InceptionResNetV2/modules.py at master · skaldek/InceptionResNetV2 · GitHub

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
import tensorflow as tf
from tensorflow.keras import layers
import hyperparams as hp


def conv2d(inp, filters, kernel, strides=1, padding='same', use_bias=False, activation='relu'):
    x = layers.Conv2D(filters, kernel, strides, padding=padding, use_bias=use_bias)(inp)
    x = layers.BatchNormalization(3, scale=False)(x)
    if activation != 'None':
        x = layers.Activation(activation)(x)
    return x


def stem_block(inp):
    x = conv2d(inp, 32, 3, strides=2, padding='valid')
    x = conv2d(x, 32, 3, padding='valid')
    x = conv2d(x, 64, 3)

    x1 = layers.MaxPool2D(3, strides=2)(x)
    x2 = conv2d(x, 96, 3, strides=2, padding='valid')

    x = layers.Concatenate(3)([x1, x2])

    x1 = conv2d(x, 64, 1)
    x1 = conv2d(x1, 96, 3, padding='valid')

    x2 = conv2d(x, 64, 1)
    x2 = conv2d(x2, 64, (7, 1))
    x2 = conv2d(x2, 64, (1, 7))
    x2 = conv2d(x2, 96, 3, padding='valid')

    x = layers.Concatenate(3)([x1, x2])

    x1 = conv2d(x, 192, 3, strides=2, padding='valid')
    x2 = layers.MaxPool2D(3, strides=2)(x)

    x = layers.Concatenate(3)([x1, x2])

    return x


def block_a(inp):
    x1 = conv2d(inp, 32, 1)

    x2 = conv2d(inp, 32, 1)
    x2 = conv2d(x2, 32, 3)

    x3 = conv2d(inp, 32, 1)
    x3 = conv2d(x3, 48, 3)
    x3 = conv2d(x3, 64, 3)

    x = layers.Concatenate(3)([x1, x2, x3])
    x = conv2d(x, 384, 1, activation='None')

    x = res_con(inp, x, hp.scale_a)

    return x


def reduction_a(inp):
    x1 = layers.MaxPool2D(3, 2)(inp)

    x2 = conv2d(inp, hp.filters_n, 3, 2, 'valid')

    x3 = conv2d(inp, hp.filters_k, 1)
    x3 = conv2d(x3, hp.filters_l, 3)
    x3 = conv2d(x3, hp.filters_m, 3, 2, 'valid')

    x = layers.Concatenate(3)([x1, x2, x3])

    return x


def block_b(inp):
    x1 = conv2d(inp, 192, 1)

    x2 = conv2d(inp, 128, 1)
    x2 = conv2d(x2, 160, (1, 7))
    x2 = conv2d(x2, 192, (7, 1))

    x = layers.Concatenate(3)([x1, x2])
    x = conv2d(x, 1152, 1, activation='None')

    x = res_con(inp, x, hp.scale_b)

    return x


def reduction_b(inp):
    x1 = layers.MaxPool2D(3, 2)(inp)

    x2 = conv2d(inp, 256, 1)
    x2 = conv2d(x2, 384, 3, 2, 'valid')

    x3 = conv2d(inp, 256, 1)
    x3 = conv2d(x3, 256, 3, 2, 'valid')

    x4 = conv2d(inp, 256, 1)
    x4 = conv2d(x4, 256, 3)
    x4 = conv2d(x4, 256, 3, 2, 'valid')

    x = layers.Concatenate(3)([x1, x2, x3, x4])

    return x


def block_c(inp):
    x1 = conv2d(inp, 192, 1)

    x2 = conv2d(inp, 192, 1)
    x2 = conv2d(x2, 224, (1, 3))
    x2 = conv2d(x2, 256, (3, 1))

    x = layers.Concatenate(3)([x1, x2])
    x = conv2d(x, 2048, 1, activation='None')

    x = res_con(inp, x, hp.scale_c)

    return x


def res_con(inp, out, scale_val):
    x = layers.Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
                      output_shape=tf.keras.backend.int_shape(inp)[1:],
                      arguments={'scale': scale_val})([inp, out])
    return x