FedLD: Tackling Data Heterogeneity in Federated Learning via Loss Decomposition

Official implementation of "Tackling Data Heterogeneity in Federated Learning via Loss Decomposition" (MICCAI 2024) https://papers.miccai.org/miccai-2024/paper/1348_paper.pdf

Pre-requisites

Set up environment

cd ./FedLD
conda create -n fedld python=3.10
conda activate fedld
pip install -r requirements.txt

Data Preparation

cd ./data

Download the prepared Retina dataset.

gdown https://drive.google.com/uc?id=1bW--_qRZnWbkb0XXvGBCSferdqXZ6pe7

Download the prepared COVID-FL dataset.

gdown https://drive.google.com/uc?id=1cuvoYvt-EVs5qtA5Xgos0yUJmfPhRbwg

Train and Test

cd ./FedLD
python federated_main.py --train_rule FedLD --dataset retina --retina_split 1 --num_users 5 --local_bs 50 --lr 0.01 --epochs 200 --local_epoch 1 --marg_control_loss True --margin_loss_penalty 0.1 --svd True --k_proportion 0.8 --device cuda:0

Citation

If you find our code or paper useful, please consider citing:

@article{zeng2024tackling,
  title={Tackling Data Heterogeneity in Federated Learning via Loss Decomposition},
  author={Zeng, Shuang and Guo, Pengxin and Wang, Shuai and Wang, Jianbo and Zhou, Yuyin and Qu, Liangqiong},
  journal={arXiv preprint arXiv:2408.12300},
  year={2024}
}