TsingZ0/FedKTL

CVPR 2024 accepted paper, An Upload-Efficient Scheme for Transferring Knowledge From a Server-Side Pre-trained Generator to Clients in Heterogeneous Federated Learning

Introduction

This is the implementation of our paper An Upload-Efficient Scheme for Transferring Knowledge From a Server-Side Pre-trained Generator to Clients in Heterogeneous Federated Learning (accepted by CVPR 2024).

Key words: pre-trained generative model, knowledge transfer, federated learning, data heterogeneity, model heterogeneity

• Poster
• Slides From another perspective: Generative Model-Assisted Collaborative Learning

Take away: We introduce FedKTL, a Federated Knowledge Transfer Loop (KTL) that (1) transfers common knowledge from a server-side pre-trained generator to client small models, regardless of the generator's pre-training datasets, and (2) shares task-specific knowledge among clients through federated learning.

An example of our FedKTL for a 3-class classification task. Rounded and slender rectangles denote models and representations, respectively; dash-dotted and solid borders denote updating and frozen components, respectively; the segmented circle represents the ETF classifier.

Citation

@inproceedings{zhang2024upload,
  title={An Upload-Efficient Scheme for Transferring Knowledge From a Server-Side Pre-trained Generator to Clients in Heterogeneous Federated Learning},
  author={Zhang, Jianqing and Liu, Yang and Hua, Yang and Cao, Jian},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2024}
}

Dataset & Environments

Due to the file size limitation, we only upload the statistics (config.json) of the Cifar10 dataset in the practical setting ($\beta=0.1$). Please refer to our popular repository PFLlib and HtFLlib to generate all the datasets and create the required python environment.

System

• main.py: System configurations.
• total.sh: Command lines to run experiments for FedKTL with default hyperparameter settings.
• flcore/:
  • clients/: The code on clients. See HtFL for baselines.
  • servers/: The code on servers. See HtFL for baselines.
    • serverktl_stable_diffusion.py: the code for using the pre-trained Stable Diffusion on the server.
    • serverktl_stylegan_3.py: The code for using the pre-trained StyleGAN3 on the server.
    • serverktl_stylegan_xl.py: The code for using the pre-trained StyleGAN-XL on the server.
  • trainmodel/: The code for some heterogeneous client models.
• stable-diffusion/ (Other text-to-image models are also supported):
  • pipelines/: The customized pipeline enables the independent operation of the Latent Diffusion Model from other components.
  • v1.5/: The folder to store the pre-trained Stable Diffusion v1.5. Large model files are not included here due to limited space. Please download primary safetensors files into sub-folders from the Hugging Face link. For further instructions on running Stable Diffusion, please see the documentation of the diffusers package.
• stylegan/:
  • stylegan-utils/: Some indispensable utils when using the pre-trained StyleGAN3 and StyleGAN-XL.
  • stylegan-3-models/: The folder to store the pre-trained StyleGAN3 models. Please download the entire models from the following links: StyleGAN3 (pre-trained on AFHQv2), StyleGAN3 (pre-trained on Bench), StyleGAN3 (pre-trained on FFHQ-U), and StyleGAN3 (pre-trained on WikiArt).
  • stylegan-xl-models/: The folder to store the pre-trained StyleGAN-XL. Please down the entire model from this link.
• utils/:
  • data_utils.py: The code to read the dataset.
  • mem_utils.py: The code to record memory usage.
  • result_utils.py: The code to save results to files.