FedDyn (WIP)

Getting Started

Downloading datasets through torchvision can be quite slow or can even fail. Therefore, it is recommended that you download the datasets prior to preparing the data.

• CIFAR10
• EMNIST

Results

Custom Datasets

ToDo:

• [x] Add MLP
• [x] Add CNN
• [ ] Write client side training -[ ] Dynamic regularization
• [x] Server side update
  • [ ] Update only parameters & not entire state dict?
• [x] Main FedDyn loop
• [x] Proper communication between server and clients
• [x] Prepare dataset
  • [x] IID
  • [x] Non-IID Balanced
  • [x] None-IID Unbalanced
• [x] Download datasets
• [ ] Metrics (bits transferred, accuracy etc)
• [ ] Parallel training

References

• Communication-Efficient Learning of Deep Networks from Decentralized Data
• Bayesian Nonparametric Federated Learning of Neural Networks