About gan loss

[zhangqianhui]

Hello @taesungp @junyanz

contrastive-unpaired-translation/models/networks.py

Line 362 in afdc8fb

self.loss = nn.BCEWithLogitsLoss()

I see you use both vanilla gan and nonsaturating with softplus

When you use BCEWithLogitsLoss:

For real samples:

L_{r} = BCE(logits1, 1) + BCE(logits, 0)
= -(1 * log(sigmoid(logits1)) + 0 * log(1 - sigmoid(logits1)) + 0 * log(sigmoid(logits2)) + 1 * (1 - sigmoid(logits2)))
= -(log(sigmoid(logits1)) + log(1 - sigmoid(logits2))))

For fake samples:

L_{f} = BCE(logits2, 1)
= -(log(sigmoid(logits2))

Obviously, it is else nonsaturating loss.

Thus, I don't understand your defination of for vanilla gan. Why do you both nonsaturating gan with two different funtions?
Maybe I misunderstand.

[taesungp]

You are absolutely right about that. It was merely an artifact of combining old codes of CycleGAN with the new pieces of codes. Sorry for the confusion.