Open MSchnei opened 2 years ago
Can we add a unit test, that checks that when loading pre-trained model weights for inference, it doesn't matter if we do or do not use spectral normalisation
I added such a test but, to my surprise, it is failing. I would not have expected the test to fail, to be honest, given that changes should only come in effect once we actually train the model but here we are only performing inference. So I am a bit puzzled at the moment and need more time to understand it.
Implement spectral normalization (SN) for ResNets and ConvNexts.