Clever approach! I would not have thought to use learner.model.add_module(), this seems very handy.
The approach that worked for me was to set learn.crit = F.binary_cross_entropy_with_logits, keeping everything the same.
1 Like