About the label about polyloss-pytorch

[abhuse]

Hi, 'channel' is number of classes, right? So your labels is already one-hot encoded?

At the moment polyloss-focal expects to labels to hold class ids (e.g. of shape [batch, H,W] in your case). Conversion to one-hot format is done in the following lines:

    class Poly1FocalLoss(nn.Module):
        def forward(self, logits, labels):
            . . . 
            if labels.ndim == 1:
                labels = F.one_hot(labels, num_classes=self.num_classes)

            # if labels are of shape [N, ...] e.g. segmentation task
            # convert to one-hot tensor of shape [N, num_classes, ...]
            else:
                labels = F.one_hot(labels.unsqueeze(1), self.num_classes).transpose(1, -1).squeeze_(-1)
            . . .

If your labels are already one-hot encoded the quick solution would be simply to comment 'if ... else' lines shown above that call F.one_hot function.

With that being said, I am going to add the option to make one-hot encoding optional for scenarios where labels are already one-hot. I think that is a common case.

[abhuse]

You can now set label_is_onehot to True when you initialize Poly1FocalLoss if your labels are one-hot encoded.