The Problem of Loss Function

Algolzw / image-restoration-sde

Image Restoration with Mean-Reverting Stochastic Differential Equations, ICML 2023. Winning solution of the NTIRE 2023 Image Shadow Removal Challenge.

https://algolzw.github.io/ir-sde/index.html

MIT License

583 stars 43 forks source link

The Problem of Loss Function #42

Open wangweiran0922 opened 1 year ago

wangweiran0922 commented 1 year ago

Thank you for opening up your work！

I have some questions about the loss function .ddpm’s label should be a prior information of gradually increasing noise. How did you modify it to calculate the loss between the enhanced image and the label image, which is reflected in the code? This place has always confused me，Looking forward to your reply，thank you！

Algolzw commented 1 year ago

Hi, the loss function is based on the Maximum Likelihood Learning from $xt$ to $x{t-1}$. You can find this part in our paper Section 3.3 and the code in optimize_parameters.