Thanks for open sourcing the awesome work. However I find that the nll_loss is averaged over batch size, which means that it is summed over all pixels and channels. I have tried training and found the loss is very large. Is it a typo or intended? Thanks.
Thanks for open sourcing the awesome work. However I find that the nll_loss is averaged over batch size, which means that it is summed over all pixels and channels. I have tried training and found the loss is very large. Is it a typo or intended? Thanks.
https://github.com/buxiangzhiren/Asymmetric_VQGAN/blob/406c5cd2e86b3e9565795cc5cd1e3aab304b1a44/ldm/modules/losses/contperceptual.py#L58C14-L58C14