This code is associated with the paper "Federated Gradient Averaging for Multi-Site Training with Momentum-Based Optimizers".
Clone this repository on all participating sites as well as on the central server. Ensure all sites have ssh access to the central server.
On the server, run server.py, passing in as an argument the dataset to use, either CT or MNIST.
On each client, run one of train_ct.py, train_mnist.py, test_ct.py, test_mnist.py, depending on the experiment. All clients must run the same script; other combinations will result in undefined behavior.
Valid MODE arguments include cyclic, federated, weightavg, local.
An example command line argument to train MNIST with cyclic weight transfer using GPU 0 and port 10203:
python train_ct.py A cyclic 0 /path/to/training/set/ 10203
Please consider citing our relevant papers:
@article{remedios2020FederatedGradientAveraging,
title={Federated Gradient Averaging for Multi-Site Training with Momentum-Based Optimizers},
author={Remedios, Samuel W and Butman, John A and Landman, Bennett A and Pham, Dzung L},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
year={2020},
organization={Springer}
}
@article{remedios2020DistributedDeepLearning,
title={Distributed deep learning across multisite datasets for generalized CT hemorrhage segmentation},
author={Remedios, Samuel W and Roy, Snehashis and Bermudez, Camilo and Patel, Mayur B and Butman, John A and Landman, Bennett A and Pham, Dzung L},
journal={Medical physics},
volume={47},
number={1},
pages={89--98},
year={2020},
publisher={Wiley Online Library}
}