• NOTE: This repository will be updated to ver 2.0 at least in August, 2022.

Federated Averaging (FedAvg) in PyTorch

An unofficial implementation of FederatedAveraging (or FedAvg) algorithm proposed in the paper Communication-Efficient Learning of Deep Networks from Decentralized Data in PyTorch. (implemented in Python 3.9.2.)

Implementation points

• Exactly implement the models ('2NN' and 'CNN' mentioned in the paper) to have the same number of parameters written in the paper.
  • 2NN: TwoNN class in models.py; 199,210 parameters
  • CNN: CNN class in models.py; 1,663,370 parameters
• Exactly implement the non-IID data split.
  • Each client has at least two digits in case of using MNIST dataset.
• Implement multiprocessing of client update and client evaluation.
• Support TensorBoard for log tracking.

Requirements

• See requirements.txt

Configurations

• See config.yaml

Run

• python3 main.py

Results

MNIST

• Number of clients: 100 (K = 100)
• Fraction of sampled clients: 0.1 (C = 0.1)
• Number of rounds: 500 (R = 500)
• Number of local epochs: 10 (E = 10)
• Batch size: 10 (B = 10)
• Optimizer: torch.optim.SGD
• Criterion: torch.nn.CrossEntropyLoss
• Learning rate: 0.01
• Momentum: 0.9
• Initialization: Xavier

Table 1. Final accuracy and the best accuracy

Table 2. Final loss and the least loss

Figure 1. MNIST 2NN model accuracy (IID: top / non-IID: bottom)

Figure 2. MNIST CNN model accuracy (IID: top / non-IID: bottom)

TODO

• [ ] Do CIFAR experiment (CIFAR10 dataset) & large-scale LSTM experiment (Shakespeare dataset)
• [ ] Learning rate scheduling
• [ ] More experiments with other hyperparameter settings (e.g., different combinations of B, E, K, and C)