Spaces:
Sleeping
Sleeping
Upload 3 files
Browse files- dataTransform.py +29 -0
- styleTransfer.py +75 -0
- vggModel.py +24 -0
dataTransform.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from PIL import Image
|
2 |
+
import torchvision.transforms as transforms #to transform the images
|
3 |
+
|
4 |
+
|
5 |
+
def load_image(image_path, device):
|
6 |
+
|
7 |
+
image_size = 356
|
8 |
+
|
9 |
+
loader = transforms.Compose(
|
10 |
+
[
|
11 |
+
transforms.Resize((image_size, image_size)), #RESIZE IMAGE
|
12 |
+
transforms.ToTensor() #TRANSFORM IMAGE TO TENSOR
|
13 |
+
]
|
14 |
+
)
|
15 |
+
|
16 |
+
image = Image.open(image_path)
|
17 |
+
image = loader(image).unsqueeze(0) #(h, c, w) -> (1, h, c, w) adds batch dim
|
18 |
+
|
19 |
+
return image.to(device)
|
20 |
+
|
21 |
+
|
22 |
+
def tensor_to_image(tensor):
|
23 |
+
tensor = tensor.clone().detach() # Ensure the tensor is detached from the graph
|
24 |
+
tensor = tensor.squeeze(0) # Remove batch dimension if present
|
25 |
+
tensor = torch.clamp(tensor, 0, 1) # Clamp the values to [0, 1] range
|
26 |
+
|
27 |
+
unloader = transforms.ToPILImage()
|
28 |
+
image = unloader(tensor.cpu())
|
29 |
+
return image
|
styleTransfer.py
ADDED
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch # for model
|
2 |
+
import numpy as np
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch.optim as optim
|
5 |
+
from PIL import Image #for importing images
|
6 |
+
import torchvision.models as models #to load vgg 19 model
|
7 |
+
import torchvision.transforms as transforms
|
8 |
+
from tqdm import tqdm
|
9 |
+
|
10 |
+
from dataTransform import load_image
|
11 |
+
from vggModel import VGGNet
|
12 |
+
|
13 |
+
def style_transfer(content_img, style_img, total_steps, alpha=1e5, beta=1e10, learning_rate=0.001):
|
14 |
+
# Preprocess the input images
|
15 |
+
|
16 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
17 |
+
print('-'*30)
|
18 |
+
print(f'Device Initialized: {device}')
|
19 |
+
print('-'*30)
|
20 |
+
content_img = load_image(content_img, device)
|
21 |
+
style_img = load_image(style_img, device)
|
22 |
+
generated_img = content_img.clone().requires_grad_(True)
|
23 |
+
optimizer = optim.Adam([generated_img], lr = learning_rate)
|
24 |
+
model = VGGNet().to(device).eval()
|
25 |
+
|
26 |
+
# print(content_img.shape)
|
27 |
+
# print(style_img.shape)
|
28 |
+
# print(generated_img.shape)
|
29 |
+
|
30 |
+
|
31 |
+
for step in tqdm(range(total_steps)):
|
32 |
+
|
33 |
+
#first we send the 3 images from the vgg network
|
34 |
+
|
35 |
+
generated_feats = model(generated_img)
|
36 |
+
original_image_feats = model(content_img)
|
37 |
+
style_feats = model(style_img)
|
38 |
+
|
39 |
+
#defining the style loss
|
40 |
+
|
41 |
+
style_loss = original_loss = 0
|
42 |
+
|
43 |
+
|
44 |
+
for gen_feat, orig_image_feat, styl_feat in zip(generated_feats, original_image_feats, style_feats): #looping over each feature
|
45 |
+
|
46 |
+
# print(gen_feat.shape)
|
47 |
+
# print(orig_image_feat.shape)
|
48 |
+
# print(styl_feat.shape)
|
49 |
+
|
50 |
+
batch, channel, height, width = gen_feat.shape
|
51 |
+
original_loss += torch.mean((gen_feat - orig_image_feat)**2)
|
52 |
+
|
53 |
+
# computing gram matrix for gen and style to compute style loss
|
54 |
+
|
55 |
+
G = gen_feat.view(channel, height*width).mm(
|
56 |
+
gen_feat.view(channel, height*width).t()
|
57 |
+
)
|
58 |
+
|
59 |
+
# correlation matrix
|
60 |
+
|
61 |
+
A = styl_feat.view(channel, height*width).mm(
|
62 |
+
styl_feat.view(channel, height*width).t()
|
63 |
+
)
|
64 |
+
|
65 |
+
style_loss += torch.mean((G-A)**2)
|
66 |
+
|
67 |
+
total_loss = alpha*original_loss + beta*style_loss
|
68 |
+
|
69 |
+
optimizer.zero_grad()
|
70 |
+
total_loss.backward()
|
71 |
+
optimizer.step()
|
72 |
+
|
73 |
+
if step == total_steps - 1:
|
74 |
+
# Postprocess and return the final generated image
|
75 |
+
return generated_img
|
vggModel.py
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch # for model
|
2 |
+
import torch.nn as nn
|
3 |
+
import torchvision.models as models #to load vgg 19 model
|
4 |
+
|
5 |
+
|
6 |
+
class VGGNet(nn.Module):
|
7 |
+
|
8 |
+
def __init__(self):
|
9 |
+
|
10 |
+
super(VGGNet, self).__init__()
|
11 |
+
self.chosen_features = ['0', '5', '10', '19', '28']
|
12 |
+
self.vgg = models.vgg19(pretrained = True).features #select only certain layers to extract fetaures
|
13 |
+
|
14 |
+
|
15 |
+
def forward(self,x):
|
16 |
+
features = [] #returns features from selected conv layers from VGG19 pretrained model
|
17 |
+
|
18 |
+
for layer_num, layer in self.vgg._modules.items():
|
19 |
+
x = layer(x)
|
20 |
+
|
21 |
+
if layer_num in self.chosen_features:
|
22 |
+
features.append(x)
|
23 |
+
|
24 |
+
return features
|