If I understand correctly:
The guidance mechanism used here trains only 1 class at a time
In other word, a mask can contains only zeros (unkown class) and 1 other class (specifically selected for this sample)
If there are other class in this mask, they are sent to 0
What is the reason behind this ?
Why can't a mask contains multiple class during training ?
for example class idx_1 on the righthalf, class idx_2 on the left half ?
Making proper use of all the label pixel available at each iteration would significantly speed up training no ?
Hello, thanks for this repo, about this line:
https://github.com/marcoaversa/diffinfinite/blob/master/dataset.py#L353
If I understand correctly: The guidance mechanism used here trains only 1 class at a time In other word, a mask can contains only zeros (unkown class) and 1 other class (specifically selected for this sample) If there are other class in this mask, they are sent to 0
What is the reason behind this ? Why can't a mask contains multiple class during training ? for example class
idx_1
on the righthalf, classidx_2
on the left half ?Making proper use of all the label pixel available at each iteration would significantly speed up training no ?