Client2Vec: Improving Federated Learning by Distribution Shifts Aware Client Indexing

If our project helps you, please give us a star ⭐ and cite our paper!

Requirements:

torch~=1.13.1
rich~=13.3.5
torchvision~=0.14.1
torchaudio~=0.13.1
numpy~=1.24.3
visdom~=0.2.4
Pillow~=9.4.0
scipy~=1.10.1
pandas~=2.0.1
matplotlib~=3.7.1
faiss-cpu~=1.7.4
scikit-learn~=1.2.2
pynvml~=11.5.0

Acknowledgement

We are grateful for the following awesome projects:

• FL-bench

Training

The code is built on top of FL-bench. To run the code, you need to first download and split the datasets using data\generate_data.py. Then

1. generate CLIP embeddings by

python datapreprocess/preprocess.py -d [dataset]

2. generate client index and sample orders by

python datapreprocess/train.py -d [dataset] -lr [learning rate] -e [epochs] -tp [global/local]
python datapreprocess\generate_sample_order.py

Please config the datapreprocess\generate_sample_order.py before run.

3. Run the experiments using the scrips like the follows:

python fedavg.py -d cifar10 -m res18 -ge 100 -tg 1 -le 5 -mom 0.0 -bs 32  -wd 5e-5  --seed 42 --MWU_aggregate 1 --MWU_tau 1.0 --local_reg 1 --sample_order_path ../../datapreprocess/sample_orders/cifar10/global/transformer/e100_lr0.01_div_1/greedy/10-mean-mean-1.0.pkl --client_index_path cifar10/global/transformer/e500_lr0.01_div_1 -lr 1e-2 --local_reg_weight 5.0 --MWU_momentum 0.9;

More examples can be found in src\server\scripts

Bibliography

If you find this repository helpful for your project, please consider citing:

@article{guo2024client2vec,
  title={Client2Vec: Improving Federated Learning by Distribution Shifts Aware Client Indexing},
  author={Guo, Yongxin and Wang, Lin and Tang, Xiaoying and Lin, Tao},
  journal={arXiv preprint arXiv:2405.16233},
  year={2024}
}