CompVis/stable-diffusion

Confused with nll_loss

Opened this issue · 1 comments

haichao2023 commented

Is this correct? Sum all pixels and divide by batch_size?

stable-diffusion/ldm/modules/losses/contperceptual.py

Line 57 in 21f890f

weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0]

onlyonewater commented

+1, I am also confused about this loss