Backbone networks for GAN evaluation

[davidHemf]

Hi & Good day!

In the section "4.3 Backbone networks for GAN evaluation" of your article, you have mentioned that you have used DINO as a metric. Does that mean you have used it (only) as a features extractor-- instead of extracting from the InceptionV3 network? Or you have performed some other analysis?

And I have some question on the studio itself, may I ask them?

For FID, or iFID, is there any flag that allow to extract features of the layers other than pool3 of InceptionV3?
How about DINO? how to pick a GAN and (simultaneously?) feed some real and fake images and store extracted features?

Thanks again for the very engaging repo!

[alex4727]

Hi!

1. Yes, we used DINO (only) as feature extractor instead of InceptionV3 network along with SwAV, Swin-T.
2. For now, we do not provide a flag that allows extracting features from layers other than pool3 of InceptionV3. It's same for other backbones as well. We do not provide feature extracting locations as flag. You'll need to change code a little bit to get from different layers.
3. Unfortunately, we do not provide functions to feed real/fake images and store their extracted features.
   Thanks for your attention on our repo!