Skip to content

The is the official implementation of ICCV 2023 paper "No Fear of Classifier Biases: Neural Collapse Inspired Federated Learning with Synthetic and Fixed Classifier".

License

Notifications You must be signed in to change notification settings

ZexiLee/ICCV-2023-FedETF

Repository files navigation

ICCV-2023-FedETF

License: MIT

This is the official implementation of ICCV 2023 paper "No Fear of Classifier Biases: Neural Collapse Inspired Federated Learning with Synthetic and Fixed Classifier (arixv)".

Paper Overview

TLDR: We devise FedETF which is inspired by the neural collapse phenomenon, showing both strong generalization and personalization performances.

Abstract: Data heterogeneity is an inherent challenge that hinders the performance of federated learning (FL). Recent studies have identified the biased classifiers of local models as the key bottleneck. Previous attempts have used classifier calibration after FL training, but this approach falls short in improving the poor feature representations caused by training-time classifier biases. Resolving the classifier bias dilemma in FL requires a full understanding of the mechanisms behind the classifier. Recent advances in neural collapse have shown that the classifiers and feature prototypes under perfect training scenarios collapse into an optimal structure called simplex equiangular tight frame (ETF). Building on this neural collapse insight, we propose a solution to the FL's classifier bias problem by utilizing a synthetic and fixed ETF classifier during training. The optimal classifier structure enables all clients to learn unified and optimal feature representations even under extremely heterogeneous data. We devise several effective modules to better adapt the ETF structure in FL, achieving both high generalization and personalization. Extensive experiments demonstrate that our method achieves state-of-the-art performances on CIFAR-10, CIFAR-100, and Tiny-ImageNet.

image

image


Codebase Overview

  • We implement our proposed FedETF and the FL baselines (FedAvg, FedDF, FedBE, FedDyn, FedProx, FedRoD, FedNH, Ditto, FedRep, FedProto, CCVR, etc.) into the same framework. We note that:
    • Running ''main.py'' for the methods.
    • For CCVR, it needs additional running ''main_runccvr.py'' for the final result since CCVR is a post-hoc method after federated training.
  • We decouple the algorithms into the ''server_method'' and the ''client_method''. Therefore, the server and client methods can be flexibly combined.
  • For convenience, we add the ''set_server_method'' function so that you can only set the client_method and the server_method can be assigned accordingly.
  • The original algorithms can be reproduced by setting:
    FedETF: {'server_method': 'fedavg', 'client_method': 'fedetf'}.
    FedAvg: {'server_method': 'fedavg', 'client_method': 'local_train'}.
    FedDF: {'server_method': 'feddf', 'client_method': 'local_train'}.
    FedBE: {'server_method': 'fedbe', 'client_method': 'local_train'}.
    FedDyn: {'server_method': 'feddyn', 'client_method': 'feddyn'}.
    FedProx: {'server_method': 'fedavg', 'client_method': 'fedprox'}.
    FedRoD: {'server_method': 'fedavg', 'client_method': 'fedrod'}.
    FedProto: {'server_method': 'fedproto', 'client_method': 'fedproto'}.
    FedNH: {'server_method': 'fednh', 'client_method': 'fednh'}.
    CCVR: {'server_method': 'ccvr', 'client_method': 'ccvr'}.
    FedRep: {'server_method': 'fedrep', 'client_method': 'fedrep'}.
    Ditto: {'server_method': 'fedavg', 'client_method': 'ditto'}.
    

Citing This Repository

Please cite our paper if you find this repo useful in your work:

@InProceedings{Li_2023_ICCV,
    author    = {Li, Zexi and Shang, Xinyi and He, Rui and Lin, Tao and Wu, Chao},
    title     = {No Fear of Classifier Biases: Neural Collapse Inspired Federated Learning with Synthetic and Fixed Classifier},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {5319-5329}
}

Contact

Please feel free to contact via email (zexi.li@zju.edu.cn) if you have further questions.

About

The is the official implementation of ICCV 2023 paper "No Fear of Classifier Biases: Neural Collapse Inspired Federated Learning with Synthetic and Fixed Classifier".

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages