Accelerated-Federated-Learning-Over-MAC-in-Heterogeneous-Networks icon indicating copy to clipboard operation
Accelerated-Federated-Learning-Over-MAC-in-Heterogeneous-Networks copied to clipboard

Published 20 hours ago • verified publisher icon ramshi236

About the params and controls explosion caused by the increase of the local_epoch

Open Woo-ll opened this issue 2 years ago • 0 comments

Hi, I have increased the local_epochs from 1 to 4 under the SCAFFOLD algorithm, and the result was non-converged, and the params and controls exploded. Wondering if it could be a problem with the code, or, could it be a problem with the SCAFFOLD algorithm itself? Thanks a lot.

running codes: femnist_dict = {"model": "mclr", "batch_size": 4, "learning_rate": 0.001, "local_epochs": 4, "L": 0, "users_per_round": 20}

the results: Figure_1 Figure_2

Woo-ll avatar Sep 01 '22 14:09 Woo-ll