when using Dice loss, should the predections be probabilites like [0.3,0.7] or I should set threshold to make them either 0 or 1 like [0,1] (trsh=0.5)

Question

I'm using Unet model and diceloss for binary segmentation with 1 channel. I'm really confused if I should give the dice loss the predections as they are (after applying sigmoid) or I have to set threshold before passing it to the loss function

I tried passing the preds directly as probabilites but i'm not sure