models: SSD
- FCN
-
Mobilenet
- optimize the gradInput in the gpu implementation
- support inputchannels != outputChannels
- naive forward implementaion
-
investigate
- shuffleNet
-
review
- RecurrentOp: update the SegmentInput and ConcatOutput function, discuss some functions(slice_ddims, etc) in dims.h with @fengjiayi.
- book:
- add video link in chinese version: https://github.com/PaddlePaddle/book/pull/341
- code review:
- update yixiu with @lizhao
- wechat PaddlePaddle: 324 fans -> 389 fans
- Paddle
- Add a gated unit helper:
- enable recurrent group to output a sequence inside it during generation:
- fix recurrent_group parsing bug:
- Models
- Add a trained text generation model to generate Chinese poetry into Paddle Models. Not finished yet. The learning performance of model trained by V2 API is not comparable with that triained by v1 API. Still checking it.
- PR review: https://github.com/PaddlePaddle/models/pull/36#pullrequestreview-50802703
- Add models repo automated test. PR:
- https://github.com/PaddlePaddle/models/pull/167
- https://github.com/PaddlePaddle/models/pull/170
- Fault Tolerant
- Fix read record failed,https://github.com/PaddlePaddle/Paddle/pull/2848
- PaddleCloud
- Fix notebook session request 404
- PR Review: https://github.com/PaddlePaddle/cloud/pull/219#pullrequestreview-50250328
- PR Review: https://github.com/PaddlePaddle/cloud/pull/212#pullrequestreview-50896371
- Bug fix, https://github.com/PaddlePaddle/cloud/pull/231
- Bug fix, https://github.com/PaddlePaddle/cloud/pull/227
- MPI
- [WIP] Paddle on P40 MPI cluster.
- PaddlePaddle Mobile
- The current progress, integrating Paddle into the face SDK. Performance issues, in the mobile big.LITTLE architecture, multi-threaded computing performance is not as high as expected. And at present, on the Huawei P9 multi-threaded performance test has not yet reached the requirements (less than 450ms).
- https://github.com/PaddlePaddle/Paddle/pull/2816 -- merge
- https://github.com/PaddlePaddle/Paddle/pull/2824 -- merge
- Bug fix
- Review
-
FCN
- Add switch order layer for FCN model
- Add voc2012 dataset for image segment[merged]
- Refine crop layer[merged]
- Debug and tune FCN model demo[WIP]
- Fix bugs[merged]
- Code review
- Add Interfaces for Tensor and simply is implementation
- [WIP] Backward implementation with @zhihong
- PR:
- Refactor:
- Add Backward Operator: https://github.com/PaddlePaddle/Paddle/pull/2949
- Fixes:
- fix unrar package in docker: https://github.com/PaddlePaddle/Paddle/pull/2902
- add start script for ltr model : https://github.com/PaddlePaddle/models/pull/173
- fix dense shape in dataset: https://github.com/PaddlePaddle/Paddle/pull/2905
- add skip layer of loading model interface: https://github.com/PaddlePaddle/Paddle/pull/2904
- fix checkpoint init error: https://github.com/PaddlePaddle/Paddle/pull/2839
- Refactor:
- Review:
- Refactor:
- fully connected op: https://github.com/PaddlePaddle/Paddle/pull/2945
- sgd operator: https://github.com/PaddlePaddle/Paddle/pull/2950
- Paddle Cloud:
- add go testing into cmake: https://github.com/PaddlePaddle/Paddle/pull/2811
- Implement distribute save model: https://github.com/PaddlePaddle/Paddle/pull/2973
- add gometaliner checker: https://github.com/PaddlePaddle/Paddle/pull/2885
- fix optimizer buffer allocation size: https://github.com/PaddlePaddle/Paddle/pull/2855
- race detector on golang https://github.com/PaddlePaddle/Paddle/pull/2856
- Refactor:
- PaddlePaddle Image Enhancement
- Faster R-CNN
- ROIPoolLayer https://github.com/PaddlePaddle/Paddle/pull/2982
- ProposalTargetLayer https://github.com/PaddlePaddle/Paddle/pull/2983
- RCNNLossLayer
- Faster R-CNN
- DS2
- Decayed ADAM (-1.3 WER), decoder refinement (with @Yibing), Mandarin training debug (with @Yaming)
- PR Review https://github.com/PaddlePaddle/models/pull/168#pullrequestreview-51093780
- Investigating ConvS2S, Transformer, SliceNet.
DS2:
-
Add the support of mfcc feature and retrain the model
-
CTC Beam search decoder in C++: covert the compuation of probs in log format
Conv seq2seq:
- Read related papers
- Align LM scoring logic with @yibing
- Cooperate with @yiqun to find numerical bug for wctc
- Polish SSD demo
https://github.com/PaddlePaddle/models/pull/137 - Correct formation for nce_cost/train.py
https://github.com/PaddlePaddle/models/pull/169 - Code review
- Done Fault tolerant on paddle cloud: https://github.com/PaddlePaddle/cloud/pull/212
- Fault tolerant job task fixes: https://github.com/PaddlePaddle/Paddle/pull/2948
- Enable notebook access dlnel.org: https://github.com/PaddlePaddle/cloud/pull/219
- BFE dosn’t support websocket need a workaround
- Clusters: Formal cluster for dlnel on Baidu Cloud, CPU cluster resource exchange request
- MPI cluster mantainous
- Add build_android task on Travis CI
- Update the dockerfile to build cross-compiling image for Android.
- Fix a bug of warp-ctc, that cause float-pointing exception
- Search for a toolchain to cross-compiling Paddle for iOS device on Linux
- genome want to use Paddle instead of caffe in their Android application, and they have some demands, https://github.com/PaddlePaddle/Paddle/issues/2966
- Test the inference speed of Paddle
- Code Review