Error when running the "finetuning_training.py" about realistic-neural-talking-head-models HOT 3 CLOSED

I found the problem, it is the self.VGG19(x) in loss_generator.py.

The input of the pretrain VGG19 should be in size 224 by 224, but the x is 256 by 256 instead, causing size mismatch problem in FC layer of VGG19.

from realistic-neural-talking-head-models.

Hi @imaxpayne!

I think the problem might not be with VGG19 as there's no error when I run this:

from torchvision.models import vgg19

VGG19 = vgg19(pretrained = True)
x = torch.randn(1,3,224,224)
VGG19(x)

But you're right about the last fc layers, I made some commits lately to loss generator to drop these layers in VGGFace after uploading finetuning_training.py. Could you pull the recent changes and tell me if the problem still persists?

As a sidenote, you may want to redo the download for the prerequisites, I also changed the source for the network as the previous one was wrong. This change only makes the network train more efficiently.

from realistic-neural-talking-head-models.

I found the problem, it is the self.VGG19(x) in loss_generator.py.

The input of the pretrain VGG19 should be in size 224 by 224, but the x is 256 by 256 instead, causing size mismatch problem in FC layer of VGG19.

i also have the same question. how to deal with the problem?thank you

from realistic-neural-talking-head-models.