-
Notifications
You must be signed in to change notification settings - Fork 20
/
model.py
139 lines (103 loc) · 5.95 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
138
139
from ops import *
class FRACTAL(object):
def __init__(self, x, scale,reuse=False):
self.input=x
self.scale=scale
self.reuse=reuse
self.build_model(reuse=self.reuse)
def build_model(self, reuse=False):
print('Build Model GEN')
with tf.variable_scope('GEN', reuse=reuse):
self.conv1 = conv2d(self.input, 64, [3, 3], scope='conv1', activation=None)
self.head = self.conv1
for idx in range(0,2):
self.head = self.dense_block(self.head, 6, 'Block' + repr(idx))
self.conv2=tf.add(self.head,self.conv1)
self.head=self.conv2
for idx in range(2,4):
self.head = self.dense_block(self.head, 6, 'Block' + repr(idx))
self.conv3=tf.add(self.head,self.conv2)
self.conv3=tf.add(self.conv3, self.conv1)
self.head=self.conv3
for idx in range(4,6):
self.head = self.dense_block(self.head, 6, 'Block' + repr(idx))
self.conv4=tf.add(self.head,self.conv3)
self.head=self.conv4
for idx in range(6,8):
self.head = self.dense_block(self.head, 6, 'Block' + repr(idx))
self.conv5=tf.add(self.head,self.conv4)
self.conv5=tf.add(self.conv5,self.conv3)
self.conv5=tf.add(self.conv5, self.conv1)
self.head=self.conv5
self.out1 = conv2d(self.head, 64, [3, 3], scope='conv2', activation=None)
self.out2 = tf.add(self.conv1, self.out1)
if self.scale==4:
self.conv_up1 = conv2d(self.out2, 64 * self.scale // 2 * self.scale //2 , [3, 3], scope='conv_up1',
activation=None)
self.conv2_1 = tf.depth_to_space(self.conv_up1, self.scale // 2)
self.conv_up2 = conv2d(self.conv2_1, 64 * self.scale// 2 * self.scale//2, [3, 3], scope='conv_up2',
activation=None)
self.conv2_2 = tf.depth_to_space(self.conv_up2, self.scale // 2)
else:
self.conv_up1 = conv2d(self.out2, 64 * self.scale * self.scale, [3, 3], scope='conv_up1',
activation=None)
self.conv2_2 = tf.depth_to_space(self.conv_up1, self.scale)
self.output = conv2d(self.conv2_2, 3, [3, 3], scope='conv_out', activation=None)
tf.add_to_collection('InNOut', self.input)
tf.add_to_collection('InNOut', self.output)
def dense_block(self, input_x, nb_layers, scope):
with tf.variable_scope(scope):
layers_concat = []
layers_concat.append(input_x)
for i in range(nb_layers - 1):
x = tf.concat(layers_concat,axis=-1)
x = conv2d(x, 64, [3,3], scope='conv'+str(i), activation='ReLU')
layers_concat.append(x)
x = tf.concat(layers_concat, axis=-1)
x = conv2d(x, 64,[1,1], scope='conv_fusion')
return tf.add(input_x,0.1*x)
class NMD(object):
def __init__(self, input):
self.input=input
self.build_model()
def build_model(self, reuse=False):
print('Build Model NMD')
with tf.variable_scope("CLASSIFIER", reuse=reuse):
self.conv1_1 = conv2d(self.input, 64, [3, 3], scope='conv1_1', activation='ReLU')
self.conv1_2 = conv2d(self.conv1_1, 64, [3, 3],scope='conv1_2', activation='ReLU')
self.pool1=maxpool(self.conv1_2)
self.conv2_1 = conv2d(self.pool1, 128, [3, 3], scope='conv2_1', activation='ReLU')
self.conv2_2 = conv2d(self.conv2_1, 128, [3, 3], scope='conv2_2', activation='ReLU')
self.pool2 = maxpool(self.conv2_2)
self.conv3_1 = conv2d(self.pool2, 256, [3, 3], scope='conv3_1', activation='ReLU')
self.conv3_2 = conv2d(self.conv3_1, 256, [3, 3], scope='conv3_2', activation='ReLU')
self.pool3 = maxpool(self.conv3_2)
self.conv4_1 = conv2d(self.pool3, 512, [3, 3], scope='conv4_1', activation='ReLU')
self.conv4_2 = conv2d(self.conv4_1, 512, [3, 3], scope='conv4_2', activation='ReLU')
self.pool4 = maxpool(self.conv4_2)
self.conv5_1 = conv2d(self.pool4, 1, [3, 3], scope='conv5_1')
self.logit = tf.reduce_mean(self.conv5_1,axis=(1,2))
self.out = sigmoid(self.logit)
class Discriminator(object):
def __init__(self, input, reuse=False):
self.input = input
self.reuse=reuse
self.build_model()
def build_model(self):
print('Build Model DIS')
with tf.variable_scope("DIS", reuse=self.reuse) as scope:
if self.reuse:
scope.reuse_variables()
self.conv1_1 = SNconv(self.input, 64, [3, 3], scope='conv1_1', activation='leakyReLU')
self.conv1_2 = SNconv(self.conv1_1, 64, [3, 3],strides=2, scope='conv1_2', activation='leakyReLU')
self.conv2_1 = SNconv(self.conv1_2, 128, [3, 3], scope='conv2_1', activation='leakyReLU')
self.conv2_2 = SNconv(self.conv2_1, 128, [3, 3], strides=2, scope='conv2_2', activation='leakyReLU')
self.conv3_1 = SNconv(self.conv2_2, 256, [3, 3], scope='conv3_1', activation='leakyReLU')
self.conv3_2 = SNconv(self.conv3_1, 256, [3, 3],strides=2, scope='conv3_2', activation='leakyReLU')
self.conv4_1 = SNconv(self.conv3_2, 512, [3, 3], scope='conv4_1', activation='leakyReLU')
self.conv4_2 = SNconv(self.conv4_1, 512, [3, 3],strides=2, scope='conv4_2', activation='leakyReLU')
self.conv5_1 = SNconv(self.conv4_2, 1024, [3, 3], scope='conv5_1', activation='leakyReLU')
self.conv5_2 = SNconv(self.conv5_1, 1024, [3, 3], strides=2, scope='conv5_2', activation='leakyReLU')
self.conv6_1 = SNconv(self.conv5_2, 1, [3, 3], scope='conv6_1', activation=None)
self.logit = tf.reduce_mean(self.conv6_1, axis=(1,2))
self.out = tf.nn.sigmoid(self.logit)