search

andreas128 / SRFlow

Official SRFlow training code: Super-Resolution using Normalizing Flow in PyTorch
Other
824 stars 112 forks source link

why only take $z_L$ into consideration instead of $z=(z_l)_l=1^L$ #33

Closed wyf0912 closed 2 years ago

wyf0912 commented 3 years ago

Thanks for sharing the code.

I have a confusion that you only maximize the term $p(z_L|x)$ in the code. However, the image is obtained by $z=(z_l)_l=1^L$.

What is the reason? Looking forward to your reply. Thanks!

KwanghyunOn commented 2 years ago

The log probability of the intermediate latent vectors (z[L-1] to z[1]) are added in the split layer. You can find the code at https://github.com/andreas128/SRFlow/blob/master/code/models/modules/Split.py#L56.

wyf0912 commented 2 years ago

Thanks a lot!