programmers_contest/model.py at master · Soonmok/programmers_contest · 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
import numpy as np
from tensorflow import keras
from tensorflow.keras.layers import GlobalAveragePooling2D, Dense, Dropout, Flatten, add, Conv2D, MaxPooling2D, \
    BatchNormalization, Activation, PReLU, ReLU

from utils import stem_block, inception_resnet_A, squeeze_excitation_layer, reduction_A, inception_resnet_B, \
    reduction_B, inception_resnet_C


def build_senet(config, input_x):
    x = stem_block(input_x, scope='stem')

    for i in range(config.trainer.senet.inception_a_iter):
        x = inception_resnet_A(x, scope='Inception_A' + str(i))
        channel = x.shape[-1]
        x = squeeze_excitation_layer(x, out_dim=channel, ratio=config.trainer.senet.reduction_ratio, number=i)

    x = reduction_A(x, scope='Reduction_A')

    channel = x.shape[-1]
    x = squeeze_excitation_layer(x, out_dim=channel, ratio=config.trainer.senet.reduction_ratio, number=6)

    for i in range(config.trainer.senet.inception_b_iter):
        x = inception_resnet_B(x, scope='Inception_B' + str(i))
        channel = x.shape[-1]
        x = squeeze_excitation_layer(x, out_dim=channel, ratio=config.trainer.senet.reduction_ratio, number=7 + i)

    x = reduction_B(x, scope='Reduction_B')

    channel = x.shape[-1]
    x = squeeze_excitation_layer(x, out_dim=channel, ratio=config.trainer.senet.reduction_ratio, number=18)

    for i in range(config.trainer.senet.inception_c_iter):
        x = inception_resnet_C(x, scope='Inception_C' + str(i))
        channel = int(np.shape(x)[-1])
        x = squeeze_excitation_layer(x, out_dim=channel, ratio=config.trainer.senet.reduction_ratio, number=19 + i)

    x = GlobalAveragePooling2D()(x)
    x = Dropout(rate=config.trainer.senet.drop_out)(x, training=config.training)
    x = Flatten()(x)

    x = Dense(config.data.class_num, activation='softmax')(x)
    return x


def get_se_model(config):
    input = keras.Input(shape=(128, 128, 3))
    output = build_senet(config, input)
    se_model = keras.Model(input, output)
    config.training = True
    return se_model


def res_net_block(input_data, filters, conv_size, training):
    x = Conv2D(filters=filters, kernel_size=conv_size, padding='same')(input_data)
    x = BatchNormalization()(x, training=training)
    x = ReLU()(x)
    x = Conv2D(filters=filters, kernel_size=conv_size, activation=None, padding='same')(x)
    x = BatchNormalization()(x, training=training)
    x = add([x, input_data])
    x = Activation('relu')(x)
    return x


def res_net_prelu_block(input_data, filters, conv_size, training=True):
    x = Conv2D(filters=filters, kernel_size=conv_size, padding='same')(input_data)
    x = PReLU()(x)
    x = Conv2D(filters=filters, kernel_size=conv_size, padding='same')(x)
    x = PReLU()(x)
    out = add([input_data, x])
    return out


def get_resnet_model(config):
    inputs = keras.Input(shape=(32, 32, 3))

    x = Conv2D(32, 3, activation='relu')(inputs)
    x = Conv2D(64, 3, activation='relu')(x)
    x = MaxPooling2D(3)(x)

    for i in range(config.trainer.resnet.num_blocks):
        x = res_net_prelu_block(x, 64, 3, training=config.training)

    x = Conv2D(64, 3, activation='relu')(x)
    x = GlobalAveragePooling2D()(x)
    x = Dense(256, activation='relu')(x)
    x = Dropout(config.trainer.drop_out)(x, training=config.training)
    outputs = Dense(config.data.class_num, activation='softmax')(x)
    res_net_model = keras.Model(inputs, outputs)
    return res_net_model


def make_layer(input_x, resnet_block, filters, blocks_num, stride):
    x = Conv2D(filters=filters, kernel_size=3, strides=stride, padding='same')(input_x)
    x = PReLU()(x)
    for i in range(blocks_num):
        x = resnet_block(x, filters=filters, conv_size=3)
    return x


def get_sphere_net(config, input_x):
    type = config.trainer.sphere.type
    # TODO: need to use xavior initializer
    if type is 20:
        layers = [1, 2, 4, 1]
    elif type is 64:
        layers = [3, 7, 16, 3]
    else:
        raise ValueError
    filter_list = [3, 64, 128, 256, 512]
    x = make_layer(input_x, res_net_prelu_block, filter_list[0], layers[0], stride=2)
    x = make_layer(x, res_net_prelu_block, filter_list[1], layers[1], stride=2)
    x = make_layer(x, res_net_prelu_block, filter_list[2], layers[2], stride=2)
    x = make_layer(x, res_net_prelu_block, filter_list[3], layers[3], stride=2)
    x = Flatten()(x)
    x = Dense(512)(x)
    x = Dense(config.data.class_num)(x)
    return x


def get_sphere_model(config):
    input_x = keras.Input(shape=(128, 128, 3))
    output = get_sphere_net(config, input_x)
    sphere_model = keras.Model(input_x, output)
    return sphere_model