You can't return the loss when in model.eval() mode at the moment. You could change the code though and I think I have seen someone that did exactly that.