Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (FedDure)

Setup

conda create --name fssl python=3.9.8
conda activate feddure

Alternatively, you can use pip install -r requirements.txt to install the dependencies.

Download Dataset

We provide experiment with two datasets: Cifar10 and CINIC10.

Training

We provide the running scripts in configs, which allow you to reproduce the results on our paper. Make sure you change the data path (root) in .yaml files.

Cifar10

Slurm cluster:

• sh train_slurm.sh vi_irdc 1 ./configs/cifar10/res9_cifar10_dir_dir_5.yaml

Workstation

• bash train.sh ./configs/cifar10/res9_cifar10_dir_dir_5.yaml

You run the commands for CINIC10 in a similar manner by changing the scripts.

References

If you use this code, please cite

@inproceedings{bai2024combating,
  title={Combating data imbalances in federated semi-supervised learning with dual regulators},
  author={Bai, Sikai and Li, Shuaicheng and Zhuang, Weiming and Zhang, Jie and Yang, Kunlin and Hou, Jun and Yi, Shuai and Zhang, Shuai and Gao, Junyu},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={10},
  pages={10989--10997},
  year={2024}
}