This repository provides the code for the paper "FedFisher: Leveraging Fisher Information for One-Shot Federated Learning" by Divyansh Jhunjhunwala, Shiqiang Wang, and Gauri Joshi, published in AISTATS 2024.
Our results can be replicated by running the file main.py. The file takes the following arguments.
• --dataset: Choice of dataset. Possible choices are FashionMNIST, SVHN, CIFAR10, CINIC10, CIFAR100, and GTSRB. Note that for CINIC10, the train and test data first needs to be downloaded from https://github.com/BayesWatch/cinic-10.
• --model: The model to use. Possible choices are LeNet, CNN, and ResNet18.
• --algs_to_run: The one-shot algorithms to run. Note that you can specify more than one algorithm. Possible choices are fedavg, otfusion, pfnm, regmean, dense, fisher_merge, fedfisher_diag and fedfisher_kfac.
• --seed: Seed for reproducibility. The default value is 0.
• --alpha: Heterogeneity parameter when splitting the dataset across clients. The default value is 0.1.
• --num_clients: Number of clients in the setup. The default value is 5.
• --num_rounds: Number of rounds of local training and aggregation. The default value is 1.
• --local_epochs: Number of local epochs run by clients. The default value is 30.
• --use_pretrained: Whether to use a pre-trained model or not. The default value is False.
An example of a command to run main.py is given below:
python main.py --dataset 'FashionMNIST' --model 'LeNet' --local_epochs 30 --algs_to_run 'fedfisher_kfac' 'fedavg' • The pfnm algorithm only works with LeNet and CNN models.
• The --use_pretrained only works when using the ResNet18 model. We have also provided the pre-trained checkpoint resnet_18_tiny_imagenet_40.pt which is a ResNet18 model pre-trained on downsampled 32x32 TinyImageNet dataset.
• We are using a modified ResNet18 architecture without BatchNorm layers to be compatible with all algorithms.
Requirements can be found in the requirements.txt file.
• The code for otfusion algorithm is adopted from https://github.com/sidak/otfusion.
• The code for regmean algorithm is adopted from https://github.com/bloomberg/dataless-model-merging.
• The code for pfnm algorithm is adopted from https://github.com/IBM/FedMA.
• The code for dense algorithm is adopted from https://github.com/zj-jayzhang/DENSE.