Hi, in the original implementation there are multiple labels (background, label 1, labels 2, etc..). In my application I have only one class besides background. Therefore I use sigmoid and have logits in the shape (B, H, W, 1). Should I expand my labels and predictions to (B, H, W, 2) where labels/pred[..., 1] = 1 - pred[..., 0]?
Or can I just compute for a single channel since background and foreground are complementary for the binary case?
Hi, in the original implementation there are multiple labels (background, label 1, labels 2, etc..). In my application I have only one class besides background. Therefore I use sigmoid and have logits in the shape (B, H, W, 1). Should I expand my labels and predictions to (B, H, W, 2) where labels/pred[..., 1] = 1 - pred[..., 0]?
Or can I just compute for a single channel since background and foreground are complementary for the binary case?
Thank you, Kind regards