VGGFace loss details.

[funnylittleman]

For perceptual loss do you use features obtained by VGGFeat class(same features as in main model) or some other model? In paper it is written that you use first 4 conv layers of VGGFace for perceptual loss. Did you mean last conv layers of first 4 blocks of VGG (as it is usually done in perceptual losses) or you mean literally first 4 conv blocks?