I am working with Nvflare for Federated Learning in medical imaging.
I would like to know if it is preferred to use DataKind.WEIGHT_DIFF with respect to DataKind.WEIGHTS for transferring FL control variables to the server. Is there a fundamental implementation difference, besides the fact of combining weights or their differences ? Are there any advantages of using WEIGHT_DIFF, like smaller errors or faster convergence ?
To sum up, is using WEIGHT_DIFF the best practice, or there is actually no difference, and it depends on the user’s preferences ?
Thanks in advance.
Gonzalo