About the perceptual loss

[mazzzystar]

In your paper, the perceptual loss was:

# vggsubmod refers to certain layer of VGG.
Lper = L1(gram(vggsubmod(x)), gram(vggsubmod(y))  

However in your implementation:

ADGAN/losses/L1_plus_perceptualLoss.py

Lines 63 to 72 in d948cb1

	fake_p2_norm = self.vgg_submodel(fake_p2_norm)
	input_p2_norm = self.vgg_submodel(input_p2_norm)
	input_p2_norm_no_grad = input_p2_norm.detach()
	if self.percep_is_l1 == 1:
	# use l1 for perceptual loss
	loss_perceptual = F.l1_loss(fake_p2_norm, input_p2_norm_no_grad) * self.lambda_perceptual
	else:
	# use l2 for perceptual loss
	loss_perceptual = F.mse_loss(fake_p2_norm, input_p2_norm_no_grad) * self.lambda_perceptual

Could you give some explanation on that ? thanks.

[menyifang]

In your paper, the perceptual loss was:

# vggsubmod refers to certain layer of VGG.
Lper = L1(gram(vggsubmod(x)), gram(vggsubmod(y))  

However in your implementation:

ADGAN/losses/L1_plus_perceptualLoss.py

Lines 63 to 72 in d948cb1

	fake_p2_norm = self.vgg_submodel(fake_p2_norm)
	input_p2_norm = self.vgg_submodel(input_p2_norm)
	input_p2_norm_no_grad = input_p2_norm.detach()
	if self.percep_is_l1 == 1:
	# use l1 for perceptual loss
	loss_perceptual = F.l1_loss(fake_p2_norm, input_p2_norm_no_grad) * self.lambda_perceptual
	else:
	# use l2 for perceptual loss
	loss_perceptual = F.mse_loss(fake_p2_norm, input_p2_norm_no_grad) * self.lambda_perceptual

Could you give some explanation on that ? thanks.

Hi @mazzzystar, thank you for your correction! Here we directly use the implement of perceptual loss from PATN and mistake it as common style loss in perceptual loss. We will correct this issue in our arxiv version.