batch dimension in eval.py about pytorch-unet HOT 4 CLOSED

The following dice_coeff(mask_pred, true_mask) call should take two batched elements, so I shouldn't have taken net(img)[0]. It makes sense that it works by removing the [0].

I don't know how it was working for me, it seems the code is wrong, but it worked with one channel in the mask (not 100% sure about that).

from pytorch-unet.

my only assumption that it could be because of flat mask , when mask shape is (h,w), not (c=1,h,w)

from pytorch-unet.

Yeah the dice loss function was designed with the behaviour of n_classes=1 in mind, so it might be different when n_classes > 1.

from pytorch-unet.

I have the same problem, when n_classes = 2, the dice loss can not converge

from pytorch-unet.