The question is about batchnorm

vision
1

I use batchnorm2d to run net,and batch_size=10.
But When I test the net performance , I set batch_size=1，the result very bad.I know the reason is batchnorm2d, and What should I do about batchnorm2d when I train and test net?

2

Have you set your model to eval mode with: model.eval()?

3

Yes I set model.cuda().eval()

4

Is this .eval() will shutdown batchnorm2d?

5

This will make sure to use the running stats on your test samples instead of calculating them from the current batch.

You might play a bit with the momentum, which might help, if your batch stats are quite different.

BatchNorm layers usually work good with a batch size of > 64 so you might try group normalization.

6

Thank you for you help