README

This repository contains the source code of experiments introduced in the paper entitled with FedAPEN: Personalized Cross-silo Federated Learning with Adaptability to Statistical Heterogeneity, which is accepted by ACM SIGKDD Conference on Knowledge Discovery and Data Mining (SIGKDD), 2023.

FedAPEN, FedEN, FML-AE, FML-EE, APFL, FML(Sha), FML(Pri), FedAvg-FT, FedAvg and Individual can be conducted by running main.py with different value of algorithm argument.

• FedAPEN: python main.py --algorithm "learned_adaptive_training" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of FedAPEN.
• FedEN: python main.py --algorithm "equal_training" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of FedAPEN.
• FML-AE: python main.py --algorithm "learned_adaptive" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of FML-AE.
• FML-EE: python main.py --algorithm "learned_adaptive" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of FML-EE.
• APFL: python main.py --algorithm "APFL" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of apfl.
• FML(Sha): python main.py --algorithm "learned_adaptive" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of FML(Sha).
• FML(Pri): python main.py --algorithm "learned_adaptive" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of FML(Pri).
• FedAvg-FT: python main.py --algorithm "fed_avg_tune" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of fed_avg_tune.
• FedAvg: python main.py --algorithm "fed_avg" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of fed_avg.
• Individual: python main.py --algorithm "individual" ..., where the records of accuracy can be revealed by tools/plot_utils.py with the name of individual.

FedRep, FedAMP and FedProx are implemented with an open source project (Temporarily hidden for review) provided in PFL-Non-IID. They can be conducted by running PFL-Non-IID/main.py with different values of the algorithm argument. To run PFL-Non-IID/main.py, you should first change the directory from root of this repository to PFL-Non-IID. Then, execute the commands listed below on your demand.

• FedRep: python main.py --algorithm "FedRep" ..., where the records of accuracy can be revealed by tools/plot_utils.py.
• FedAMP: python main.py --algorithm "FedAMP" ..., where the records of accuracy can be revealed by tools/plot_utils.py.
• FedProx: python main.py --algorithm "FedProx" ..., where the records of accuracy can be revealed by tools/plot_utils.py.

The functionalities of directories and files contained in this repository are listed below:

• models: Defining the architectures of models.
• tools: Containing some tools helping with this repository.
  • nn_utils.py: utilities for neural networks.
  • plot_utils.py: tools to calculate BMTA and plot the trend of mean accuracy.
  • utils.py: tools to decay learning rate.
• PFL-Non-IID: This directory contains an open-source repository for some personalized federated learning approaches. Note that FedRep, FedAMP and FedProx are implemented in this repository.
• data_loader.py: loading datasets and partitioning them to these clients.
• main.py: the entrance of this project.
• node.py: implementation of functionalities for a node of FL, including the server node and client node.
• recorder.py: recording the test accuracy of each client during training.

Requirements

Please see requirements.txt for details.

Note

Some of the PFL baseline implementations (in PFL-Non-IID) are referred to Personalized Federated Learning Platform.

Citation

If you find our work useful for your research or development, please kindly cite our paper.

@inproceedings{qin2023fedapen,
  title={FedAPEN: Personalized Cross-silo Federated Learning with Adaptability to Statistical Heterogeneity},
  author={Qin, Zhen and Deng, Shuiguang and Zhao, Mingyu and Yan, Xueqiang},
  booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={1954--1964},
  year={2023}
}