Why training only 1 class at a time ?

[joihn]

Hello, thanks for this repo,
about this line:

mask=torch.where(mask==self.subclasses[self.tmp_index-1], self.tmp_index, 0)

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, 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 ?

[joihn]

Hum, maybe it's for minimizing intra-mask class imbalance influence ?

[bhosalems]

that is how a method is set I think, if we see the generation pipeline, it predicts noise in a training iteration for every class before adding them all together.