how to use with sigmoid as activation function when meeting binary classification segmentation task about boundary-loss HOT 3 CLOSED

So, I dont know is it right,

pc = probs[:, self.idc, ...].type(torch.float32)
dc = dist_maps[:, self.idc, ...].type(torch.float32)

change to

pc = probs[:, 0, ...].type(torch.float32)
dc = dist_maps[:, self.idc, ...].type(torch.float32)

from boundary-loss.

Also, I want to know, in the onehot2dis function, dtype is none so the res would be all integer, if it is ok I change it to float cause all my target is only few pixel, and I think this would be more precisely

from boundary-loss.

Hi,

If you are using a sigmoid, then I am assuming that you are in a binary case. In that case the one-hot encoding of the label is redundant, but to keep changes to a minimum, for you case, you could go with:

pc = probs[:, 0, ...].type(torch.float32)
dc = dist_maps[:, 1, ...].type(torch.float32)

as self.idc would always have the value of [1] (it is mostly useful in multi-class).

Also, I want to know, in the onehot2dis function, dtype is none so the res would be all integer, if it is ok I change it to float cause all my target is only few pixel, and I think this would be more precisely

Yes that would work. I think I added the dtype option at the time, so that the code could handle distances maps with varying resolution across axises.

But to be useful, you would need to pass a resolution argument, otherwise you would only fill a float tensor with integers.

Cheers,

Hoel

from boundary-loss.