I am working on a semantic segmentation problem for a medical dataset. I am using soft dice loss that focus on the positive class pixels, as implemented here: https://arxiv.org/abs/2209.06078 . Here is my code: