-
Notifications
You must be signed in to change notification settings - Fork 24
/
Copy pathdata.py
192 lines (160 loc) · 7.42 KB
/
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
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import torch
import numpy as np
from torch.utils.data import Dataset
import librosa
from glob import glob
import random
from skimage.transform import resize
import pandas as pd
from random import sample
SR=16000
class SpeechDataset(Dataset):
def __init__(self, mode, label_words_dict, wav_list, add_noise, preprocess_fun, preprocess_param = {}, sr=SR, resize_shape=None, is_1d=False):
"""Args:
mode: train or evaluate or test
label_words_dict: a dict of words for labels
wav_list: a list of wav file paths
add_noise: boolean. if background noise should be added
preprocess_fun: function to load/process wav file
preprocess_param: params for preprocess_fun
sr: default 16000
resize_shape: None. only for 2d cnn.
is_1d: boolean. if it is going to be 1d cnn or 2d cnn
"""
self.mode = mode
self.label_words_dict = label_words_dict
self.wav_list = wav_list
self.add_noise = add_noise
self.sr = sr
self.n_silence = int(len(wav_list) * 0.09)
self.preprocess_fun = preprocess_fun
self.preprocess_param = preprocess_param
# read all background noise here
self.background_noises = [librosa.load(x, sr=self.sr)[0] for x in glob("../input/train/audio/_background_noise_/*.wav")]
self.resize_shape = resize_shape
self.is_1d = is_1d
def get_one_noise(self):
"""generates one single noise clip"""
selected_noise = self.background_noises[random.randint(0, len(self.background_noises) - 1)]
# only takes out 16000
start_idx = random.randint(0, len(selected_noise) - 1 - self.sr)
return selected_noise[start_idx:(start_idx + self.sr)]
def get_mix_noises(self, num_noise=1, max_ratio=0.1):
result = np.zeros(self.sr)
for _ in range(num_noise):
result += random.random() * max_ratio * self.get_one_noise()
return result / num_noise if num_noise > 0 else result
def get_one_word_wav(self, idx):
wav = librosa.load(self.wav_list[idx], sr=self.sr)[0]
if len(wav) < self.sr:
wav = np.pad(wav, (0, self.sr - len(wav)), 'constant')
return wav[:self.sr]
def get_silent_wav(self, num_noise=1, max_ratio=0.5):
return self.get_mix_noises(num_noise=num_noise, max_ratio=max_ratio)
def timeshift(self, wav, ms=100):
shift = (self.sr * ms) // 1000
shift = random.randint(-shift, shift)
a = -min(0, shift)
b = max(0, shift)
data = np.pad(wav, (a, b), "constant")
return data[:len(data) - a] if a else data[b:]
def get_noisy_wav(self, idx):
scale = random.uniform(0.75, 1.25)
num_noise = random.choice([1, 2])
max_ratio = random.choice([0.1, 0.5, 1, 1.5])
mix_noise_proba = random.choice([0.1, 0.3])
shift_range = random.randint(80, 120)
one_word_wav = self.get_one_word_wav(idx)
if random.random() < mix_noise_proba:
return scale * (self.timeshift(one_word_wav, shift_range) + self.get_mix_noises(
num_noise, max_ratio))
else:
return one_word_wav
def __len__(self):
if self.mode == 'test':
return len(self.wav_list)
else:
return len(self.wav_list) + self.n_silence
def __getitem__(self, idx):
"""reads one sample"""
if idx < len(self.wav_list):
wav_numpy = self.preprocess_fun(
self.get_one_word_wav(idx) if self.mode != 'train' else self.get_noisy_wav(idx),
**self.preprocess_param)
if self.resize_shape:
wav_numpy = resize(wav_numpy, (self.resize_shape, self.resize_shape), preserve_range=True)
wav_tensor = torch.from_numpy(wav_numpy).float()
if not self.is_1d:
wav_tensor = wav_tensor.unsqueeze(0)
if self.mode == 'test':
return {'spec': wav_tensor, 'id': self.wav_list[idx]}
label = self.label_words_dict[self.wav_list[idx].split("/")[-2]] if self.wav_list[idx].split(
"/")[-2] in self.label_words_dict else len(self.label_words_dict)
return {'spec': wav_tensor, 'id': self.wav_list[idx], 'label': label}
else:
"""generates silence here"""
wav_numpy = self.preprocess_fun(self.get_silent_wav(
num_noise=random.choice([0, 1, 2, 3]),
max_ratio=random.choice([x / 10. for x in range(20)])), **self.preprocess_param)
if self.resize_shape:
wav_numpy = resize(wav_numpy, (self.resize_shape, self.resize_shape), preserve_range=True)
wav_tensor = torch.from_numpy(wav_numpy).float()
if not self.is_1d:
wav_tensor = wav_tensor.unsqueeze(0)
return {'spec': wav_tensor, 'id': 'silence', 'label': len(self.label_words_dict) + 1}
def get_label_dict():
words = ['yes', 'no', 'up', 'down', 'left', 'right', 'on', 'off', 'stop', 'go']
label_to_int = dict(zip(words, range(len(words))))
int_to_label = dict(zip(range(len(words)), words))
int_to_label.update({len(words): 'unknown', len(words) + 1: 'silence'})
return label_to_int, int_to_label
def get_wav_list(words, unknown_ratio=0.2):
full_train_list = glob("../input/train/audio/*/*.wav")
full_test_list = glob("../input/test/audio/*.wav")
# sample full train list
sampled_train_list = []
for w in full_train_list:
l = w.split("/")[-2]
if l not in words:
if random.random() < unknown_ratio:
sampled_train_list.append(w)
else:
sampled_train_list.append(w)
return sampled_train_list, full_test_list
def get_sub_list(num, sub_path):
lst = []
df = pd.read_csv(sub_path)
words = ['yes', 'no', 'up', 'down', 'left', 'right', 'on', 'off', 'stop', 'go', 'silence', 'unknown']
each_num = int(num * 0.085)
for w in words:
tmp = df['fname'][df['label'] == w].sample(each_num).tolist()
lst += ["../input/test/audio/" + x for x in tmp]
return lst
def get_semi_list(words, sub_path, unknown_ratio=0.2, test_ratio=0.2):
train_list, _ = get_wav_list(words=words, unknown_ratio=unknown_ratio)
test_list = get_sub_list(num=int(len(train_list) * test_ratio), sub_path=sub_path)
lst = train_list + test_list
return sample(lst, len(lst))
def preprocess_mfcc(wave):
spectrogram = librosa.feature.melspectrogram(wave, sr=SR, n_mels=40, hop_length=160, n_fft=480, fmin=20, fmax=4000)
idx = [spectrogram > 0]
spectrogram[idx] = np.log(spectrogram[idx])
dct_filters = librosa.filters.dct(n_filters=40, n_input=40)
mfcc = [np.matmul(dct_filters, x) for x in np.split(spectrogram, spectrogram.shape[1], axis=1)]
mfcc = np.hstack(mfcc)
mfcc = mfcc.astype(np.float32)
return mfcc
def preprocess_mel(data, n_mels=40, normalization=False):
spectrogram = librosa.feature.melspectrogram(data, sr=SR, n_mels=n_mels, hop_length=160, n_fft=480, fmin=20, fmax=4000)
spectrogram = librosa.power_to_db(spectrogram)
spectrogram = spectrogram.astype(np.float32)
if normalization:
spectrogram = spectrogram.spectrogram()
spectrogram -= spectrogram
return spectrogram
def preprocess_wav(wav, normalization=True):
data = wav.reshape(1, -1)
if normalization:
mean = data.mean()
data -= mean
return data