A from-scratch PyTorch implementation of the Federated Averaging (FedAvg) algorithm introduced in Communication-Efficient Learning of Deep Networks from Decentralized Data.
- Simulated client-server federated learning on MNIST
- Local client training with SGD
- Weighted model aggregation (FedAvg)
- Modular architecture with separate Client, Server, and Model classes
- Timestamped model checkpointing
- Separate training and evaluation scripts
McMahan, B., Moore, E., Ramage, D., Hampson, S., & Aguera y Arcas, B. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data.