Closed HazardFY closed 2 years ago
Yes, from our empirical experiments, using BatchNorm might hurt the performance in FL. Since there are learnable parameters in BatchNorm, averaging those BatchNorm statistic is not optimal for FL. We usually avoid to use it. But some works, such as FedBN, show local BatchNorm can overcome non-iid in personalized FL scenario. See https://arxiv.org/pdf/2102.07623.pdf.
Thanks for your sharing! But I do notice that the Unet uses instanceNorm instead of BatchNorm2d and I don't understand why. Is that related to the framework of federated learning?