This repository contains the PyTorch implementation of FADAS: Towards Federated Adaptive Asynchronous Optimization (accepted by ICML 2024).

Pytorch 2.0.1 CUDA 12.0

To run the experiment for FADAS without delay-adaptive:

python3 asyncfl_forward.py --model=resnet --dataset=cifar10 --gpu=0 --iid=0 --optimizer=fedams --local_lr=0.1 --lr=0.0001 --num_users=50 --frac=0.5 --local_bs=50 --local_ep=2 --epochs=500 --scale=1 --update_freq=5 --dir_alpha=0.3 --delay_type=large_delay

To run the experiment for FADAS with delay-adaptive:

python3 g_delay_adaptive.py --model=resnet --dataset=cifar10 --gpu=0 --iid=0 --optimizer=fedams --local_lr=0.1 --lr=0.001 --num_users=50 --frac=0.5 --local_bs=50 --local_ep=2 --epochs=500 --scale=1 --update_freq=5 --dir_alpha=0.3 --delay_type=large_delay --tauc=1 

The default values for various paramters parsed to the experiment are given in options.py.

• -dataset: Default: 'cifar10'. Options: 'mnist', 'fmnist', 'cifar100'.
• -model: Default: 'cnn'. Options: 'mlp', 'resnet', 'convmixer'.
• -gpu: To use cuda, set to a specific GPU ID.
• -epochs: Number of rounds of training.
• -local_ep: Number of local epochs.
• -local_lr: Learning rate for local update.
• -lr: Learning rate for global update.
• -local_bs: Local update batch size.
• -iid: Default set to IID. Set to 0 for non-IID.
• -num_users: Number of users. Default is 100.
• -frac: Fraction of users to be used for federated updates. Default is 0.1.
• -optimizer: Default: 'fedavg'. Options: 'fedadam', 'fedams'.
• -update_freq: The buffer size for the asynchronous update. Default: 5.
• -scale: The parameter for Dirichlet sampling in the delay simulation. Default: 1.0.
• -delay_type: The type of large or mild delay for the experiments. Set delay_type=large_delay for the large worst case delay
• -tauc: The threshold for delay adaptive learning rate

Please check our paper for technical details and full results. If you find this repo useful for your research, please consider citing the paper.

@inproceedings{
  wang2024fadas,
  title={{FADAS}: Towards Federated Adaptive Asynchronous Optimization},
  author={Yujia Wang and Shiqiang Wang and Songtao Lu and Jinghui Chen},
  booktitle={Forty-first International Conference on Machine Learning},
  year={2024},
  url={https://openreview.net/forum?id=j56JAd29uH}
}