RecipFL

How Few Davids Improve One Goliath: Federated Learning in Resource-Skewed Edge Computing Environments. WWW 2024. [paper][video]

Run Experiments

Datasets

The datasets will be downloaded automatically to the data folder upon their first use. To change the directory, adjust --data_dir in run_client.py accordingly.

Environment

In our experiments, we use the following specific versions of Python libraries:

networkx==1.11
numpy==1.13.3
pandas==0.22.0
scikit_learn==1.4.0
torch==1.3.0
torch_summary==1.4.5
torchsummary==1.5.1
torchvision==0.4.1
tqdm==4.63.0
adapter-transformers==3.1.0

Run

1. Launch client(s) to await requests from the server. Specify the port by --port and the gpu device by -g. Note that total_clients will be automatically allocated across connected ports, so it is possible to simulate all clients in a single process. Upon execution, the command provides the client IP, which is required as an argument (--client_ip) in the next step. For example, to launch clients on port 8361 and GPU 6, run:

   python run_client.py --port 8361 -g 6

2. Start the training on server. Specify the task by -t, the gpu device by -g, the client IP by --client_ip (replace xxx.xxx.xx.xxx with the IP obtained from the first step), the client ports by --cp (should match the ports used in the first step). Configure other experiment hyperparameters as needed (e.g., --distill, --scaling. See run_server.py for details). For example, to launch server training on CIFAR-10 using GPU 6, run:

   python run_server.py -t cifar10 -g 6 --distill --scaling width --client_ip xxx.xxx.xx.xxx --cp 8361 --cp 8362

Citation

Please cite the following paper if you found our framework useful. Thanks!

@inproceedings{zhang2024few,
  title={How Few Davids Improve One Goliath: Federated Learning in Resource-Skewed Edge Computing Environments},
  author={Zhang, Jiayun and Li, Shuheng and Huang, Haiyu and Wang, Zihan and Fu, Xiaohan and Hong, Dezhi and Gupta, Rajesh K and Shang, Jingbo},
  booktitle={Proceedings of the ACM on Web Conference 2024},
  pages={2976--2985},
  year={2024}
}