-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
163 lines (108 loc) · 5.09 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
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
#!/usr/bin/python
# -*- coding: utf-8 -*-
from keras.models import Model
from keras.layers import Activation, Input, concatenate, Dropout, Dense, Flatten
from keras.layers import MaxPooling2D, Conv2D, AveragePooling2D, BatchNormalization
def create(classes_num=1000, image_height=299, image_width=299, image_channel=3):
inputs = Input((image_height, image_width, image_channel))
# 299 x 299 x 3
net = block_stem(inputs)
# 4 x Inception-A ( Output: 35 x 35 x 384 )
for i in range(4):
net = block_inception_a(net)
# Reduction-A ( Output: 17 x 17 x 1024 )
net = block_reduction_a(net)
# 7 x Inception-B ( Output: 17 x 17 x 1024 )
for i in range(7):
net = block_inception_b(net)
# Reduction-B ( Output: 8 x 8 x 1536 )
net = block_reduction_b(net)
# 3 x Inception-C ( Output: 8 x 8 x 1536 )
for i in range(3):
net = block_inception_c(net)
# Average Pooling ( Output: 1536 )
net = AveragePooling2D((8, 8))(net)
# Dropout ( keep 0.8 )
net = Dropout(0.2)(net)
net = Flatten()(net)
# Output
outputs = Dense(units=classes_num, activation='softmax')(net)
return Model(inputs, outputs, name='Inception-v4')
def block_stem(inputs):
net = conv2d(inputs, 32, (3, 3), strides=(2, 2), padding='valid')
net = conv2d(net, 32, (3, 3), padding='valid')
net = conv2d(net, 64, (3, 3))
branch_1 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(net)
branch_2 = conv2d(net, 96, (3, 3), strides=(2, 2), padding='valid')
net = concatenate([branch_1, branch_2])
branch_1 = conv2d(net, 64, (1, 1))
branch_1 = conv2d(branch_1, 96, (3, 3), padding='valid')
branch_2 = conv2d(net, 64, (1, 1))
branch_2 = conv2d(branch_2, 64, (7, 1))
branch_2 = conv2d(branch_2, 64, (1, 7))
branch_2 = conv2d(branch_2, 96, (3, 3), padding='valid')
net = concatenate([branch_1, branch_2])
branch_1 = conv2d(net, 192, (3, 3), strides=(2, 2), padding='valid') # different from the paper
branch_2 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(net)
net = concatenate([branch_1, branch_2])
return net
def block_inception_a(inputs):
branch_1 = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(inputs)
branch_1 = conv2d(branch_1, 96, (1, 1))
branch_2 = conv2d(inputs, 96, (1, 1))
branch_3 = conv2d(inputs, 64, (1, 1))
branch_3 = conv2d(branch_3, 96, (3, 3))
branch_4 = conv2d(inputs, 64, (1, 1))
branch_4 = conv2d(branch_4, 96, (3, 3))
branch_4 = conv2d(branch_4, 96, (3, 3))
return concatenate([branch_1, branch_2, branch_3, branch_4])
def block_inception_b(inputs):
branch_1 = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(inputs)
branch_1 = conv2d(branch_1, 128, (1, 1))
branch_2 = conv2d(inputs, 384, (1, 1))
branch_3 = conv2d(inputs, 192, (1, 1))
branch_3 = conv2d(branch_3, 224, (1, 7))
branch_3 = conv2d(branch_3, 256, (7, 1)) # different from the paper
branch_4 = conv2d(inputs, 192, (1, 1))
branch_4 = conv2d(branch_4, 192, (1, 7))
branch_4 = conv2d(branch_4, 224, (7, 1))
branch_4 = conv2d(branch_4, 224, (1, 7))
branch_4 = conv2d(branch_4, 256, (7, 1))
return concatenate([branch_1, branch_2, branch_3, branch_4])
def block_inception_c(inputs):
branch_1 = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(inputs)
branch_1 = conv2d(branch_1, 256, (1, 1))
branch_2 = conv2d(inputs, 256, (1, 1))
branch_3 = conv2d(inputs, 384, (1, 1))
branch_3_1 = conv2d(branch_3, 256, (1, 3))
branch_3_2 = conv2d(branch_3, 256, (3, 1))
branch_4 = conv2d(inputs, 384, (1, 1))
branch_4 = conv2d(branch_4, 448, (1, 3))
branch_4 = conv2d(branch_4, 512, (3, 1))
branch_4_1 = conv2d(branch_4, 256, (3, 1))
branch_4_2 = conv2d(branch_4, 256, (1, 3))
return concatenate([branch_1, branch_2, branch_3_1, branch_3_2, branch_4_1, branch_4_2])
def block_reduction_a(inputs):
branch_1 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(inputs)
branch_2 = conv2d(inputs, 384, (3, 3), strides=(2, 2), padding='valid')
branch_3 = conv2d(inputs, 192, (1, 1))
branch_3 = conv2d(branch_3, 224, (3, 3))
branch_3 = conv2d(branch_3, 256, (3, 3), strides=(2, 2), padding='valid')
return concatenate([branch_1, branch_2, branch_3])
def block_reduction_b(inputs):
branch_1 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(inputs)
branch_2 = conv2d(inputs, 192, (1, 1))
branch_2 = conv2d(branch_2, 192, (3, 3), strides=(2, 2), padding='valid')
branch_3 = conv2d(inputs, 256, (1, 1))
branch_3 = conv2d(branch_3, 256, (1, 7))
branch_3 = conv2d(branch_3, 320, (7, 1))
branch_3 = conv2d(branch_3, 320, (3, 3), strides=(2, 2), padding='valid')
return concatenate([branch_1, branch_2, branch_3])
def conv2d(net, filters, kernel_size, strides=(1, 1), padding='same'):
net = Conv2D(filters=filters, kernel_size=kernel_size, strides=strides, padding=padding, use_bias=False)(net)
net = BatchNormalization()(net)
net = Activation('relu')(net)
return net
if __name__ == '__main__':
model = create()
model.summary()