audio-denoising/model_unet.py at master · DZDL/audio-denoising · 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
import numpy as np
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Conv2D, LeakyReLU, MaxPooling2D, Dropout, concatenate, UpSampling2D
from tensorflow.keras.optimizers import Adam
from tensorflow.keras import backend
import tensorflow as tf
print(tf.__version__)

#Unet network
def unet(pretrained_weights = None,input_size = (128,128,1)):
    #size filter input
    size_filter_in = 16
    #normal initialization of weights
    kernel_init = 'he_normal'
    #To apply leaky relu after the conv layer
    activation_layer = None
    inputs = Input(input_size)
    conv1 = Conv2D(size_filter_in, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(inputs)
    conv1 = LeakyReLU()(conv1)
    conv1 = Conv2D(size_filter_in, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv1)
    conv1 = LeakyReLU()(conv1)
    pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
    conv2 = Conv2D(size_filter_in*2, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(pool1)
    conv2 = LeakyReLU()(conv2)
    conv2 = Conv2D(size_filter_in*2, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv2)
    conv2 = LeakyReLU()(conv2)
    pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
    conv3 = Conv2D(size_filter_in*4, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(pool2)
    conv3 = LeakyReLU()(conv3)
    conv3 = Conv2D(size_filter_in*4, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv3)
    conv3 = LeakyReLU()(conv3)
    pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
    conv4 = Conv2D(size_filter_in*8, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(pool3)
    conv4 = LeakyReLU()(conv4)
    conv4 = Conv2D(size_filter_in*8, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv4)
    conv4 = LeakyReLU()(conv4)
    drop4 = Dropout(0.5)(conv4)
    pool4 = MaxPooling2D(pool_size=(2, 2))(drop4)

    conv5 = Conv2D(size_filter_in*16, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(pool4)
    conv5 = LeakyReLU()(conv5)
    conv5 = Conv2D(size_filter_in*16, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv5)
    conv5 = LeakyReLU()(conv5)
    drop5 = Dropout(0.5)(conv5)

    up6 = Conv2D(size_filter_in*8, 2, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(UpSampling2D(size = (2,2))(drop5))
    up6 = LeakyReLU()(up6)
    merge6 = concatenate([drop4,up6], axis = 3)
    conv6 = Conv2D(size_filter_in*8, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(merge6)
    conv6 = LeakyReLU()(conv6)
    conv6 = Conv2D(size_filter_in*8, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv6)
    conv6 = LeakyReLU()(conv6)
    up7 = Conv2D(size_filter_in*4, 2, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(UpSampling2D(size = (2,2))(conv6))
    up7 = LeakyReLU()(up7)
    merge7 = concatenate([conv3,up7], axis = 3)
    conv7 = Conv2D(size_filter_in*4, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(merge7)
    conv7 = LeakyReLU()(conv7)
    conv7 = Conv2D(size_filter_in*4, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv7)
    conv7 = LeakyReLU()(conv7)
    up8 = Conv2D(size_filter_in*2, 2, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(UpSampling2D(size = (2,2))(conv7))
    up8 = LeakyReLU()(up8)
    merge8 = concatenate([conv2,up8], axis = 3)
    conv8 = Conv2D(size_filter_in*2, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(merge8)
    conv8 = LeakyReLU()(conv8)
    conv8 = Conv2D(size_filter_in*2, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv8)
    conv8 = LeakyReLU()(conv8)

    up9 = Conv2D(size_filter_in, 2, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(UpSampling2D(size = (2,2))(conv8))
    up9 = LeakyReLU()(up9)
    merge9 = concatenate([conv1,up9], axis = 3)
    conv9 = Conv2D(size_filter_in, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(merge9)
    conv9 = LeakyReLU()(conv9)
    conv9 = Conv2D(size_filter_in, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv9)
    conv9 = LeakyReLU()(conv9)
    conv9 = Conv2D(2, 3, activation = activation_layer, padding = 'same', kernel_initializer = kernel_init)(conv9)
    conv9 = LeakyReLU()(conv9)
    conv10 = Conv2D(1, 1, activation = 'tanh')(conv9)

    model = Model(inputs,conv10)

    model.compile(optimizer = 'adam', loss = tf.keras.losses.Huber(), metrics = ['mae'])

    #model.summary()

    if(pretrained_weights):
    	model.load_weights(pretrained_weights)

    return model