search

genforce / mganprior

[CVPR 2020] Image Processing Using Multi-Code GAN Prior
https://genforce.github.io/mganprior/
288 stars 44 forks source link

0.001 * vgg loss #16

Closed zhang-lingyun closed 3 years ago

zhang-lingyun commented 3 years ago

why you did this in you vgg loss return self.loss(x_features, gt_features, reduction='mean') * 0.001? 0.001 shrink the loss a lot. what the popurse behind this? thanks .

JasonGUTU commented 3 years ago

Hi,

We did this because the original VGG network we used is transferred from Caffe model. The value of the input is [0, 255] and we need to give a small weight to it. If you use the VGG network from pytorch-pretraining VGG. This 0.001 could be something like 1. or 0.1.

zhang-lingyun commented 3 years ago

I see, thanks .