Loss functions for binary semantic segmentation

[japanj]

Hi,

I saw that the binary semantic segmentation task will be available in next release (0.7.0). I saw the configure_losses function and wonder is it possible to set mode='binary' for JaccardLoss and FocalLoss if task='binary'? Or are there any other reasons why you don't specially specified mode='binary' for them in case of binary semantic segmentation?

def configure_losses(self) -> None:
        """Initialize the loss criterion."""
        ignore_index: int | None = self.hparams['ignore_index']
        match self.hparams['loss']:
            case 'ce':
                ignore_value = -1000 if ignore_index is None else ignore_index
                self.criterion: nn.Module = nn.CrossEntropyLoss(
                    ignore_index=ignore_value, weight=self.hparams['class_weights']
                )
            case 'bce':
                self.criterion = nn.BCEWithLogitsLoss()
            case 'jaccard':
                # JaccardLoss requires a list of classes to use instead of a class
                # index to ignore.
                classes = [
                    i for i in range(self.hparams['num_classes']) if i != ignore_index
                ]

                self.criterion = smp.losses.JaccardLoss(
                    mode='multiclass', classes=classes
                )
            case 'focal':
                self.criterion = smp.losses.FocalLoss(
                    'multiclass', ignore_index=ignore_index, normalized=True
                )

Thanks!

[adamjstewart]

Good catch, I think I completely missed that! Would you like to open a PR to fix this? Just replace mode='multiclass' with mode=self.hparams['task'] like I did in ClassificationTask.

[japanj]

Hi @adamjstewart,

Sure, I created #2690 to fix this.