The code runs on Python 3. To install the dependencies, run

pip3 install -r requirements.txt

For MNIST, SVHN, CIFAR-10, and CIFAR-100, the datasets will be downloaded automatically by the torchvision package.

For FEMNIST, manually download the dataset as instructed in https://github.com/TalwalkarLab/leaf and put the train and test datasets extended by .json into dataset_data_files/femnist/train and dataset_data_files/femnist/test respectively.

For CelebA, manually download the celebrity faces dataset img_align_celeba.zip from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html, extract the files, and place the standalone .jpg files into dataset_data_files/celeba/raw/img_align_celeba. Then, manually download the train and test files extended by .json as instructed by https://github.com/TalwalkarLab/leaf and put them into dataset_data_files/celeba/train and dataset_data_files/celeba/test respectively.

For Shakespeare, manually download the dataset as instructed in https://github.com/TalwalkarLab/leaf and put the train and test datasets extended by .json into dataset_data_files/shakespeare/train and dataset_data_files/shakespeare/test respectively.

To test the code:

python simulation.py

All configuration options are given in config.py which also explains the different setups that the code can run with.

The results are saved as CSV files in the results folder. The CSV files should be deleted before starting a new round of experiment. Otherwise, the new results will be appended to the existing file.

Part of this code is inspired by https://github.com/IBM/adaptive-federated-learning,

https://github.com/jz9888/federated-learning,

https://github.com/bolianchen/Data-Free-Learning-of-Student-Networks,

and https://github.com/xternalz/WideResNet-pytorch.

Any suggestion or issue is welcome. Pull request is very welcome.