search

med-air / FedBN

[ICLR'21] FedBN: Federated Learning on Non-IID Features via Local Batch Normalization
227 stars 36 forks source link

The train/val split in domainnet #6

Open FrankZhangRp opened 1 year ago

FrankZhangRp commented 1 year ago

we find that the train sets are kept the same size on all domains as the 'min_data_len' and we think there is a mistake: https://github.com/med-air/FedBN/blob/1caf83cf95e6859b712e1e80c8264009a5067162/singleset/single_domainnet.py#L120

FrankZhangRp commented 1 year ago

if the 'trainset' is only 5% of the total dataset, the training process is too toy and unstable with about 100+ samples per client without pretrained weights. We think the train/val split may be 95% / 5%, but in this code is 5% / 5%.