Repository navigation
DiceLoss add weight per class parameter #7065
Description
Currently DiceFocalLoss(), has "focal_weight", which reweights class contributions (but only for the focal loss part).
Similarly DiceCELoss(), has "ce_weight" parameter for the same thing of CE loss part.
- Let's standardize the naming of this argument to be the same for all losses, e.g. simply "weight"
- Let's add this "weight" parameters to DiceLoss() class, if provided to re-weights contributions of individual per-component dices. And modify DiceFocalLoss() and DiceCELoss() to propagate this "weight" to the Dice loss initialization.
This can be achieved by modifying DiceLoss() after this line
Line 182 in 14fcf72
with an additional code snippet (or similar)
if self.weight is not None:
# make sure the lengths of weights are equal to the number of classes
class_weight: Optional[torch.Tensor] = None
num_of_classes = target.shape[1]
if isinstance(self.weight, (float, int)):
class_weight = torch.as_tensor([self.weight] * num_of_classes)
else:
class_weight = torch.as_tensor(self.weight)
if class_weight.shape[0] != num_of_classes:
raise ValueError(
"""the length of the `weight` sequence should be the same as the number of classes.
If `include_background=False`, the weight should not include
the background category class 0."""
)
if class_weight.min() < 0:
raise ValueError("the value/values of the `weight` should be no less than 0.")
f = f * class_weight.to(f)
thank you
Activity
fix Project-MONAI#7065
Add weight in `DiceLoss` (#7098)
fix Project-MONAI#7065
Refactor `self.weight` in `DiceLoss` and `FocalLoss` (#7158)