Hi, I issue a bug for Using pre-trained networks.

[disanda]

Hi, I issue some bugs for Using pre-trained networks. could you help to deal with this?

(1)
ex: "ffhq-config-e-Gorig-Dskip"
there are like this below:

(2)
meanwhile, some pre-trained net can not be converted
ex:"ffhq-config-e-Gresnet-Dorig-Bug"

eccured runtime error:
"RuntimeError: The size of tensor a (256) must match the size of tensor b (512) at non-singleton dimension 1"

(3)
btw, In my view,

the pre-trained net: 'car-config-e-Gorig-Dorig' are the same net structures (both G & D) with 'car-config-e', right?

[adriansahlman]

Hey, thanks for the report! I will have to look into this in the future. I am afraid I do not have time to fix these issues right now as I am fully swamped by work 7 days per week :(

If "orig" stands for the original models used in the previous version of styleGAN it might be that this repo does not support them. Although I do not remember as it was almost a year ago I wrote the code. Ill take a look at it at a later time! Maybe try using models that dont have "orig" in the name in the meanwhile. They are not part of the top-performing models anyway.

According to the paper

The combination of generator with output skips and residual discriminator corresponds to configuration E in the main result table.

which means that car-config-e is the same as car-config-e-Gskip-Dresnet.

Hope this helps :)