For training process,
line 148 SR = self.unet(images), SR_probs = F.sigmoid(SR), why just use SR_probs for loss computing. But for ACC, F1 etc. metrics just use SR.
But for evaluating process,
line 209 SR = F.sigmoid(self.unet(images)), just directly use SR_probs for computing.
Is there something wrong?
For training process, line 148 SR = self.unet(images), SR_probs = F.sigmoid(SR), why just use SR_probs for loss computing. But for ACC, F1 etc. metrics just use SR. But for evaluating process, line 209 SR = F.sigmoid(self.unet(images)), just directly use SR_probs for computing. Is there something wrong?