-
Notifications
You must be signed in to change notification settings - Fork 3
/
net.py
119 lines (90 loc) · 4.39 KB
/
net.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
from keras.layers import Input, Dense, Conv2D, Activation
from keras.layers import Add, Multiply, Conv2DTranspose, Concatenate
from keras.layers import BatchNormalization, Flatten
from keras.models import Model
from keras.layers.advanced_activations import LeakyReLU
from keras import backend as K
import math
import tensorflow as tf
from util import *
def discriminator():
input_len = 128
target = Input(shape=(input_len, input_len, 3))
reconstruction = Input(shape=(input_len, input_len, 3))
x = Concatenate()([target, reconstruction])
x = LeakyReLU(alpha=0.2)(BatchNormalization()(Conv2D(32, kernel_size=4, strides=2)(x)))
x = LeakyReLU(alpha=0.2)(BatchNormalization()(Conv2D(64, kernel_size=4, strides=2)(x)))
x = LeakyReLU(alpha=0.2)(BatchNormalization()(Conv2D(128, kernel_size=4, strides=2)(x)))
x = LeakyReLU(alpha=0.2)(BatchNormalization()(Conv2D(256, kernel_size=4, strides=2)(x)))
x = Conv2D(1, kernel_size=1)(x)
x = Flatten()(x)
x = Dense(2, activation='sigmoid')(x)
model = Model([target, reconstruction], x)
return model
def generator():
input_len = 128
input = Input(shape=(input_len, input_len, 3))
# f function param
f_ch = 32
# g function param
c = 16
w = 16
h = 16
g1_conv = input_len - 1 - h
g2_conv = (int(input_len / 2) - 1) - 1 - h
g3_conv = int((int(input_len / 2) - 1) / 2) - 1 - 1 - h
g4_deconv = h - ( int((int((int(input_len / 2) - 1) / 2) - 1) / 2) - 1 - 1 - 2)
g5_deconv = h - ( int((int((int((int(input_len / 2) - 1) / 2) - 1) / 2) - 1) / 2) - 1 - 1 - 2)
g6_deconv = h - ( int((int((int((int((int(input_len / 2) - 1) / 2) - 1) / 2) - 1) / 2) - 1) / 2) - 1 - 1)
f1 = LeakyReLU(alpha=0.2)(Conv2D(f_ch, kernel_size=3)(input))
g1 = Conv2D(c, kernel_size=g1_conv)(f1)
x2 = Conv2D(3, kernel_size=4, strides=2)(input)
f2 = LeakyReLU(alpha=0.2)(Conv2D(f_ch, kernel_size=3)(x2))
g2 = Conv2D(c, kernel_size=g2_conv)(f2)
x3 = Conv2D(3, kernel_size=4, strides=2)(x2)
f3 = LeakyReLU(alpha=0.2)(Conv2D(f_ch, kernel_size=3)(x3))
g3 = Conv2D(c, kernel_size=g3_conv)(f3)
x4 = Conv2D(3, kernel_size=4, strides=2)(x3)
f4 = LeakyReLU(alpha=0.2)(Conv2D(f_ch, kernel_size=3)(x4))
g4 = Conv2DTranspose(c, kernel_size=g4_deconv)(f4)
x5 = Conv2D(3, kernel_size=4, strides=2)(x4)
f5 = LeakyReLU(alpha=0.2)(Conv2D(f_ch, kernel_size=3)(x5))
g5 = Conv2DTranspose(c, kernel_size=g5_deconv)(f5)
x6 = Conv2D(3, kernel_size=4, strides=2)(x5)
f6 = LeakyReLU(alpha=0.2)(Conv2D(f_ch, kernel_size=1)(x6))
g6 = Conv2DTranspose(c, kernel_size=g6_deconv)(f6)
fe = Add()([g1, g2, g3, g4, g5, g6])
def acr(weight_matrix): # adaptive_codelength_regularization
alpha = 0.01
c, h, w = definedCHW()
x = K.round(32 * weight_matrix + 0.5) / 32
num = tf.log(K.sum(K.abs(x)))
den = tf.log(tf.constant(10, dtype=num.dtype))
l1 = num / den
return alpha * l1 / (c * h * w)
g = Conv2D(c, kernel_size=3, name='code', activity_regularizer=acr)(fe) # add regularization
g_d = Conv2DTranspose(c, kernel_size=3)(g)
g6_d = Conv2D(f_ch, kernel_size=g6_deconv)(g_d)
f6_d = LeakyReLU(alpha=0.2)(Conv2DTranspose(3, kernel_size=1)(g6_d))
x6_d = Conv2DTranspose(3, kernel_size=4, strides=2)(f6_d)
g5_d = Conv2D(f_ch, kernel_size=g5_deconv)(g_d)
f5_d = LeakyReLU(alpha=0.2)(Conv2DTranspose(3, kernel_size=3)(g5_d))
x6_f5_d = Add()([x6_d, f5_d])
x5_d = Conv2DTranspose(3, kernel_size=4, strides=2)(x6_f5_d)
g4_d = Conv2D(f_ch, kernel_size=g4_deconv)(g_d)
f4_d = LeakyReLU(alpha=0.2)(Conv2DTranspose(3, kernel_size=3)(g4_d))
x5_f4_d = Add()([x5_d, f4_d])
x4_d = Conv2DTranspose(3, kernel_size=4, strides=2)(x5_f4_d)
g3_d = Conv2DTranspose(f_ch, kernel_size=g3_conv)(g_d)
f3_d = LeakyReLU(alpha=0.2)(Conv2DTranspose(3, kernel_size=3)(g3_d))
x4_f3_d = Add()([x4_d, f3_d])
x3_d = Conv2DTranspose(3, kernel_size=5, strides=2)(x4_f3_d) # kernel_size=5
g2_d = Conv2DTranspose(f_ch, kernel_size=g2_conv)(g_d)
f2_d = LeakyReLU(alpha=0.2)(Conv2DTranspose(3, kernel_size=3)(g2_d))
x3_f2_d = Add()([x3_d, f2_d])
x2_d = Conv2DTranspose(3, kernel_size=4, strides=2)(x3_f2_d)
g1_d = Conv2DTranspose(f_ch, kernel_size=g1_conv)(g_d)
f1_d = LeakyReLU(alpha=0.2)(Conv2DTranspose(3, kernel_size=3)(g1_d))
output = Add()([x2_d, f1_d])
model = Model(input, output)
return model