-
Notifications
You must be signed in to change notification settings - Fork 0
/
style_transfer.py
205 lines (142 loc) · 6.42 KB
/
style_transfer.py
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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
# -*- coding: utf-8 -*-
import torch,torchvision
import numpy as np
from matplotlib import pyplot as plt
import os
from PIL import Image
from skimage import io
import copy
import tqdm
tensor_to_numpy = lambda t:t.detach().cpu().numpy()
device = 'cpu'
"""Define functions for converting the input to the CNN back into a visualizable image"""
mean,std = (0.485, 0.456, 0.406),(0.229, 0.224, 0.225)
def denormalize_tensor(t,
mean = mean,
std = std):
mean = torch.tensor(mean).float().unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
std = torch.tensor(std).float().unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
mean = mean.type_as(t).to(t.device)
std = std.type_as(t).to(t.device)
dnrm_t = t * std + mean
return dnrm_t
def tensor_to_im(t,
mean = mean,
std=std):
t = denormalize_tensor(t,
mean=mean,
std=std)
t_np = tensor_to_numpy(t)
im = np.transpose(t_np,(0,2,3,1))
return im
""" for Gram Matrix"""
def gram_matrix(t):
b,c,h,w = t.shape
f = t.view(b,c,-1)
g = torch.einsum('bxf,bzf->bxz',[f,f])
g = g*(1./np.prod(t.shape)) # to account for difference in number of channels and pixels at every layer of the CNN
return g
""" Forward Hook """
def fwdHook(self,input,output):
self.feat = output
pass
def style_transfer(model,
transforms,
content,
style,
nepochs = 2400,
style_lambda = 1000000
):
if model_name == 'vgg19':
layers_for_style = [0,
2,
5,
7,
10]
layers_for_content = [0]
layers_of_interest = layers_for_style[:]
layers_of_interest.extend(layers_for_content[:])
hooked_layers = []
model_children = list(model.children())
for li in layers_of_interest:
l = model_children[li]
hooked_layers.append(l.register_forward_hook(fwdHook))
"""Declare the style and content images, the optimizee and the optimizer"""
output = content.detach().clone() # the output starts from the content image
content,style,output = content.to(device),style.to(device),output.to(device) # push everything onto the device
output = output.requires_grad_(True) # requires_grad should be the last operation on the optimizee
# opt = torch.optim.LBFGS([output]) # similar to the tutorial, using a L-BFGS optimizer
opt = torch.optim.Adam([output],
lr = 1e-2)
"""Optimization loop and visualization"""
trends = {'total_loss':[],
'content_loss':[],
'style_loss':[],}
for e in tqdm.tqdm(range(nepochs)):
'''----- style -----'''
_ = model(style) # pass the style image through the CNN
style_feat = [model_children[li].feat for li in layers_for_style] # get the feats from the style layers
style_gram = [gram_matrix(f) for f in style_feat] # calculate the gram matrix over these layers
'''----- content -----'''
_ = model(content) # pass the content image through the CNN
content_feat = model_children[layers_for_content[0]].feat # there is only one content layer
'''----- output -----'''
_ = model(output) # pass the optimizee through the CNN
output_style_feat = [model_children[li].feat for li in layers_for_style] # get the optimizee style feats
output_style_gram = [gram_matrix(f) for f in output_style_feat] # get the optimizee gram matrices
# output_content_feat = [model_children[li].feat for li in layers_for_content]
output_content_feat = model_children[layers_for_content[0]].feat # get the optimizee content features
'''----- losses -----'''
content_loss = torch.nn.functional.mse_loss(content_feat.view(-1),output_content_feat.view(-1))
style_loss = sum([torch.nn.functional.mse_loss(st_gr.view(-1),op_gr.view(-1)) for st_gr,op_gr in zip(style_gram,output_style_gram)])
total_loss = content_loss + style_lambda * style_loss
'''----- update -----'''
opt.zero_grad()
total_loss.backward()
opt.step()
'''----- bookkeeping -----'''
trends['total_loss'].append(tensor_to_numpy(total_loss))
trends['content_loss'].append(tensor_to_numpy(content_loss))
trends['style_loss'].append(tensor_to_numpy(style_loss))
pass
"""Save the final result to disk"""
result = tensor_to_im(output)[0]
if 'save_to_disk':
io.imsave('style_transfered.png',result)
return result,trends
if __name__ == '__main__':
model_specs = {'alexnet':(227,227),
'vgg19':(224,224),
}
model_name = 'vgg19'
transforms = torchvision.transforms.Compose([torchvision.transforms.Resize(model_specs[model_name]),
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225)),
])
"""Model specifications and transform:
sizes of images, mean and variance
"""
"""See the images"""
im_scream,im_bridge_girl = io.imread('scream.jpeg'), io.imread('girl_on_bridge.jpg')
plt.figure()
plt.imshow(im_scream)
plt.title('Scream')
plt.figure()
plt.imshow(im_bridge_girl)
plt.title('Girl On Bridge')
"""make tensors from images, resizing and transforming them."""
im_scream_pil,im_bridge_girl_pil = Image.fromarray(im_scream),Image.fromarray(im_bridge_girl)
scream = transforms(im_scream_pil).unsqueeze(0)
bridge_girl = transforms(im_bridge_girl_pil).unsqueeze(0)
content,style = bridge_girl,scream
"""Get the model"""
model = torchvision.models.vgg19(pretrained=True).features.to(device).eval() # get the model in eval mode
results,trends = style_transfer(model,
transforms,
content,
style,
nepochs = 20,
style_lambda = 1000000
)
#"""Some setup code in case we want to make a time lapse of style transfer"""