Per-pixel focal loss

[RubenS02]

Should the focal loss not be applied per pixel? The reduction occurs when computing the crossentropy loss but shouldn't it occur at the very end?

criterion = nn.CrossEntropyLoss(weight=self.weight, ignore_index=self.ignore_index,
                                    size_average=self.size_average)

logpt = -criterion(logit, target.long()) # logpt is a scalar
pt = torch.exp(logpt)
loss = -((1 - pt) ** gamma) * logpt

should maybe become

criterion = nn.CrossEntropyLoss(weight=self.weight, ignore_index=self.ignore_index,
                                    reduction="none") 

logpt = -criterion(logit, target.long()) # logpt is a tensor
pt = torch.exp(logpt)
loss = -((1 - pt) ** gamma) * logpt
if self.size_average:
    loss = loss.mean()

[liuwenran]

agree with @RubenS02

[ycc66104116]

hi, i'm also curious about this issue. and want to know if this adjustment work successfully?

[RubenS02]

Hi @ycc66104116 , i did not test it, this repo is obsolete now you can find much better models