-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathav_cnn_x_mlp.py
156 lines (124 loc) · 5.5 KB
/
av_cnn_x_mlp.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
from keras.models import Model
from keras.layers import Dense, Dropout, Flatten, Reshape
from keras.layers import Convolution2D, MaxPooling2D, Deconvolution2D
from keras.layers import merge, Input
from keras.layers.advanced_activations import PReLU
from keras.layers.normalization import BatchNormalization
from Datasets.data_config import *
def get_model(args):
# Dataset config
assert args.dataset.lower() == 'avletters'
config = data_constants['avletters']
inputCNNshape = config['inputCNNshape']
inputMLPshape = config['inputMLPshape']
nb_classes = config['nb_classes']
# Build the CNN - pre-cross-connections
inputCNN = Input(shape=inputCNNshape)
inputNorm = BatchNormalization()(inputCNN)
conv = Convolution2D(16, 3, 3, border_mode='same', activation='relu')(inputNorm)
conv = Convolution2D(16, 3, 3, border_mode='same', activation='relu')(conv)
conv = BatchNormalization()(conv)
pool = MaxPooling2D((2,2), strides=(2, 2))(conv)
pool = Dropout(0.25)(pool)
# Build the MLP - pre-cross-connections
inputMLP = Input(shape=inputMLPshape)
fcMLP = Dense(128, activation='relu')(inputMLP)
fcMLP = BatchNormalization()(fcMLP)
fcMLP = Dropout(0.5)(fcMLP)
# Add the 1st round of cross-connections - CNN to MLP
x21 = Convolution2D(16, 1, 1, border_mode='same')(pool)
x21 = PReLU()(x21)
x21 = BatchNormalization()(x21)
x21 = Flatten()(x21)
x21 = Dense(64)(x21)
x21 = PReLU()(x21)
# Add 1st shortcut (residual connection) from CNN input to MLP
short1_2dto1d = MaxPooling2D((4,4), strides=(4,4))(inputNorm)
short1_2dto1d = Dropout(0.25)(short1_2dto1d)
short1_2dto1d = Flatten()(short1_2dto1d)
short1_2dto1d = Dense(128)(short1_2dto1d)
short1_2dto1d = PReLU()(short1_2dto1d)
# MLP to CNN - cross-connections
x12 = Dense(33*23)(fcMLP)
x12 = PReLU()(x12)
x12 = BatchNormalization()(x12)
x12 = Dropout(0.5)(x12)
x12 = Reshape((33,23,1))(x12)
x12 = Deconvolution2D(16, 8, 8, output_shape=(None, 40, 30, 16), border_mode='valid')(x12)
x12 = PReLU()(x12)
# 1st shortcut (residual connection) from MLP input to CNN
short1_1dto2d = Dense(33*23)(inputMLP)
short1_1dto2d = PReLU()(short1_1dto2d)
short1_1dto2d = BatchNormalization()(short1_1dto2d)
short1_1dto2d = Dropout(0.25)(short1_1dto2d)
short1_1dto2d = Reshape((33,23,1))(short1_1dto2d)
short1_1dto2d = Deconvolution2D(16, 8, 8, output_shape=(None, 40, 30, 16), border_mode='valid')(
short1_1dto2d)
short1_1dto2d = PReLU()(short1_1dto2d)
short1_1dto2d = BatchNormalization()(short1_1dto2d)
# CNN - post-cross-connections 1
pool = merge([pool, short1_1dto2d], mode='sum')
x12map = merge([x12, pool], mode='concat', concat_axis=3)
x12map = BatchNormalization()(x12map)
x12map = Dropout(0.5)(x12map)
conv = Convolution2D(32, 3, 3, border_mode='same', activation='relu')(x12map)
conv = Convolution2D(32, 3, 3, border_mode='same', activation='relu')(conv)
conv = BatchNormalization()(conv)
pool = MaxPooling2D((2,2), strides=(2, 2))(conv)
pool = Dropout(0.25)(pool)
# MLP - post-cross-connections 1
fcMLP = merge([fcMLP, short1_2dto1d], mode='sum')
x21map = merge([x21, fcMLP], mode='concat', concat_axis=1)
x21map = BatchNormalization()(x21map)
x21map = Dropout(0.5)(x21map)
fcMLP = Dense(128, activation='relu')(x21map)
fcMLP = BatchNormalization()(fcMLP)
fcMLP = Dropout(0.5)(fcMLP)
# Add the 2nd round of cross-connections - CNN to MLP
x21 = Convolution2D(32, 1, 1, border_mode='same')(pool)
x21 = PReLU()(x21)
x21 = BatchNormalization()(x21)
x21 = Flatten()(x21)
x21 = Dense(128)(x21)
x21 = PReLU()(x21)
# Add 2nd shortcut (residual connection) from CNN input to MLP
short2_2dto1d = MaxPooling2D((8,8), strides=(8,4))(inputNorm)
short2_2dto1d = Dropout(0.25)(short2_2dto1d)
short2_2dto1d = Flatten()(short2_2dto1d)
short2_2dto1d = Dense(128)(short2_2dto1d)
short2_2dto1d = PReLU()(short2_2dto1d)
# MLP to CNN - cross-connections
x12 = Dense(17*12)(fcMLP)
x12 = PReLU()(x12)
x12 = BatchNormalization()(x12)
x12 = Dropout(0.5)(x12)
x12 = Reshape((17,12,1))(x12)
x12 = Deconvolution2D(32, 4, 4, output_shape=(None, 20, 15, 32), border_mode='valid')(x12)
x12 = PReLU()(x12)
# 2nd shortcut (residual connection) from MLP input to CNN
short2_1dto2d = Dense(17*12)(inputMLP)
short2_1dto2d = PReLU()(short2_1dto2d)
short2_1dto2d = BatchNormalization()(short2_1dto2d)
short2_1dto2d = Dropout(0.5)(short2_1dto2d)
short2_1dto2d = Reshape((17,12,1))(short2_1dto2d)
short2_1dto2d = Deconvolution2D(32, 4, 4, output_shape=(None, 20, 15, 32), border_mode='valid')(
short2_1dto2d)
short2_1dto2d = PReLU()(short2_1dto2d)
# CNN - post cross-connections 2
pool = merge([pool, short2_1dto2d], mode='sum')
x12map = merge([x12, pool], mode='concat', concat_axis=3)
x12map = BatchNormalization()(x12map)
x12map = Dropout(0.5)(x12map)
reshape = Flatten()(x12map)
fcCNN = Dense(256, activation='relu')(reshape)
# Merge the CNN and MLP models
fcMLP = merge([fcMLP, short2_2dto1d], mode='sum')
merged = merge([x21, fcCNN, fcMLP], mode='concat')
merged = BatchNormalization()(merged)
merged = Dropout(0.5)(merged)
fc = Dense(512, activation='relu')(merged)
fc = Dropout(0.5)(fc)
out = Dense(nb_classes, activation='softmax')(fc)
# Return the model object
model = Model(input=[inputCNN, inputMLP], output=out)
return model