Training mask vs testing mask

I noticed that training masks do not contain only two distinct values (0 for background and 1 for lesion) but instead multiple float values. Does testing masks have the same property? Do you threshold testing masks when compute dice score? Thank you!

There should be only one sample that has a non-binary mask, and it should be getting fixed in the next revision.

The full evaluation container can be found here, but the short answer is "no." Submitted masks are clipped to 0-1, but values in that range are left as-is for Dice. The intention is to allow for uncertainty in predictions. You can apply your own threshold before submitting if you prefer.