def compute_vgg_loss(self, vgg, img, target):
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
img_fea = vgg(img_vgg)
target_fea = vgg(target_vgg)
return torch.mean(
(self.instancenorm(img_fea) - self.instancenorm(target_fea))**2)
def compute_vgg_loss(self, vgg, img, target, hyperparameters):
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
img_fea = vgg(img_vgg)
target_fea = vgg(target_vgg)
total_loss = 0
for i in range(len(img_fea)):
total_loss += hyperparameters['feature_weights'][i] * torch.mean(
torch.abs(img_fea[i] - target_fea[i]))
return total_loss
def compute_vgg_loss(self, vgg, img, target, all=0):
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
# img_fea = vgg(img_vgg)
# target_fea = vgg(target_vgg)
img_fea_dict = self.VggExtract(img_vgg)
target_fea_dict = self.VggExtract(target_vgg)
loss=0
if all:
# for feature in img_fea_dict:
# loss+= torch.mean((img_fea_dict[feature] - (target_fea_dict[feature])) ** 2)
loss += torch.mean((img_fea_dict['relu4_3'] - (target_fea_dict['relu4_3'])) ** 2)
else:
loss += torch.mean(
(self.instancenorm(img_fea_dict['relu4_3']) - self.instancenorm(target_fea_dict['relu4_3'])) ** 2)
return loss
def compute_vgg_loss(self, vgg, img, target):
"""
Compute the domain-invariant perceptual loss
Arguments:
vgg {model} -- popular Convolutional Network for Classification and Detection
img {torch.Tensor} -- image before translation
target {torch.Tensor} -- image after translation
Returns:
torch.Float -- domain invariant perceptual loss
"""
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
img_fea = vgg(img_vgg)
target_fea = vgg(target_vgg)
return torch.mean(
(self.instancenorm(img_fea) - self.instancenorm(target_fea))**2)
def get_image(img_rows, img_cols, image_file_path):
"""
TODO Doctring.
"""
# DO NOT CHANGE THE ORDER OF THE NEXT 4 LINES - JUST. DO. NOT.
img = image.load_img(image_file_path, target_size=(img_rows, img_cols))
img = image.img_to_array(img)
img = utils.vgg_preprocess(img)
img = np.expand_dims(img, axis=0)
return img
def compute_vgg_loss(self, img, target, mask):
"""
Compute the domain-invariant perceptual loss
Arguments:
vgg {model} -- popular Convolutional Network for Classification and Detection
img {torch.Tensor} -- image before translation
target {torch.Tensor} -- image after translation
Returns:
torch.Float -- domain invariant perceptual loss
"""
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
# Mask input to VGG:
img_vgg = img_vgg * (1.0 - mask)
target_vgg = target_vgg * (1.0 - mask)
loss_G_VGG = self.criterionVGG(img_vgg, target_vgg)
return loss_G_VGG
def compute_vgg_loss(self, vgg, img, target):
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
img_fea = vgg(img_vgg)
target_fea = vgg(target_vgg)
return torch.mean(torch.abs(img_fea - target_fea))
def compute_vgg_loss(self, vgg, img, target):
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
img_fea = vgg(img_vgg)
target_fea = vgg(target_vgg)
return torch.mean((self.instancenorm(img_fea) - self.instancenorm(target_fea)) ** 2)
def compute_vgg_loss_new(self, vgg, img, target):
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
img_feat = vgg(img_vgg)
target_feat = vgg(target_vgg)
return self.recon_criterion(img_feat, target_feat)
def compute_vgg_loss(self, vgg, img, target):
img_vgg = vgg_preprocess(img)
target_vgg = vgg_preprocess(target)
img_fea = vgg.features(img_vgg)
target_fea = vgg.features(target_vgg)
return contextual_loss(img_fea, target_fea)