search

med-air / FedBN

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

About how to write febbn test code #9

Open lloveapple opened 1 year ago

lloveapple commented 1 year ago

I have tried various methods to test the code of fedbn (use the trained model, input a picture for prediction), fedbn can not give the correct result, but fedavg is normal

Never2102 commented 7 months ago

Excuse me,have you solved your problem? I got the same question which is if the layers of BN did not join in the aggration, the globle model had no paramters about this layers, so how to test

digbangbang commented 3 months ago

+1, I suppose it use the client to test the acc, maybe in average

MeiruiJiang commented 3 weeks ago

Hi,

Thanks for your interest in our work. When doing testing, we will use client models for each training client. For external clients, we suggest using the averaged BN parameters from training clients, or you can re-estimate the new BN statistics during the test time, please refer to our other work (https://arxiv.org/abs/2110.09974).