-
Notifications
You must be signed in to change notification settings - Fork 228
/
audio_data.py
158 lines (132 loc) · 5.8 KB
/
audio_data.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
import os
import os.path
import math
import threading
import torch
import torch.utils.data
import numpy as np
import librosa as lr
import bisect
class WavenetDataset(torch.utils.data.Dataset):
def __init__(self,
dataset_file,
item_length,
target_length,
file_location=None,
classes=256,
sampling_rate=16000,
mono=True,
normalize=False,
dtype=np.uint8,
train=True,
test_stride=100):
# |----receptive_field----|
# |--output_length--|
# example: | | | | | | | | | | | | | | | | | | | | |
# target: | | | | | | | | | |
self.dataset_file = dataset_file
self._item_length = item_length
self._test_stride = test_stride
self.target_length = target_length
self.classes = classes
if not os.path.isfile(dataset_file):
assert file_location is not None, "no location for dataset files specified"
self.mono = mono
self.normalize = normalize
self.sampling_rate = sampling_rate
self.dtype = dtype
self.create_dataset(file_location, dataset_file)
else:
# Unknown parameters of the stored dataset
# TODO Can these parameters be stored, too?
self.mono = None
self.normalize = None
self.sampling_rate = None
self.dtype = None
self.data = np.load(self.dataset_file, mmap_mode='r')
self.start_samples = [0]
self._length = 0
self.calculate_length()
self.train = train
print("one hot input")
# assign every *test_stride*th item to the test set
def create_dataset(self, location, out_file):
print("create dataset from audio files at", location)
self.dataset_file = out_file
files = list_all_audio_files(location)
processed_files = []
for i, file in enumerate(files):
print(" processed " + str(i) + " of " + str(len(files)) + " files")
file_data, _ = lr.load(path=file,
sr=self.sampling_rate,
mono=self.mono)
if self.normalize:
file_data = lr.util.normalize(file_data)
quantized_data = quantize_data(file_data, self.classes).astype(self.dtype)
processed_files.append(quantized_data)
np.savez(self.dataset_file, *processed_files)
def calculate_length(self):
start_samples = [0]
for i in range(len(self.data.keys())):
start_samples.append(start_samples[-1] + len(self.data['arr_' + str(i)]))
available_length = start_samples[-1] - (self._item_length - (self.target_length - 1)) - 1
self._length = math.floor(available_length / self.target_length)
self.start_samples = start_samples
def set_item_length(self, l):
self._item_length = l
self.calculate_length()
def __getitem__(self, idx):
if self._test_stride < 2:
sample_index = idx * self.target_length
elif self.train:
sample_index = idx * self.target_length + math.floor(idx / (self._test_stride-1))
else:
sample_index = self._test_stride * (idx+1) - 1
file_index = bisect.bisect_left(self.start_samples, sample_index) - 1
if file_index < 0:
file_index = 0
if file_index + 1 >= len(self.start_samples):
print("error: sample index " + str(sample_index) + " is to high. Results in file_index " + str(file_index))
position_in_file = sample_index - self.start_samples[file_index]
end_position_in_next_file = sample_index + self._item_length + 1 - self.start_samples[file_index + 1]
if end_position_in_next_file < 0:
file_name = 'arr_' + str(file_index)
this_file = np.load(self.dataset_file, mmap_mode='r')[file_name]
sample = this_file[position_in_file:position_in_file + self._item_length + 1]
else:
# load from two files
file1 = np.load(self.dataset_file, mmap_mode='r')['arr_' + str(file_index)]
file2 = np.load(self.dataset_file, mmap_mode='r')['arr_' + str(file_index + 1)]
sample1 = file1[position_in_file:]
sample2 = file2[:end_position_in_next_file]
sample = np.concatenate((sample1, sample2))
example = torch.from_numpy(sample).type(torch.LongTensor)
one_hot = torch.FloatTensor(self.classes, self._item_length).zero_()
one_hot.scatter_(0, example[:self._item_length].unsqueeze(0), 1.)
target = example[-self.target_length:].unsqueeze(0)
return one_hot, target
def __len__(self):
test_length = math.floor(self._length / self._test_stride)
if self.train:
return self._length - test_length
else:
return test_length
def quantize_data(data, classes):
mu_x = mu_law_encoding(data, classes)
bins = np.linspace(-1, 1, classes)
quantized = np.digitize(mu_x, bins) - 1
return quantized
def list_all_audio_files(location):
audio_files = []
for dirpath, dirnames, filenames in os.walk(location):
for filename in [f for f in filenames if f.endswith((".mp3", ".wav", ".aif", "aiff"))]:
audio_files.append(os.path.join(dirpath, filename))
if len(audio_files) == 0:
print("found no audio files in " + location)
return audio_files
def mu_law_encoding(data, mu):
mu_x = np.sign(data) * np.log(1 + mu * np.abs(data)) / np.log(mu + 1)
return mu_x
def mu_law_expansion(data, mu):
s = np.sign(data) * (np.exp(np.abs(data) * np.log(mu + 1)) - 1) / mu
return s