forked from oreilly-japan/deep-learning-from-scratch-3
-
Notifications
You must be signed in to change notification settings - Fork 0
/
style_transfer.py
147 lines (121 loc) · 4.62 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
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import dezero
import dezero.functions as F
from dezero import Variable
from dezero.models import VGG16
use_gpu = dezero.cuda.gpu_enable
lr = 5.0
iterations = 2001
model_input_size = (224, 224)
style_weight = 1.0
content_weight = 1e-4
total_varitaion_weight = 1e-6
content_layers = ['conv5_2']
style_layers = ['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1']
content_url = 'https://github.com/oreilly-japan/deep-learning-from-scratch-3/raw/images/zebra.jpg'
style_url = 'https://raw.githubusercontent.com/jcjohnson/neural-style/master/examples/inputs/starry_night_google.jpg'
class VGG16(VGG16):
def extract(self, x):
c1_1 = F.relu(self.conv1_1(x))
c1_2 = F.relu(self.conv1_2(c1_1))
p1 = F.average_pooling(c1_2, 2, 2)
c2_1 = F.relu(self.conv2_1(p1))
c2_2 = F.relu(self.conv2_2(c2_1))
p2 = F.average_pooling(c2_2, 2, 2)
c3_1 = F.relu(self.conv3_1(p2))
c3_2 = F.relu(self.conv3_2(c3_1))
c3_3 = F.relu(self.conv3_3(c3_2))
p3 = F.average_pooling(c3_3, 2, 2)
c4_1 = F.relu(self.conv4_1(p3))
c4_2 = F.relu(self.conv4_2(c4_1))
c4_3 = F.relu(self.conv4_3(c4_2))
p4 = F.average_pooling(c4_3, 2, 2)
c5_1 = F.relu(self.conv5_1(p4))
c5_2 = F.relu(self.conv5_2(c5_1))
c5_3 = F.relu(self.conv5_3(c5_2))
return {'conv1_1':c1_1, 'conv1_2':c1_2, 'conv2_1':c2_1, 'conv2_2':c2_2,
'conv3_1':c3_1, 'conv3_2':c3_2, 'conv3_3':c3_3, 'conv4_1':c4_1,
'conv5_1':c5_1, 'conv5_2':c5_2, 'conv5_3':c5_3}
# Setup for content & style image
content_path = dezero.utils.get_file(content_url)
style_path = dezero.utils.get_file(style_url)
content_img = Image.open(content_path)
content_size = content_img.size
style_img = Image.open(style_path)
content_img = VGG16.preprocess(content_img, size=model_input_size)[np.newaxis] # preprocess for VGG
style_img = VGG16.preprocess(style_img, size=model_input_size)[np.newaxis]
content_img, style_img = Variable(content_img), Variable(style_img)
model = VGG16(pretrained=True)
#gen_data = np.random.uniform(-20, 20, (1, 3, img_resize[0], img_resize[1])).astype(np.float32)
gen_data = content_img.data.copy()
gen_img = dezero.Parameter(gen_data)
gen_model = dezero.models.Model()
gen_model.param = gen_img
optimizer = dezero.optimizers.AdaGrad(lr=lr).setup(gen_model)
if use_gpu:
model.to_gpu()
gen_img.to_gpu()
content_img.to_gpu()
style_img.to_gpu()
with dezero.no_grad():
content_features = model.extract(content_img)
style_features = model.extract(style_img)
def deprocess_image(x, size=None):
if use_gpu:
x = dezero.cuda.as_numpy(x)
if x.ndim == 4:
x = np.squeeze(x)
x = x.transpose((1,2,0))
x += np.array([103.939, 116.779, 123.68])
x = x[:,:,::-1] # BGR -> RGB
x = np.clip(x, 0, 255).astype('uint8')
img = Image.fromarray(x, mode="RGB")
if size:
img = img.resize(size)
return img
def gram_mat(x):
N, C, H, W = x.shape
features = x.reshape(C, -1)
gram = F.matmul(features, features.T)
return gram.reshape(1, C, C)
def style_loss(style, comb):
S = gram_mat(style)
C = gram_mat(comb)
N, ch, H, W = style.shape
return F.mean_squared_error(S, C) / (4 * (ch * W * H)**2)
def content_loss(base, comb):
return F.mean_squared_error(base, comb) / 2
def total_varitaion_loss(x):
a = (x[:, :, :-1, :-1] - x[:, :, 1:, : -1]) ** 2
b = (x[:, :, :-1, :-1] - x[:, :, : -1, 1:]) ** 2
return F.sum(a + b)
def loss_func(gen_features, content_features, style_features, gen_img):
loss = 0
# content loss
for layer in content_features:
loss += content_weight / len(content_layers) * \
content_loss(content_features[layer], gen_features[layer])
# style loss
for layer in style_features:
loss += style_weight / len(style_layers) * \
style_loss(style_features[layer], gen_features[layer])
# total variation loss
loss += total_varitaion_weight * total_varitaion_loss(gen_img)
return loss
print_interval = 100 if use_gpu else 1
for i in range(iterations):
model.cleargrads()
gen_img.cleargrad()
gen_features = model.extract(gen_img)
loss = loss_func(gen_features, content_features, style_features, gen_img)
loss.backward()
optimizer.update()
if i % print_interval == 0:
print('{} loss: {:.0f}'.format(i, float(loss.data)))
if i % 100 == 0:
img = deprocess_image(gen_img.data, content_size)
plt.imshow(np.array(img))
plt.show()
#img.save("style_transfer_{}.png".format(str(i)))