Paddle-fl

Major Features

* Support CUDA for ABY3 protocol.

* Suport NCCL communication for CUDA mode.

* cuda_demo added.

* Several bugs fixed.

Known Issues

* `CPUPlace` and PrivC protocol are not avaliable in gpu version, use cpu version image/whl instead.

- Adds a new two-party MPC learning protocol: PrivC
- Adds examples of YoutubeDNN with_movielens on ABY3, and Linear & Logistic Regression on PrivC
- Provides APIs of online data sharing and revealing
- Supports underlying communication using GRPC
- Fixes several bugs
- Document updated

- Add more MPC operators: conv2d, max pooling, softmax_with_cross_entropy, embedding, batch_norm, adam_optimizer, XavierInitializer, etc.
- Support more MPC models: CNN, FM, etc.
- Add more data process API: mean normal, one hot encoding.
- Add more model evaluation API: KS statistic, Auc, F1-score.
- Add paddle-serving API in data parallel to support model service after training.
- More examples to follow: fm_with_criteo, lenet_with_mnist, model encryption/decryption, psi demo, deploy_serving_after_training.

- Refactor the code for future update and maintenance.
- Add Federated Learning with MPC, which supports horizontal, vertical and transfer Federated Learning.
- Add load & transfer program from normal model to PaddleFL, supporting more models and scenarios.
- Add document and instructions for all demos to make them easy to follow.
- Provide official docker image, pull and use PaddleFL easily.

Support Kubernetes easy deployment
Add api for LEAF dataset which is in federated settings, supporting benchmark experiments.
Add FL-scheduler, acting as a central controller during the training phase.
Add FL-Submitter to support cluster task submission
Add secure aggregation algorithm
Support more optimizers in PaddleFL such as Adam