README.md

Out-of-Distribution Generalization of Federated Learning via Implicit Invariant Relationships

Official PyTorch implementation for the ICML 2023 paper Out-of-Distribution Generalization of Federated Learning via Implicit Invariant Relationships. Our code structure is founded on DomainBed, an out-of-distribution generalization benchmark designed for the centralized scenario, but has been refactored to suit the federated learning.

Requirements

This codebase is written in Python3, specifically developed using Python 3.8.12. We utilize PyTorch version 1.10.0 and CUDA version 10.2. To install the necessary Python packages, you can use:

pip install -r requirements.txt

How to Run

Firstly, download the dataset, such as RotatedMNIST:

python domainbed/download.py --data_dir=./data --dataset RotatedMNIST

Next, train the model. For instance, to execute the FedIIR algorithm on the RotatedMNIST dataset:

python main.py --dataset RotatedMNIST --algorithm FedIIR --num_clients 50 --output_dir train_output

The results will be saved in train_output/RotatedMNIST/FedIIR, including model files *.pkl, the result file *.csv, the standard output file *_stdout, and the standard error file *_stderr. It is worth noting that the main file main.py conducts experiments by treating each domain within the dataset in turn as a test domain once.

The main file main.py also allows overloading the experiment configuration using parser arguments, some of the important ones are as follows:

• --num_clients: the number of total clients
• --sample_num: the number of sampled clients in one communication round
• --global_epochs : the number of global communication rounds
• --local_epochs: the number of local update epochs
• --holdout_fraction: the fraction of each domain set holdout for validation
• --trial_seed: the trial number (used for seeding split dataset)
• --seed: the seed for everything else

Reference Github

• https://github.com/facebookresearch/DomainBed
• https://github.com/inouye-lab/FedDG_Benchmark
• https://github.com/atuannguyen/FedSR

Citation

Please consider citing our paper as

@InProceedings{pmlr-v202-guo23b,
  title = 	 {Out-of-Distribution Generalization of Federated Learning via Implicit Invariant Relationships},
  author =       {Guo, Yaming and Guo, Kai and Cao, Xiaofeng and Wu, Tieru and Chang, Yi},
  booktitle = 	 {Proceedings of the 40th International Conference on Machine Learning},
  year = 	 {2023}
}