Affine and momentum of BatchNorm layer

2

You can call bn.eval() which will use the running stats and will not update them. The affine parameters will still be trained.