Open fiskrt opened 1 year ago
Hi, is there a reason why the dice smoothing is only applied in the denominator:
https://github.com/aim-uofa/AdelaiDet/blob/f60e856dcf3bc8ac5c023e14caa7551c93df3c63/adet/modeling/condinst/dynamic_mask_head.py#L57-L58
and not like loss = 1. - ((2 * intersection + eps) / (union + eps))
loss = 1. - ((2 * intersection + eps) / (union + eps))
This causes all zero-input predictions to always maximize the loss even if it's completely correct.
Hi, is there a reason why the dice smoothing is only applied in the denominator:
https://github.com/aim-uofa/AdelaiDet/blob/f60e856dcf3bc8ac5c023e14caa7551c93df3c63/adet/modeling/condinst/dynamic_mask_head.py#L57-L58
and not like
loss = 1. - ((2 * intersection + eps) / (union + eps))
This causes all zero-input predictions to always maximize the loss even if it's completely correct.