The model is not "turned off during training". With dropout=1.0, for dropout layers you'll get all zero at train and, apparently, identity at test. I don't think pytorch should have allowed dropout=1.0. It should be ValueError, not sure I get the reasoning there.
Leave a Reply