Federated Learning with Flexible Control (FlexFL)

This is the official code of the paper S. Wang, J. Perazzone, M. Ji, K. Chan, "Federated learning with flexible control," in IEEE International Conference on Computer Communications (INFOCOM), May 2023.

@inproceedings{wang2022federated,
  title={Federated Learning with Flexible Control},
  author={Wang, Shiqiang and Perazzone, Jake and Ji, Mingyue and Chan, Kevin S},
  booktitle={IEEE International Conference on Computer Communications (INFOCOM)},
  year={2022}
}

The code was run successfully in the following environment: Python 3.8, PyTorch 1.7, Torchvision 0.8.1

See config.py for all the configurations. Some examples are as follows.

FashionMNIST dataset with 20 random seeds (proposed method in the first line and baseline with k_r=0.01 in the second line):

python3 main.py -data fashion -compression-adaptive-method lyapunov -seeds 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20 -out results_fashion_proposed

python3 main.py -data fashion -compression-adaptive-method fixed-0.01 -seeds 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20 -out results_fashion_baseline_0.01

CIFAR-10 dataset with 5 random seeds (proposed method in the first line and baseline with k_r=0.01 in the second line):

python3 main.py -data cifar -compression-adaptive-method lyapunov -seeds 1,2,3,4,5 -out results_cifar10_proposed

python3 main.py -data cifar -compression-adaptive-method fixed-0.01 -seeds 1,2,3,4,5 -out results_cifar10_baseline_0.01

The results are saved in results_*.csv by default and the prefix results can be changed to another value by specifying the -out argument.

This code was inspired by and derived from past work with other collaborators, such as:

• https://github.com/PengchaoHan/EasyFL/
• https://github.com/IBM/adaptive-federated-learning