When using single-channel masks or atlases, it's value was being applied onto the predictions as int (thus multiplying) rather than boolean values (thus masking).
This PR forces the mask into a boolean by setting it to multichannel_mask[..., 0:1] > 0, fixing the issue.
When using single-channel masks or atlases, it's value was being applied onto the predictions as int (thus multiplying) rather than boolean values (thus masking).
This PR forces the mask into a boolean by setting it to
multichannel_mask[..., 0:1] > 0
, fixing the issue.