forked from keums/SingingVoiceDetection
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel_SVAD.py
67 lines (53 loc) · 3 KB
/
model_SVAD.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
from keras.regularizers import l2
from keras.layers import LeakyReLU
from keras.models import Model
from keras.layers.core import Dense, Activation
from keras.layers import Input,Conv1D,Conv2D, MaxPooling2D, GlobalAveragePooling2D, BatchNormalization,Dropout,concatenate
def SVAD_CONV_MultiLayer():
input = Input(shape=(75, 80, 1))
block1= Conv2D(64, (3, 3), name='conv1_1', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(input)
block1 = BatchNormalization()(block1)
block1 = LeakyReLU(0.01)(block1)
block1 = Conv2D(64, (3, 3), name='conv1_2', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(block1)
block1 = BatchNormalization()(block1)
block1 = LeakyReLU(0.01)(block1)
block1 = MaxPooling2D((3, 3))(block1)
block1= Dropout(0.4)(block1)
block2 = Conv2D(128, (3, 3), name='conv2_1', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(block1)
block2 = BatchNormalization()(block2)
block2 = LeakyReLU(0.01)(block2)
block2 = Conv2D(128, (3, 3), name='conv2_2', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(block2)
block2 = BatchNormalization()(block2)
block2 = LeakyReLU(0.01)(block2)
block2 = MaxPooling2D((2, 2))(block2)
block2 = Dropout(0.4)(block2)
block3 = Conv2D(256, (3, 3), name='conv3_1', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(block2)
block3 = BatchNormalization()(block3)
block3 = LeakyReLU(0.01)(block3)
block3 = Conv2D(256, (3, 3), name='conv3_2', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(block3)
block3 = BatchNormalization()(block3)
block3 = LeakyReLU(0.01)(block3)
block3 = MaxPooling2D((2, 2))(block3)
block3 = Dropout(0.4)(block3)
block4 = Conv2D(256, (3, 3), name='conv4_1', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(block3)
block4 = BatchNormalization()(block4)
block4 = LeakyReLU(0.01)(block4)
block4 = Conv2D(256, (3, 3), name='conv4_2', padding='same', kernel_initializer='he_normal',kernel_regularizer=l2(0.))(block4)
block4 = BatchNormalization()(block4)
block4 = LeakyReLU(0.01)(block4)
block4 = MaxPooling2D((2, 2))(block4)
block4 = Dropout(0.4)(block4)
max_11 = MaxPooling2D((8, 8))(block1)
max_22 = MaxPooling2D((4, 4))(block2)
max_33 = MaxPooling2D((2, 2))(block3)
lastLayer = concatenate([max_11, max_22,max_33,block4],axis = -1)
conv1x1 = Conv2D(64, (1, 1), name='conv1x1_1', padding='same', kernel_initializer='he_normal')(lastLayer)
conv1x1= BatchNormalization()(conv1x1)
conv1x1 = LeakyReLU(0.01)(conv1x1)
conv1x1 = Conv2D(1, (1, 1), name='conv1x1_2', padding='same', kernel_initializer='he_normal')(conv1x1)
conv1x1 = BatchNormalization()(conv1x1)
conv1x1 = LeakyReLU(0.01)(conv1x1)
gap = GlobalAveragePooling2D()(conv1x1)
output = Activation("sigmoid",name="sigmoid")(gap)
model = Model(inputs = input, outputs = output)
return model