[ICCV2023] FedPerfix: Towards Partial Model Personalization of Vision Transformers in Federated Learning

The code is adapted from https://github.com/TsingZ0/PFLlib

We are still working on improving the readability of the codes and removing unused commands. Please stay tuned!

Environments

python=3.8.0
torch=0.13.1
cuda=11.6

For the dependency:

pip install -r requirements.txt

Preparing Data

Option 1. Create your own split

Change the settings in dataset/generate_{$DATASET_NAME}.py
python generate_cifar100.py noniid - dir # for practical noniid and unbalanced scenario (It may take a while to split the data)

Option 2. Use our split

Download from Google Drive

Setup the wandb

We use wandb to log our results. If you don't want to use it, you can remove the related codes and the results are also saved in ./logs

To reproduce our main results

Go to the system folder

cd system

Run the following commands

Set the dataset and FL settings
data=cifar100 # Name of the data folder
nb=100 # number of classes
nc=64 # number of clients
jr=0.125 # client sample rate
nt=4 # number of thread
config=FedPerfix # Choose from FedAVG, Local, APFL, PerAVG, FedBN, FedBABU, FedRep, VanillaAttention, and FedPerfix.

python main.py -mtd $config -data $data -nc $nc -jr $jr -nb $nb -nt $nt

For the ablation and other results, please check the details of the code.

For Customize settings

Use --local_parts to specify the local parts of the model
Use --basic_model/-vt to change the backbone of the model

Citation

@inproceedings{sun2023fedperfix,
  title={FedPerfix: Towards Partial Model Personalization of Vision Transformers in Federated Learning},
  author={Sun, Guangyu and Mendieta, Matias and Luo, Jun and Wu, Shandong and Chen, Chen},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={4988--4998},
  year={2023}
}

imguangyu/FedPerfix