Your loss Implementation might be wrong

[NaiL23]

You implement loss = self.alpha * (1-y) * distance**2 + \ self.beta * y * (torch.max(torch.zeros_like(distance), self.margin - distance)**2) as your contrastive loss, however, in your dataset split and preprocessing script, you label (genuine, genuine) as 1 and (genuine,forged) as 0, which means when y=0, your loss = alpha * distance between pairs and it will be minimized, but hopefully they should be as far as possible.

loss = self.alpha * y * distance**2 + \ self.beta * (1-y) * (torch.max(torch.zeros_like(distance), self.margin - distance)**2) would be a correct implementation.

[VinhLoiIT]

It makes sense. Actually, the (assumed) author implemented it here, but I don't remember why it is different in my code.

Thanks for pointing it out.

[VinhLoiIT]

As I read the paper, the loss function in the current source code is the same as the formulas (1) in the paper. However, the author said "y is a binary indicator function denoting whether the two samples belong to the same class or not" that it is quite ambiguous