model.n_classes assumed differently for binary classification.

[mojganmadadi]

Hi,

As far as I saw, for binary classification the model.n_classes should be 2, As input of the args parameters in here shows that. But throughout the code, in several places like loss function initialization and loss calculation it is asumed to be 1 for binary classification. Maybe I am wrong. Could you let me know what is the reason? Is it a bug or is there any reason behind it? thanks in advance.

[milesial]

Hi, it doesn't really matter if you use 1 or 2 classes for binary classification.
1 will use a single channel mask then sigmoid and threshold.
2 will use a 2-channels mask then argmax.
The losses for both are equivalent. CE with 2 classes is equivalent to BCE.

[gordon-n-stevenson]

Fixes a bug in the Dice loss that was n_classes not work

#446