pipilurj/dynafed

DYNAFED: Tackling Client Data Heterogeneity with Global Dynamics

This repository contains the source code for the paper DYNAFED: Tackling Client Data Heterogeneity with Global Dynamics. Our paper is accepted by CVPR2023 and is available on arXiv: link.

Table of Contents

• Installation
• Reproducing Results
• Citation

Installation

To use this project, you will need to install the following packages:

• PyTorch: pip install torch
• wandb: pip install wandb
• scikit-learn: pip install scikit-learn

Reproducing Results

To reproduce the results from our paper, follow these steps:

1. Download the datasets (fmnist, cifar, cinic10).
2. Train the model by running the following commands:

# cifar10 experiments
bash experiments/cifar10/cifar10_0.01_serverdistill.sh
# cifar100 experiments
bash experiments/cifar100/cifar100_0.01_serverdistill.sh
# cinic10 experiments
bash experiments/cinic10/cinic10_0.01_serverdistill.sh

Example Results

Credits

We would like to give credit to the following repositories for their code and resources that we used in our project:

• Dataset Distillation by Matching Training Trajectories - we were inspired from the source code for distilling data from expert trajectories.

Citation

If you use our code or data in your research, please cite our paper. You can use the following BibTeX entry:

@article{pi2022dynafed,
  title={DYNAFED: Tackling Client Data Heterogeneity with Global Dynamics},
  author={Pi, Renjie and Zhang, Weizhong and Xie, Yueqi and Gao, Jiahui and Wang, Xiaoyu and Kim, Sunghun and Chen, Qifeng},
  journal={arXiv preprint arXiv:2211.10878},
  year={2022}
}

dynafed's Issues

跑cifar10,异构程度0.01的脚本，我跑出来的实验结果有点奇怪

你好，首先很感谢你能开源Dynafed的源码，Dynafed是一篇很有意思且值得再深入研究的论文，根据全局模型参数轨迹生成仿真数据，这的idea很棒。
因此我把Dynafed当作基线实验，但是我本人复现不出来论文里的实验效果，想知道是不是哪个参数设置错误导致的，我通过运行experiments/cifar10/cifar10_0.01_serverdistill.sh这个脚本，最后得到的精度是在56到58之间，但论文里的精度是在65左右，是这个脚本哪个参数设置错了吗，您能帮忙看一下吗