Code for paper "A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging"
This is the official code for the paper S. Wang, M. Ji, "A lightweight method for tackling unknown participation statistics in federated averaging," in International Conference on Learning Representations (ICLR) 2024.
The code was run successfully in the following environment: Python 3.9, PyTorch 1.13.1, Torchvision 0.14.1
See config.py for all the configurations. Some examples on reproducing the CIFAR-10 results (with 5 different random seeds) and Bernoulli participation are as follows.
python3 main.py -data cifar10 -weighting adaptive -iters-total 50000 -seeds 1,2,3,4,5 -lr 0.1
python3 main.py -data cifar10 -weighting adaptive -k-adaptive 50 -iters-total 50000 -seeds 1,2,3,4,5 -lr 0.1
python3 main.py -data cifar10 -weighting average_participating -iters-total 50000 -seeds 1,2,3,4,5 -lr 0.0562 -lr-global 1.78
python3 main.py -data cifar10 -weighting average_all -iters-total 50000 -seeds 1,2,3,4,5 -lr 0.1 -lr-global 10
python3 main.py -data cifar10 -weighting fedvarp -iters-total 50000 -seeds 1,2,3,4,5 -lr 0.0562
python3 main.py -data cifar10 -weighting mifa -iters-total 50000 -seeds 1,2,3,4,5 -lr 0.0562
python3 main.py -data cifar10 -weighting known_prob -iters-total 50000 -seeds 1,2,3,4,5 -lr 0.0562
By default, the results are saved in results_*.csv.