[NeurIPS 2023]Federated Learning with Bilateral Curation for Partially Class-Disjoint Data
- PyTorch >= 1.0.0
- torchvision >= 0.2.1
- scikit-learn >= 0.23.1
Here we provide the implementation on SVHN, Cifar-10 and Cifar100 datasets. The three datasets will be automatically downloaded in your datadir.
As for model used in the paper, we use the same model structure ResNet18 modified for 32x32 input as MOON.
Parameter | Description |
---|---|
proxy |
Type of last layer of classifier you used (cls for FedAvg and etf for our FedGELA). |
model |
The model architecture. Options: simple-cnn , resnet18 . |
dataset |
Dataset to use. Options: CIFAR10 . CIFAR100 , SVHN |
lr |
Learning rate. |
batch-size |
Batch size. |
epochs |
Number of local epochs. |
n_parties |
Number of parties. |
party_per_round |
number of active clients in each round. |
comm_round |
Number of communication rounds. |
beta |
The concentration parameter of the Dirichlet distribution for non-IID partition. Setting 100000 as IID |
datadir |
The path of the dataset. |
logdir |
The path to store the logs. |
seed |
The initial seed. |
temperature |
Temperature in MOON and Ew in our paper. |
mu |
Param of baselines. |
Here is an example to run FedGELA on CIFAR10 with ResNet18:
python FedGELA.py --dataset=cifar10 \
--partition='dirichlet' \
--temperature=0.00001 \
--lr=0.01 \
--epochs=10 \
--model=resnet18 \
--comm_round=100 \
--n_parties=50 \
--beta=0.5 \
--party_per_round=10 \
--logdir='./logs/' \
--datadir='./data/' \
--proxy='etf' \
We borrow some codes from MOON and FedSkip.
If you have any problem with this code, please feel free to contact zqfan_knight@sjtu.edu.cn or ziqingfan0331@gmail.com.
If you find it useful, please cite as following:
@inproceedings{fan2023federated,
title={Federated Learning with Bilateral Curation for Partially Class-Disjoint Data},
author={Fan, Ziqing and Zhang, Ruipeng and Yao, Jiangchao and Han, Bo and Zhang, Ya and Wang, Yanfeng},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023}
}