FPPL: An Efficient and Non-IID Robust Federated Continual Learning Framework

Introduction

This repository contains the official source code for the paper FPPL: An Efficient and Non-IID Robust Federated Continual Learning Framework.

@article{he2024fppl,
  title   = {FPPL: An Efficient and Non-IID Robust Federated Continual Learning Framework},
  author  = {Yuchen He and Chuyun Shen and Xiangfeng Wang and Bo Jin},
  year    = {2024},
  journal = {arXiv preprint arXiv:2411.01904}
}

Requirements

To ensure smooth execution of the code, we recommend setting up a dedicated environment using conda.

Steps:

1. First, make sure you have Anaconda or Miniconda installed.

2. Create a new conda environment:

   conda create -n FPPL python==3.9.18

3. Activate the environment:

   conda activate FPPL

4. Install the required packages:

   pip install -r requirements.txt

Datasets

CIFAR-100 dataset can be automatically downloaded with torchvision.datasets.

ImageNet-R and CUB-200 dataset can be downloaded from the link provided in LAMDA-PILOT. Please specify the folder of your dataset in src/utils/conf.py.

Experiments

Set the [DATASET] and [MODEL] options using the filenames of the .json files in the configs folder.

python src/main_fcl.py --dataset [DATASET] --model [MODEL] --device 0

Examples:

python src/main_fcl.py --dataset imagenetr_T20_beta5e-1 --model fppl --device 0

---

python src/main_fcl.py --dataset imagenetr_T20_beta5e-1 --model fppl_wo_debias --device 0

---

python src/main_fcl.py --dataset cifar100_T10_NONIID_cprompt_setting --model fppl_cprompt_setting --device 0

Acknowledgments

This repo is heavily based on LAMDA-PILOT and MarsFL, many thanks.