Awesome

Introduction

This is the official implementation of our paper GPFL: Simultaneously Learning Generic and Personalized Feature Information for Personalized Federated Learning (accepted by ICCV 2023).

• Poster

Citation

@inproceedings{zhang2023gpfl,
  title={GPFL: Simultaneously Learning Global and Personalized Feature Information for Personalized Federated Learning},
  author={Zhang, Jianqing and Hua, Yang and Wang, Hao and Song, Tao and Xue, Zhengui and Ma, Ruhui and Cao, Jian and Guan, Haibing},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={5041--5051},
  year={2023}
}

From the perspective of feature extraction, most existing pFL methods only focus on extracting global or personalized feature information during local training, which fails to meet the collaborative learning and personalization goals of pFL. To address this, we propose a new pFL method, named GPFL, to simultaneously learn global and personalized feature information on each client. We evaluate our GPFL in the label skew, feature shift, and real world settings.

Specifically, we devise a Conditional Valve (CoV) after the feature extractor ($\phi$) to create a global guidance route and a personalized task route in the client model. Then, we learn the global feature information with the guidance of global category embeddings using the Global Category Embedding (GCE) and learn personalized feature information through personalized tasks, as shown in the following figure, where globally shared modules have solid borders.

By simultaneously considering global guidance and personalized tasks, GPFL can learn discriminative features and distinguish the data distribution of each client, revealing excellent personalization performance of GPFL in feature extraction. We show the t-SNE visualization of the learned feature representations as follows.

Please refer to our paper https://arxiv.org/pdf/2308.10279.pdf for more details.

Datasets and Environments

Due to the file size limitation, we only upload the mnist dataset with the practical label skew setting ($\beta=0.1$). One can generate all the required datasets and environment settings via our another project PFLlib.

System

• main.py: configurations of GPFL.
• run_me.sh: start GPFL.
• env_linux.yaml: python environment to run GPFL on Linux.
• ./flcore:
  • ./clients/clientgp.py: the code on the client.
  • ./servers/servergp.py: the code on the server.
  • ./trainmodel/models.py: the code for backbones.
• ./utils:
  • data_utils.py: the code to read the dataset.

Training and Evaluation

All codes corresponding to GPFL are stored in ./system. Just run the following commands.

cd ./system
sh run_me.sh