Hi, just to clarify this is not my implementation, I found it somewhere in kaggle, I cant find the link now. I will try to answer your questions nonetheless:
- Alpha is hyperparameter that you can tune to assign more importance to samples from class A or B. I dont know anything about binary segmentation so correct me if i am wrong, but I assume that it must have at least 2 classes. The class you are trying to segment and background.
- The epsilon is used to avoid numerical instability if probability equals 0. In this case
torch.exp()
will deal with that.
You can find another, perhaps more clear, implementation here