Open B-1368 opened 2 years ago
This is an interesting question that I'm interested in. While I'm not the author, I believe this checkpoint was trained on a similar network related to pix2pixHD (or some other image to image translation). The part of the code you are referring to is heavily referenced from here, which is probably how this network was trained.
In the training process, we will only get net_G and net_D. However, when testing it seems another model is required. L57 of
test_one_image.py
islatend_id = model.netArc(img_id_downsample)
, butmodel.netArc
is loaded from arcface_checkpoint.tar you share. I cannot find any definition about the model architecture of that. So is it defined in this repo? If not, will you release the architecture? Thanks!