-
Notifications
You must be signed in to change notification settings - Fork 75
/
fr_model.py
143 lines (111 loc) · 4.05 KB
/
fr_model.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
import numpy as np
import chainer
from chainer import Variable
import chainer.functions as F
import chainer.links as L
from chainer import computational_graph
from chainer import cuda
from chainer import optimizers
from chainer import serializers
class FRModel(chainer.Chain):
def __init__(self, top="patchwise"):
super(FRModel, self).__init__(
# feature extraction
conv1 = L.Convolution2D(3, 32, 3, pad=1),
conv2 = L.Convolution2D(32, 32, 3, pad=1),
conv3 = L.Convolution2D(32, 64, 3, pad=1),
conv4 = L.Convolution2D(64, 64, 3, pad=1),
conv5 = L.Convolution2D(64, 128, 3, pad=1),
conv6 = L.Convolution2D(128, 128, 3, pad=1),
conv7 = L.Convolution2D(128, 256, 3, pad=1),
conv8 = L.Convolution2D(256, 256, 3, pad=1),
conv9 = L.Convolution2D(256, 512, 3, pad=1),
conv10 = L.Convolution2D(512, 512, 3, pad=1),
# quality regression
fc1 = L.Linear(512 * 3, 512),
fc2 = L.Linear(512, 1)
)
self.top = top
if top == "weighted":
fc1_a = L.Linear(512 * 3, 512)
fc2_a = L.Linear(512, 1)
self.add_link("fc1_a", fc1_a)
self.add_link("fc2_a", fc2_a)
def extract_features(self, x, train=True):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
self.h1 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv3(h))
h = F.relu(self.conv4(h))
self.h2 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv5(h))
h = F.relu(self.conv6(h))
self.h3 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv7(h))
h = F.relu(self.conv8(h))
self.h4 = h
h = F.max_pooling_2d(h,2)
h = F.relu(self.conv9(h))
h = F.relu(self.conv10(h))
self.h5 = h
h = F.max_pooling_2d(h,2)
return h
def forward(self, x_data, x_ref_data, y_data, train=True,
n_patches_per_image=32):
xp = cuda.cupy
if not isinstance(x_data, Variable):
x = Variable(x_data)
else:
x = x_data
x_data = x.data
self.n_images = y_data.shape[0]
self.n_patches = x_data.shape[0]
self.n_patches_per_image = n_patches_per_image
x_ref = Variable(x_ref_data)
h = self.extract_features(x)
self.h = h
h_ref = self.extract_features(x_ref)
h = F.concat((h-h_ref, h, h_ref))
h_ = h # save intermediate features
h = F.dropout(F.relu(self.fc1(h)), ratio=0.5)
h = self.fc2(h)
if self.top == "weighted":
a = F.dropout(F.relu(self.fc1_a(h_)), ratio=0.5)
a = F.relu(self.fc2_a(a))+0.000001
t = Variable(y_data)
self.weighted_loss(h, a, t)
elif self.top == "patchwise":
a = Variable(xp.ones_like(h.data))
t = Variable(xp.repeat(y_data, n_patches_per_image))
self.patchwise_loss(h, a, t)
if train:
return self.loss
else:
return self.loss, self.y
def patchwise_loss(self, h, a, t):
self.loss = F.sum(abs(h - F.reshape(t, (-1,1))))
self.loss /= self.n_patches
if self.n_images > 1:
h = F.split_axis(h, self.n_images, 0)
a = F.split_axis(a, self.n_images, 0)
else:
h, a = [h], [a]
self.y = h
self.a = a
def weighted_loss(self, h, a, t):
self.loss = 0
if self.n_images > 1:
h = F.split_axis(h, self.n_images, 0)
a = F.split_axis(a, self.n_images, 0)
t = F.split_axis(t, self.n_images, 0)
else:
h, a, t = [h], [a], [t]
for i in range(self.n_images):
y = F.sum(h[i]*a[i], 0) / F.sum(a[i], 0)
self.loss += abs(y - F.reshape(t[i], (1,)))
self.loss /= self.n_images
self.y = h
self.a = a