search

pkuxmq / Invertible-Image-Rescaling

[ECCV 2020, IJCV 2022] Invertible Image Rescaling
Apache License 2.0
630 stars 86 forks source link

About loss_ce? #22

Closed codyshen0000 closed 3 years ago

codyshen0000 commented 3 years ago

Hi , I'm confused why use ''torch.sum(z**2)'' to calculate loss_ce. Besides, how can it be transformed from calculating loss_distr on Y x Z to calculating Loss_ce only on Z space?

pkuxmq commented 3 years ago

Please see Eq.(9) in the paper. z follows Gaussian distribution, so its log probability is ||z||^2. CE loss does not necessarily guarantee the distribution of (Y, Z) and is only a weaker surrogate loss for stable training (refer to https://github.com/pkuxmq/Invertible-Image-Rescaling/issues/14).