-
Notifications
You must be signed in to change notification settings - Fork 5.6k
2017 08 23
Tao Luo edited this page Dec 9, 2019
·
1 revision
Key concepts structuring
Code polishment
- https://github.com/PaddlePaddle/Paddle/issues/3622
- https://github.com/PaddlePaddle/Paddle/pull/3581
Add new op
Draft design doc of Python Package
New Operator
- scale, identity https://github.com/PaddlePaddle/Paddle/pull/3592
- minus https://github.com/PaddlePaddle/Paddle/pull/3634
Enhancement of current codebase:
- Resolve a cycle depenedence https://github.com/PaddlePaddle/Paddle/pull/3587
- Rename Net::AddOp --> Net::AppendOp https://github.com/PaddlePaddle/Paddle/pull/3585
Tiny fixes:
- https://github.com/PaddlePaddle/Paddle/pull/3556
- https://github.com/PaddlePaddle/Paddle/pull/3553
- GNR model:
- add the last layer for GNR model: cross entropy over beams with early stop:
- follow comments to update sequence_slice_layer:
- update GNR models:
- all of the implementations have been finished, now I begin to focus on tuning the model.
- complete mnist_demo:
- fix
sgd_op
: - fix
rowwise_add_grad_op
: - reduce some compile warning:
- Fix rowwise add:
- many discussion about python api and the work for next Phase:
-
paddle refactor
- discussion about the design of python syntax wrapper with others
- discussion about refactor tasks
-
paddle models
- large model bug:
- https://github.com/PaddlePaddle/Paddle/pull/3518 (merged)
- Test on dnn model (Baidu Feed, WIP)
- paddle cloud:
- Clusters:
- CPU Cluster: Changed to another 144 nodes, reinstalling OS.
- DLNEL Cluster: GPU resource not releasing bug-[WIP], feature requests.
-
Image:
- Make cudnn convolution layer and projection support for dilation:
-
PaddleCloud:
- Run recognize_digits and understand_sentiment demo with fault tolerant mode:
- Discuss Python API design
- code review
- dynet api survey
- refine random related operators
- Add with refactoring cmake option
Code Reviews:
- https://github.com/PaddlePaddle/Paddle/pull/3620#discussion_r134664308
- https://github.com/PaddlePaddle/Paddle/pull/3592#discussion_r134376072
- Mobile
- Review
-
PaddlePaddle Refactoring
- Add a lookup table op and a CUDA helper
- Merge GPU kernel for cross entropy: https://github.com/PaddlePaddle/Paddle/pull/3501
- Move pybind from package paddle/framework into paddle/pybind:
- code review:
-
Other:
- Tune relative precision for unit test img_conv2 in test_NetworkCompare:
- Code Review:
- dilation conv: https://github.com/PaddlePaddle/Paddle/pull/3629
- ROIPooling: https://github.com/PaddlePaddle/Paddle/pull/2982
- Faster-RCNN Related: https://github.com/PaddlePaddle/Paddle/pull/3496
- 3DConv/Deconv:
- 3DPooling:
- Refactorization:
discuss about the Python API with guys.
- Clarify the concept of
Block
,Network
andrun target
interface. - discuss how initialized op and forward/backward/optimize op should be organized. survey the NCCL library
- Clarify the concept of
- PR & fixes unify operator Input/Output name: design doc. https://github.com/PaddlePaddle/Paddle/pull/3452 add unsigned value in attr https://github.com/PaddlePaddle/Paddle/pull/3623 fix ndcg metrics https://github.com/PaddlePaddle/models/pull/215
- Reviews
init minst demo https://github.com/PaddlePaddle/Paddle/pull/3564
fix rowwise add https://github.com/PaddlePaddle/Paddle/pull/3604 Gather operator https://github.com/PaddlePaddle/Paddle/pull/3540 refine random operator https://github.com/PaddlePaddle/Paddle/pull/3596 fix cudnn path https://github.com/PaddlePaddle/Paddle/pull/3581
- DS2:
- Update network structure (Add GRU support) https://github.com/PaddlePaddle/models/pull/214
- Tuning bigger model: WER for Librispeech -1.7%
- RL:
- Implement A3C with PyTorch (As a preparation for its paddle version)
- Made training plan and documented the training process
- Received training from Helin
- Big picture understanding about Paddlepaddle
- Paddle cloud architect
- Design doc review and update
- DLNEL production system maintaining
- PaddleCloud
- Repo: Custom Kubernetes Scheduler Demo, https://github.com/Yancey1989/k8s-scheduler
- PR: GPU limit, https://github.com/PaddlePaddle/cloud/pull/326
- issue: improve logs storage/view, https://github.com/PaddlePaddle/cloud/issues/335
- issue: Prevent data leaks, https://github.com/PaddlePaddle/cloud/issues/332
- PaddlePaddle Image Enhancement
- Faster R-CNN: refine the ROIPoolLayer.
- Cooperation with Video Analysis Team in IDL:
- Add ScaleShiftLayer to get rid of differences with the Pytorch model. https://github.com/PaddlePaddle/Paddle/pull/3560
- Debug the Attention Net and validate the forward precession. Additionally, another network used in the ensemble model has gotten the same evaluating result with the Pytorch implementation.
- Code Review
- Other
- [WIP] refine Huber loss, add huber_regression_cost: https://github.com/PaddlePaddle/Paddle/pull/3571
- [WIP] V2 API save and load param header: https://github.com/PaddlePaddle/Paddle/pull/3619
- [bug fix] help @yanchunwei to fix dssm infer error with gru: https://github.com/PaddlePaddle/models/pull/212
- code review:
- Add a sequence slice layer: https://github.com/PaddlePaddle/Paddle/pull/3367#pullrequestreview-57111394
- Remove flag use_mkldnn_wgt (Intel MKLDNN): https://github.com/PaddlePaddle/Paddle/pull/3548
- DS2
- Train unicom data on paddle cloud
- Learn paddle cloud
- Prepare environment
- Submit job
- Enhance decoder with @yibing
- Support character based LM
- Support vocabulary filtering
- https://github.com/PaddlePaddle/models/pull/139
- Train unicom data on paddle cloud
- Adaptedto the BatchNorm structure to support 3D data
- Add config parser for Conv3D and DeConv3D
- fix 3D-maxPooling function
- Cooperate with gemone for the Album project
- Profile Paddle, and look into why Paddle is slower than Caffe
- build Paddle for armeabi
- build Paddle with clang
-
Optimize the beam search decoder for deployment
-
Simplify the build process
-
Add the support of multiple threads, reducing evaluation time to 10min
-
Choose better data structure (trie tree) and decoding logic (finite state transducer)
-
Carry on some experiments on changing the text data preprocessing
- Refactor discussion (Python interface)
- Paddle Cloud test demo converge, fixes:
- Bring Chen Xi up to date
- Refactor discussion (GAN and Python API)
- Paddle operators
- gather_op, forward backward
- scatter_op, forward backward
- fix gradient check for in-place operators