-
Notifications
You must be signed in to change notification settings - Fork 0
/
wavenet.py
167 lines (132 loc) · 4.86 KB
/
wavenet.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
157
158
159
160
161
162
163
164
165
166
__author__ = "WoongwonLee@j-marple.com"
import librosa
from torch.autograd import Variable
import torch.utils.data as data_utils
import torch.nn.functional as F
import torch.optim as optim
import torch.nn as nn
import torch
import numpy as np
sampling_rate = 16000
num_layer = 10
receptive_field = (num_layer) - 2
num_stack = 3
class CausalConv1d(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, dilation):
super(CausalConv1d, self).__init__()
self.kernel_size = kernel_size
self.dilation = dilation
self.conv1d = torch.nn.Conv1d(in_channels,
out_channels,
kernel_size,
padding=(kernel_size - 1),
dilation=dilation)
def forward(self, input):
conv1d_out = self.conv1d(input)
# remove k-1 values from the end:
if self.kernel_size > 1:
return conv1d_out[:, :, 0:-(self.kernel_size - 1)]
else:
return conv1d_out
class ResidualBlock(nn.Module):
def __init__(self, num_filters, kernel_size, dilation):
super(ResidualBlock, self).__init__()
self.kernel_size = kernel_size
self.dilation = dilation
self.causal = CausalConv1d(num_filters, num_filters,
kernel_size, dilation=dilation)
self.conv1d = nn.Conv1d(num_filters, num_filters, 1)
self.activation_f = nn.Tanh()
self.activation_g = nn.Sigmoid()
def forward(self, input):
# for residual connection
origin_input = input
# gate
conv = self.causal.forward(input)
tanh = self.activation_f(conv)
sigmoid = self.activation_g(conv)
out = tanh * sigmoid
# output of residual block
res = self.conv1d(out)
res += origin_input
# skip connection
skip = self.conv1d(out)
return res, skip
class Wavenet(nn.Module):
def __init__(self):
super(Wavenet, self).__init__()
self.dilation = 1
self.causal = CausalConv1d(in_channels=1,
out_channels=256,
kernel_size=2,
dilation=1)
self.residual = ResidualBlock(num_filters=256,
kernel_size=2,
dilation=self.dilation)
self.relu = nn.ReLU()
self.conv1d = nn.Conv1d(256, 256, 1, padding=1)
self.softmax = nn.Softmax()
def forward(self, input):
skip_connections = []
# output[1, 256, 12814]
output = self.causal.forward(input)
for j in range(num_stack):
self.dilation = 1
for i in range(num_layer):
output, skip = self.residual.forward(output)
self.dilation *= 2
skip_connections.append(skip)
# skip_connections[30, 1, 256, 12814]
skip_connections = torch.stack(skip_connections)
# output[1, 256, 12814]
output = torch.sum(skip_connections, dim=0)
# output[1, 256, 12814]
output = self.relu(output)
output = self.conv1d(output)
output = output[:, :, 1:-1]
# output[1, 256, 12814]
output = self.relu(output)
output = self.conv1d(output)
output = output[:, :, 1:-1]
# output[1, 256, 12814]
# output = output.view(output, )
output = self.softmax(output)
return output
def mu_law(data, mu):
data = data.astype('float64', casting='safe')
max = np.max(data)
min = np.min(data)
data = (data - min) / (max - min/2.) - 1.
data = np.sign(data) * (np.log(1 + mu * np.abs(data))
/ np.log(1 + mu))
return data
def digitize_audio(audio):
audio = (audio + 1.)/2.
max_value = np.iinfo('uint8').max
audio *= max_value
audio = audio.astype('uint8')
return audio
def import_audio(audio_path):
audio, sr = librosa.load(audio_path, sr=sampling_rate)
audio = mu_law(audio, mu=255)
inputs = digitize_audio(audio)
return inputs
def data_generator(audio):
datas, labels = [], []
for i in range(np.shape(audio)[2] - receptive_field - 1):
datas.append(audio[:, :, i: i + receptive_field])
labels.append(audio[:, :, (i + 1) * receptive_field + 1])
return datas, labels
if __name__ == "__main__":
model = Wavenet()
# import audio and quantize in to 255 integers
inputs = import_audio('voice.wav')
print('input shape is ', np.shape(inputs))
inputs = np.reshape(inputs, [1, 1, np.shape(inputs)[0]])
datas, labels = data_generator(inputs)
inputs = torch.from_numpy(inputs)
model.train()
model.cpu()
inputs = Variable(inputs).float().cpu()
outputs = model.forward(inputs)
print(outputs.cpu())