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

Loss function #13

Closed minkyujeon closed 5 years ago

minkyujeon commented 5 years ago

In loss.py, class LossD(nn.Module): def init(self, device): super(LossD, self).init() self.dtype = torch.FloatTensor if device == 'cpu' else torch.cuda.FloatTensor self.zero = torch.zeros([1]).type(self.dtype)

def forward(self, r_x, r_x_hat):
    #r_x : D(x_t,y_t,i)'s result
    #r_x_hat : D(x_hat, y_t, i)'s result
    return (1 + r_x_hat) + (1 - r_x)

-> In paper, LossD is hingeLoss (L_DSC) , but I think it's not hinge loss.

Did I miss something?

subrtade662 commented 5 years ago

I agree, I've tried to use MarginRankingLoss but the discriminator is still too weak.

castelo-software commented 5 years ago

You're correct, I am not using hinge loss. That's simply because I never got around to fixing this. PyTorch doesn't provide a hinge loss function besides torch.nn.HingeEmbeddingLoss, which doesn't fit this use case, and I had some trouble using the max function for tensors. In order to get the script running I simply removed the max. But this is not the way the loss function is supposed to look!

minkyujeon commented 5 years ago

I used Hingeloss, and verified the loss whether it is appropriate in my task. If you dont mind, Could you leave the comment for my task. It is confusing to me if I use D loss value, and whether it is changed with negative value or not.

castelo-software commented 5 years ago

Yeah it's fine. I saw your pull request and I will merge it.