search

Yang-Liu1082 / InvDN

Implementation for the paper: Invertible Denoising Network: A Light Solution for Real Noise Removal (CVPR2021).
Apache License 2.0
185 stars 35 forks source link

Sth about the Loss Function #9

Open ZhouYiiFeng opened 3 years ago

ZhouYiiFeng commented 3 years ago

Hi,

I notice that this model is based on the "flow-based generative model", however, the loss does not include the logdet loss. The cal_jacobian is default set to False. Is this cause by you replaced some features of the forward output by the normal distribution?

I‘m new in learning the "flow-based generative model", maybe I got it wrong, could you help me? Thx!

Yang-Liu1082 commented 2 years ago

It is just because there is no need to calculate the Jacobian. Calculating Jacobian is to get the accurate log-likelihood. However, in our task, we don't care about the log-likelihood of z. Setting it false can speed up the training.

ZhouYiiFeng commented 2 years ago

Thanks for your reply, I still have some other questions, would you like to add my wechat? My wechat ID is: wx_joeyf

Yang-Liu1082 commented 2 years ago

Sure.

From: JoeyF Zhou @.> Sent: Friday, 9 July 2021 4:57 PM To: Yang-Liu1082/InvDN @.> Cc: Yang Liu @.>; Comment @.> Subject: Re: [Yang-Liu1082/InvDN] Sth about the Loss Function (#9)

Thanks for your reply, I still have some other questions, would you like to add my wechat? My wechat ID is: wx_joeyf

— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/Yang-Liu1082/InvDN/issues/9#issuecomment-876961556, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ATMHHUDOSSDBJSZ35SOWJ33TW2MVLANCNFSM47T4TWKA.