predicted mask format

[Feanor007]

During training, should I convert the output from the Unet to a binary mask prior feeding into soft_dice_cldice? I used raw logits, but the training result is really bad.
However when I tried to convert to a binary mask, I lost the grad on the tensor.

[jocpae]

Hi, thanks for reaching out. To make it short, you should not convert the output to a binary mask. Did you train using only the dice loss as well (alpha=0? What do your results look like in this case? :)

[Feanor007]

@jocpae Thanks for replying. I think I fixed this by just only applying sigmoid to the logits. You are right, I shouldn't convert the output to only 0 and 1. I trained with alpha = 0.5. I found that the results are better when combine soft_dice_cldice with other loss such as bce

[jocpae]

Good to hear :)
I do not know what kind of data you are working on but in our experience, it was always a good idea to test a wide range of alpha values empirically to find out what best suits your dataset. I am happy to comment on visual results if you would like to share them :)