Create meldataset.py
Browse files
Hiformer_Checkpoint_Libri_24khz/meldataset.py
ADDED
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
import torch
|
5 |
+
import torch.utils.data
|
6 |
+
import numpy as np
|
7 |
+
from librosa.util import normalize
|
8 |
+
from scipy.io.wavfile import read
|
9 |
+
import torchaudio
|
10 |
+
import librosa
|
11 |
+
from librosa.filters import mel as librosa_mel_fn
|
12 |
+
|
13 |
+
MAX_WAV_VALUE = 32768.0
|
14 |
+
import soundfile as sf
|
15 |
+
|
16 |
+
|
17 |
+
def normalize_audio(wav):
|
18 |
+
return wav / torch.max(torch.abs(torch.from_numpy(wav))) # Correct peak normalization
|
19 |
+
|
20 |
+
def load_wav_librosa(full_path):
|
21 |
+
data, sampling_rate = librosa.load(full_path, sr=24_000)
|
22 |
+
return data, sampling_rate
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
def load_wav_scipy(full_path):
|
27 |
+
sampling_rate, data = read(full_path)
|
28 |
+
return data, sampling_rate
|
29 |
+
|
30 |
+
def load_wav(full_path):
|
31 |
+
try:
|
32 |
+
return load_wav_scipy(full_path)
|
33 |
+
except:
|
34 |
+
# print('using librosa...')
|
35 |
+
return load_wav_librosa(full_path)
|
36 |
+
|
37 |
+
|
38 |
+
def dynamic_range_compression(x, C=1, clip_val=1e-5):
|
39 |
+
return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
|
40 |
+
|
41 |
+
|
42 |
+
def dynamic_range_decompression(x, C=1):
|
43 |
+
return np.exp(x) / C
|
44 |
+
|
45 |
+
|
46 |
+
def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
|
47 |
+
return torch.log(torch.clamp(x, min=clip_val) * C)
|
48 |
+
|
49 |
+
|
50 |
+
def dynamic_range_decompression_torch(x, C=1):
|
51 |
+
return torch.exp(x) / C
|
52 |
+
|
53 |
+
|
54 |
+
def spectral_normalize_torch(magnitudes):
|
55 |
+
output = dynamic_range_compression_torch(magnitudes)
|
56 |
+
return output
|
57 |
+
|
58 |
+
|
59 |
+
def spectral_de_normalize_torch(magnitudes):
|
60 |
+
output = dynamic_range_decompression_torch(magnitudes)
|
61 |
+
return output
|
62 |
+
|
63 |
+
|
64 |
+
mel_basis = {}
|
65 |
+
hann_window = {}
|
66 |
+
|
67 |
+
|
68 |
+
def mel_spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
|
69 |
+
|
70 |
+
|
71 |
+
# y = torch.clamp(y, -1, 1)
|
72 |
+
|
73 |
+
|
74 |
+
# if torch.min(y) < -1.:
|
75 |
+
# # y = torch.clamp(y, min = -1)
|
76 |
+
# # print('min value is ', torch.min(y))
|
77 |
+
# if torch.max(y) > 1.:
|
78 |
+
# y = torch.clamp(y, max = -1)
|
79 |
+
# print('max value is ', torch.max(y))
|
80 |
+
|
81 |
+
global mel_basis, hann_window
|
82 |
+
if fmax not in mel_basis:
|
83 |
+
mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
|
84 |
+
mel_basis[str(fmax)+'_'+str(y.device)] = torch.from_numpy(mel).float().to(y.device)
|
85 |
+
hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
|
86 |
+
|
87 |
+
y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
|
88 |
+
y = y.squeeze(1)
|
89 |
+
|
90 |
+
# complex tensor as default, then use view_as_real for future pytorch compatibility
|
91 |
+
spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[str(y.device)],
|
92 |
+
center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=True)
|
93 |
+
spec = torch.view_as_real(spec)
|
94 |
+
spec = torch.sqrt(spec.pow(2).sum(-1)+(1e-9))
|
95 |
+
|
96 |
+
spec = torch.matmul(mel_basis[str(fmax)+'_'+str(y.device)], spec)
|
97 |
+
spec = spectral_normalize_torch(spec)
|
98 |
+
|
99 |
+
return spec
|
100 |
+
|
101 |
+
|
102 |
+
|
103 |
+
to_mel = torchaudio.transforms.MelSpectrogram(n_mels=80, n_fft=2048, win_length=1200, hop_length=300)
|
104 |
+
|
105 |
+
|
106 |
+
# to_mel = torchaudio.transforms.MelSpectrogram(n_mels=80, n_fft=2048, win_length=1200, hop_length=300)
|
107 |
+
|
108 |
+
mean, std = -4, 4
|
109 |
+
5
|
110 |
+
def preproces(wave,to_mel=to_mel, device='cpu'):
|
111 |
+
|
112 |
+
to_mel = to_mel.to(device)
|
113 |
+
# wave_tensor = torch.from_numpy(wave).float()
|
114 |
+
mel_tensor = to_mel(wave)
|
115 |
+
mel_tensor = (torch.log(1e-5 + mel_tensor) - mean) / std
|
116 |
+
return mel_tensor
|
117 |
+
|
118 |
+
|
119 |
+
def get_dataset_filelist(a):
|
120 |
+
with open(a.input_training_file, 'r', encoding='utf-8') as fi:
|
121 |
+
training_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + ('' if '' not in x else ''))
|
122 |
+
for x in fi.read().split('\n') if len(x) > 0]
|
123 |
+
|
124 |
+
with open(a.input_validation_file, 'r', encoding='utf-8') as fi:
|
125 |
+
validation_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + ('' if '' not in x else ''))
|
126 |
+
for x in fi.read().split('\n') if len(x) > 0]
|
127 |
+
return training_files, validation_files
|
128 |
+
|
129 |
+
|
130 |
+
class MelDataset(torch.utils.data.Dataset):
|
131 |
+
def __init__(self, training_files, segment_size, n_fft, num_mels,
|
132 |
+
hop_size, win_size, sampling_rate, fmin, fmax, split=True, shuffle=True, n_cache_reuse=1,
|
133 |
+
device=None, fmax_loss=None, fine_tuning=False, base_mels_path=None):
|
134 |
+
self.audio_files = training_files
|
135 |
+
random.seed(1234)
|
136 |
+
if shuffle:
|
137 |
+
random.shuffle(self.audio_files)
|
138 |
+
self.segment_size = segment_size
|
139 |
+
self.sampling_rate = sampling_rate
|
140 |
+
self.split = split
|
141 |
+
self.n_fft = n_fft
|
142 |
+
self.num_mels = num_mels
|
143 |
+
self.hop_size = hop_size
|
144 |
+
self.win_size = win_size
|
145 |
+
self.fmin = fmin
|
146 |
+
self.fmax = fmax
|
147 |
+
self.fmax_loss = fmax_loss
|
148 |
+
self.cached_wav = None
|
149 |
+
self.n_cache_reuse = n_cache_reuse
|
150 |
+
self._cache_ref_count = 0
|
151 |
+
self.device = device
|
152 |
+
self.fine_tuning = fine_tuning
|
153 |
+
self.base_mels_path = base_mels_path
|
154 |
+
|
155 |
+
def __getitem__(self, index):
|
156 |
+
filename = self.audio_files[index]
|
157 |
+
if self._cache_ref_count == 0:
|
158 |
+
audio, sampling_rate = load_wav(filename)
|
159 |
+
audio = audio / MAX_WAV_VALUE
|
160 |
+
if not self.fine_tuning:
|
161 |
+
audio = normalize(audio) * 0.95
|
162 |
+
self.cached_wav = audio
|
163 |
+
if sampling_rate != self.sampling_rate:
|
164 |
+
audio = librosa.resample(audio, orig_sr= sampling_rate, target_sr= self.sampling_rate)
|
165 |
+
# raise ValueError("{} SR doesn't match target {} SR, {}".format(
|
166 |
+
# sampling_rate, self.sampling_rate, filename))
|
167 |
+
self._cache_ref_count = self.n_cache_reuse
|
168 |
+
else:
|
169 |
+
audio = self.cached_wav
|
170 |
+
self._cache_ref_count -= 1
|
171 |
+
|
172 |
+
audio = torch.FloatTensor(audio)
|
173 |
+
audio = audio.unsqueeze(0)
|
174 |
+
|
175 |
+
if not self.fine_tuning:
|
176 |
+
if self.split:
|
177 |
+
if audio.size(1) >= self.segment_size:
|
178 |
+
max_audio_start = audio.size(1) - self.segment_size
|
179 |
+
audio_start = random.randint(0, max_audio_start)
|
180 |
+
audio = audio[:, audio_start:audio_start+self.segment_size]
|
181 |
+
else:
|
182 |
+
audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), 'constant')
|
183 |
+
|
184 |
+
# mel = mel_spectrogram(audio, self.n_fft, self.num_mels,
|
185 |
+
# self.sampling_rate, self.hop_size, self.win_size, self.fmin, self.fmax,
|
186 |
+
# center=False)
|
187 |
+
|
188 |
+
mel = preproces(audio)
|
189 |
+
else:
|
190 |
+
mel = np.load(
|
191 |
+
os.path.join(self.base_mels_path, os.path.splitext(os.path.split(filename)[-1])[0] + '.npy'))
|
192 |
+
mel = torch.from_numpy(mel)
|
193 |
+
|
194 |
+
if len(mel.shape) < 3:
|
195 |
+
mel = mel.unsqueeze(0)
|
196 |
+
|
197 |
+
if self.split:
|
198 |
+
frames_per_seg = math.ceil(self.segment_size / self.hop_size)
|
199 |
+
|
200 |
+
if audio.size(1) >= self.segment_size:
|
201 |
+
mel_start = random.randint(0, mel.size(2) - frames_per_seg - 1)
|
202 |
+
mel = mel[:, :, mel_start:mel_start + frames_per_seg]
|
203 |
+
audio = audio[:, mel_start * self.hop_size:(mel_start + frames_per_seg) * self.hop_size]
|
204 |
+
else:
|
205 |
+
mel = torch.nn.functional.pad(mel, (0, frames_per_seg - mel.size(2)), 'constant')
|
206 |
+
audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), 'constant')
|
207 |
+
|
208 |
+
mel_loss = mel_spectrogram(audio, self.n_fft, self.num_mels,
|
209 |
+
self.sampling_rate, self.hop_size, self.win_size, self.fmin, self.fmax_loss,
|
210 |
+
center=False)
|
211 |
+
|
212 |
+
# mel_loss = mel_spectrogram(audio)
|
213 |
+
if mel.shape[-1] != mel_loss.shape[-1]:
|
214 |
+
mel = mel[..., :mel_loss.shape[-1]]
|
215 |
+
|
216 |
+
return (mel.squeeze(), audio.squeeze(0), filename, mel_loss.squeeze())
|
217 |
+
|
218 |
+
def __len__(self):
|
219 |
+
return len(self.audio_files)
|