Spaces:
Sleeping
Sleeping
File size: 2,892 Bytes
b762e56 86e64e9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 |
import torch
import torchvision
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
class VGG19Feats(torch.nn.Module):
def __init__(self, requires_grad=False):
super(VGG19Feats, self).__init__()
vgg = torchvision.models.vgg19(pretrained=True).to(device) #.cuda()
# vgg.eval()
vgg_pretrained_features = vgg.features.eval()
self.requires_grad = requires_grad
self.slice1 = torch.nn.Sequential()
self.slice2 = torch.nn.Sequential()
self.slice3 = torch.nn.Sequential()
self.slice4 = torch.nn.Sequential()
self.slice5 = torch.nn.Sequential()
for x in range(3):
self.slice1.add_module(str(x), vgg_pretrained_features[x])
for x in range(3, 8):
self.slice2.add_module(str(x), vgg_pretrained_features[x])
for x in range(8, 13):
self.slice3.add_module(str(x), vgg_pretrained_features[x])
for x in range(13, 22):
self.slice4.add_module(str(x), vgg_pretrained_features[x])
for x in range(22, 31):
self.slice5.add_module(str(x), vgg_pretrained_features[x])
if not self.requires_grad:
for param in self.parameters():
param.requires_grad = False
def forward(self, img):
conv1_2 = self.slice1(img)
conv2_2 = self.slice2(conv1_2)
conv3_2 = self.slice3(conv2_2)
conv4_2 = self.slice4(conv3_2)
conv5_2 = self.slice5(conv4_2)
out = [conv1_2, conv2_2, conv3_2, conv4_2, conv5_2]
return out
class VGGPerceptualLoss(torch.nn.Module):
def __init__(self):
super(VGGPerceptualLoss, self).__init__()
self.vgg = VGG19Feats().to(device)
self.criterion = torch.nn.functional.l1_loss
self.register_buffer("mean", torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
self.register_buffer("std", torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
self.weights = [1.0/2.6, 1.0/4.8, 1.0/3.7, 1.0/5.6, 1.0*10/1.5]
def forward(self, input_img, target_img):
if input_img.shape[1] != 3:
input_img = input_img.repeat(1, 3, 1, 1)
target_img = target_img.repeat(1, 3, 1, 1)
input_img = (input_img - self.mean) / self.std
target_img = (target_img - self.mean) / self.std
x_vgg, y_vgg = self.vgg(input_img), self.vgg(target_img)
loss = {}
loss['pt_c_loss'] = self.weights[0] * self.criterion(x_vgg[0], y_vgg[0])+\
self.weights[1] * self.criterion(x_vgg[1], y_vgg[1])+\
self.weights[2] * self.criterion(x_vgg[2], y_vgg[2])+\
self.weights[3] * self.criterion(x_vgg[3], y_vgg[3])+\
self.weights[4] * self.criterion(x_vgg[4], y_vgg[4])
loss['pt_s_loss'] = 0.0
return loss |