-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathmodel.py
138 lines (116 loc) · 5.65 KB
/
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
import numpy as np
import tensorflow as tf
import tensorflow.contrib.slim as slim
class Generator:
def __init__(self, learning_rate=1e-4, num_blocks=6):
self.learning_rate = learning_rate
self.num_blocks = num_blocks
def pelu(self, x):
with tf.variable_scope(x.op.name + '_activation', initializer=tf.constant_initializer(1.0), reuse=tf.AUTO_REUSE):
shape = x.get_shape().as_list()[1:]
alpha = tf.get_variable('alpha', 1, constraint=lambda t: tf.maximum(t, 0.1))
beta = tf.get_variable('beta', 1, constraint=lambda t: tf.maximum(t, 0.1))
positive = tf.nn.relu(x) * alpha / (beta + 1e-9)
negative = alpha * (tf.exp((-tf.nn.relu(-x)) / (beta + 1e-9)) - 1)
return negative + positive
def adaptive_global_average_pool_2d(self, x):
c = x.get_shape()[-1]
ADAP2d = tf.reshape(tf.reduce_mean(x, axis=[1, 2]), (-1, 1, 1, c))
return ADAP2d
def channel_attention(self, x, f, reduction):
skip_conn = tf.identity(x, name='identity')
x = self.adaptive_global_average_pool_2d(x)
x = tf.layers.conv2d(x, kernel_size=1, filters=f//reduction, strides=1, padding='same')
x = self.pelu(x)
x = tf.layers.conv2d(x, kernel_size=1, filters=f, strides=1, padding='same')
x = tf.nn.sigmoid(x)
CA = tf.multiply(skip_conn, x)
return CA
def ResidualBlock(self, x, kernel_size, filters, strides=1):
x = tf.layers.conv2d(x, kernel_size=1, filters=filters, strides=1, padding='same')
skip = x
x1 = x
for i in range(3):
tm1 = slim.conv2d(x1, num_outputs=filters, kernel_size=[3, 3], stride=1)
tm1 = self.pelu(tm1)
tm1 = slim.conv2d(tm1, num_outputs=filters, kernel_size=[1, 1], stride=1)
tm1 = self.pelu(tm1)
tm1 = slim.conv2d(tm1, num_outputs=filters, kernel_size=[1, 1], stride=1)
tm1 = self.channel_attention(tm1, f=filters, reduction=4)
x1 = tf.concat([x1,tm1], axis=3)
x2 = x
for i in range(3):
tm2 = slim.conv2d(x2, num_outputs=filters, kernel_size=[3, 3], stride=1)
tm2 = self.pelu(tm2)
tm2 = slim.conv2d(tm2, num_outputs=filters, kernel_size=[1, 1], stride=1)
tm2 = self.pelu(tm2)
tm2 = slim.conv2d(tm2, num_outputs=filters, kernel_size=[1, 1], stride=1)
tm2 = self.channel_attention(tm2, f=filters, reduction=4)
x2 = tf.concat([x2,tm2], axis=3)
x3 = x
for i in range(3):
tm3 = slim.conv2d(x3, num_outputs=filters, kernel_size=[3, 3], stride=1)
tm3 = self.pelu(tm3)
tm3 = slim.conv2d(tm3, num_outputs=filters, kernel_size=[1, 1], stride=1)
tm3 = self.pelu(tm3)
tm3 = slim.conv2d(tm3, num_outputs=filters, kernel_size=[1, 1], stride=1)
tm3 = self.channel_attention(tm3, f=filters, reduction=4)
x3 = tf.concat([x3,tm3], axis=3)
x5 = tf.concat(values=[x1, x2, x3], axis=3, name='stack0')
x6 = tf.layers.conv2d(x5, kernel_size=1, filters=filters, strides=strides, padding='same', use_bias=False)
x7 = skip + x6
return x7
def Upsample2xBlock(self, x, kernel_size, filters, strides):
#size = tf.shape(x)
#h = size[1]
#w = size[2]
#x = tf.image.resize_nearest_neighbor(x, size=[h * 3, w * 3], align_corners=False, name=None)
x = tf.layers.conv2d(x, kernel_size=kernel_size, filters=filters, strides=strides, padding='same')
x = tf.depth_to_space(x, 2)
x = self.pelu(x)
return x
def ThermalSR(self, x, reuse=False, isTraining=True):
with tf.variable_scope("ThermalSR", reuse=reuse) as scope:
x4 = tf.layers.conv2d(x, kernel_size=7, filters=64, strides=1, padding='same')
x4 = self.pelu(x4)
skip = x4
# Global Residual Learning
size = tf.shape(x)
h = size[1]
w = size[2]
x_GRL = tf.image.resize_bicubic(x, size=[h * 4, w * 4], align_corners=False, name=None)
x_GRL = tf.layers.conv2d(x_GRL, kernel_size=1, filters=64, strides=1, padding='same')
x_GRL = self.pelu(x_GRL)
x_GRL = tf.layers.conv2d(x_GRL, kernel_size=1, filters=16, strides=1, padding='same')
x_GRL = self.pelu(x_GRL)
x_GRL = tf.layers.conv2d(x_GRL, kernel_size=1, filters=3, strides=1, padding='same')
x_GRL = self.pelu(x_GRL)
for i in range(4):
x4 = self.ResidualBlock(x4, kernel_size=1, filters=64, strides=1)
x4 = tf.layers.conv2d(x4, kernel_size=1, filters=64, strides=1, padding='same', use_bias=False)
x4 = self.pelu(x4)
x4 = tf.concat([x4, skip], axis=3)
x4 = tf.layers.conv2d(x4, kernel_size=3, filters=64, strides=1, padding='same', use_bias=False)
x4 = self.pelu(x4)
x4 = x4 + skip
with tf.variable_scope('Upsamplingconv_stage_1'):
xUP = self.Upsample2xBlock(x4, kernel_size=3, filters=64, strides = 1)
xUP = tf.layers.conv2d(xUP, kernel_size=1, filters=64, strides=1, padding='same', use_bias=False)
xUP = self.pelu(xUP)
skip1 = xUP
for i in range(2):
x5 = self.ResidualBlock(xUP, kernel_size=1, filters=32, strides=1)
x5 = tf.layers.conv2d(x5, kernel_size=1, filters=32, strides=1, padding='same', use_bias=False)
x5 = self.pelu(x5)
x5 = tf.concat([x5, skip1], axis=3)
x5 = tf.layers.conv2d(x5, kernel_size=3, filters=64, strides=1, padding='same', use_bias=False)
x5 = self.pelu(x5)
x5 = x5 + skip1
with tf.variable_scope('Upsamplingconv_stage_2'):
x6 = self.Upsample2xBlock(x5, kernel_size=3, filters=64, strides = 1)
x6 = tf.layers.conv2d(x6, kernel_size=3, filters=64, strides=1, padding='same', name='forward_4')
x6 = self.pelu(x6)
x6 = tf.layers.conv2d(x6, kernel_size=3, filters=3, strides=1, padding='same', name='forward_5')
x6 = self.pelu(x6)
x_final = x6 + x_GRL
return x_final