BT-FER-WILD/holonet.py at master · rolshoven/BT-FER-WILD · 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
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
from tensorflow.python.keras.layers import Input, Conv2D, BatchNormalization, Activation, MaxPooling2D, Flatten, Lambda, \
    Dense, Add, Concatenate, UpSampling2D
from tensorflow.python.keras.models import Model
from tensorflow.python.keras.regularizers import l2
from tensorflow.python.keras.optimizers import Adam
from tensorflow.nn import crelu
from layers import GlobalCovPooling2D
from losses import island_crossentropy_loss
from model import stn_on_input


def phase_convolutional_block(X):
    X = Conv2D(8, (7, 7), padding='same', kernel_initializer='he_normal')(X)
    X = BatchNormalization()(X)
    X = Lambda(crelu)(X)
    X = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(X)
    return X


def phase_residual_block_1(X):
    X_main = Conv2D(12, (1, 1), padding='same', kernel_initializer='he_normal')(X)
    X_main = BatchNormalization()(X_main)
    X_main = Conv2D(12, (3, 3), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_main = Lambda(crelu)(X_main)
    X_main = Conv2D(32, (1, 1), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_res_branch = Conv2D(32, (1, 1), padding='same', kernel_initializer='he_normal')(X)
    X_res_branch = BatchNormalization()(X_res_branch)
    X_combined = Add()([X_main, X_res_branch])
    X_main = Conv2D(12, (1, 1), padding='same', kernel_initializer='he_normal')(X_combined)
    X_main = BatchNormalization()(X_main)
    X_main = Conv2D(12, (3, 3), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_main = Lambda(crelu)(X_main)
    X_main = Conv2D(32, (1, 1), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_combined = Add()([X_main, X_combined])
    return X_combined


def phase_residual_block_2(X):
    X_main = Conv2D(16, (1, 1), strides=(2, 2), padding='same', kernel_initializer='he_normal')(X)
    X_main = BatchNormalization()(X_main)
    X_main = Conv2D(16, (3, 3), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_main = Lambda(crelu)(X_main)
    X_main = Conv2D(48, (1, 1), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_res_branch = Conv2D(48, (1, 1), strides=(2, 2), padding='same', kernel_initializer='he_normal')(X)
    X_res_branch = BatchNormalization()(X_res_branch)
    X_combined = Add()([X_main, X_res_branch])
    X_main = Conv2D(16, (1, 1), padding='same', kernel_initializer='he_normal')(X_combined)
    X_main = BatchNormalization()(X_main)
    X_main = Conv2D(16, (3, 3), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_main = Lambda(crelu)(X_main)
    X_main = Conv2D(48, (1, 1), padding='same', kernel_initializer='he_normal')(X_main)
    X_main = BatchNormalization()(X_main)
    X_combined = Add()([X_main, X_combined])
    return X_combined


def inception_residual_block(X):
    X1 = Conv2D(24, (1, 1), strides=(2, 2), padding='same', kernel_initializer='he_normal')(X)
    X1 = BatchNormalization()(X1)
    X2 = Conv2D(16, (1, 1), strides=(2, 2), padding='same', kernel_initializer='he_normal')(X)
    X2 = BatchNormalization()(X2)
    X2 = Conv2D(32, (3, 3), padding='same', kernel_initializer='he_normal')(X2)
    X2 = BatchNormalization()(X2)
    X3 = Conv2D(12, (1, 1), strides=(2, 2), padding='same', kernel_initializer='he_normal')(X)
    X3 = BatchNormalization()(X3)
    X3 = Conv2D(16, (3, 3), padding='same', kernel_initializer='he_normal')(X3)
    X3 = BatchNormalization()(X3)
    X3 = Conv2D(16, (3, 3), padding='same', kernel_initializer='he_normal')(X3)
    X3 = BatchNormalization()(X3)
    X4 = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(X)
    X4 = Conv2D(32, (1, 1), padding='same', kernel_initializer='he_normal')(X4)
    X4 = BatchNormalization()(X4)
    concat = Concatenate()([X1, X2, X3, X4])
    concat = Conv2D(64, (1, 1), padding='same', kernel_initializer='he_normal')(concat)
    concat = BatchNormalization()(concat)
    X_res_branch = Conv2D(64, (1, 1), strides=(2, 2), padding='same', kernel_initializer='he_normal')(X)
    X_res_branch = BatchNormalization()(X_res_branch)
    X_combined = Add()([concat, X_res_branch])
    X_combined = MaxPooling2D((2, 2), strides=(2, 2), padding='same')(X_combined)
    return X_combined


def HoloNet(input_shape=(192, 192, 3), regularizers=(0.1, 0.1), classes=8):
    image = Input(shape=input_shape)
    X = phase_convolutional_block(image)
    X = phase_residual_block_1(X)
    X = phase_residual_block_2(X)
    X = inception_residual_block(X)
    X = Flatten()(X)
    X = Dense(1024, kernel_regularizer=l2(regularizers[0]))(X)
    X = Dense(classes, kernel_regularizer=l2(regularizers[1]))(X)
    prediction = Activation('softmax')(X)
    return Model(inputs=image, outputs=prediction)


def HoloNetCov(input_shape=(192, 192, 3), regularizers=(0.1, 0.1), classes=8):
    """ Holonet combined with covariance pooling. """
    image = Input(shape=input_shape)
    X = phase_convolutional_block(image)
    X = phase_residual_block_1(X)
    X = phase_residual_block_2(X)
    X = inception_residual_block(X)
    X = GlobalCovPooling2D(num_iter=5)(X)
    X = Dense(classes, kernel_regularizer=l2(regularizers[1]))(X)
    prediction = Activation('softmax')(X)
    return Model(inputs=image, outputs=prediction)


def get_holonet_stn(input_shape=(192, 192, 3), regularizers=(0.2, 0.2)):
    """ A Holonet with a spatial transformer. """
    image = Input(shape=input_shape)
    X = stn_on_input(image)
    X = phase_convolutional_block(X)
    X = phase_residual_block_1(X)
    X = phase_residual_block_2(X)
    X = inception_residual_block(X)
    X = Flatten()(X)
    X = Dense(1024, kernel_regularizer=l2(regularizers[0]))(X)
    X = Dense(8, kernel_regularizer=l2(regularizers[1]))(X)
    prediction = Activation('softmax')(X)
    return Model(inputs=image, outputs=prediction)


def get_holonet_feature_extractor(input_shape=(192, 192, 3)):
    """ Holonet feature extractor without classification module. """
    image = Input(shape=input_shape)
    X = phase_convolutional_block(image)
    X = phase_residual_block_1(X)
    X = phase_residual_block_2(X)
    X = inception_residual_block(X)
    X = Flatten()(X)
    return Model(inputs=image, outputs=X)


def HoloNetIL(input_shape=(192, 192, 3), num_classes=8, regularizers=(0.2, 0.2), lr=0.01, balance=0.01):
    """ Holonet with Island Loss. """
    image = Input(shape=input_shape)
    X = stn_on_input(image)
    X = phase_convolutional_block(X)
    X = phase_residual_block_1(X)
    X = phase_residual_block_2(X)
    X = inception_residual_block(X)
    X = Flatten()(X)
    fts = Dense(1024, kernel_regularizer=l2(regularizers[0]))(X)
    X = Dense(num_classes, kernel_regularizer=l2(regularizers[1]))(fts)
    prediction = Activation('softmax')(X)
    HoloNet = Model(inputs=image, outputs=prediction)
    HoloNet.compile(optimizer=Adam(lr=lr),
                    loss=island_crossentropy_loss(fts, num_classes=num_classes, balance=balance),
                    metrics=['accuracy'])
    return HoloNet