How to deal with multiple 1 value in the ground truth mask when you use Sotfmax cross entropy?
Superlee506 opened this issue · 1 comments
Superlee506 commented
Because the mask is only 1 point set to 1 but after crop, resize, round it may have more than 1 point set to 1 as in L526-L533. The softmax loss can handle the only one value with 1.
Superlee506 commented
use tf. ceil()