search

grey-eye / talking-heads

Our implementation of "Few-Shot Adversarial Learning of Realistic Neural Talking Head Models" (Egor Zakharov et al.)
GNU General Public License v3.0
593 stars 110 forks source link

Questions about implementation detail #20

Closed BaldrLector closed 5 years ago

BaldrLector commented 5 years ago

Hi, is there pre-trained weight provided? I try to download database and run the code, but the result not so amazing like origin paper said. so is there any trick to fine-turning result?

you don't use hinge loss at Dsc, which seems to differ from origin paper.

def forward(self, r_x, r_x_hat): return (1 + r_x_hat) + (1 - r_x)

actually, it should be def forward(self, r_x, r_x_hat): return relu(1 + r_x_hat) + relu(1 - r_x)

Besides, not only the first frame could be used to compute loss, but other frames that used to computer e_mean also could be used to compute loss at training phase, the K input frame x_i be generative with e_mean and y_i, it's right?

BaldrLector commented 5 years ago

Besides, model-collapse happened when I train the net, are there tricks to avoid that? Thanks