forked from yxlllc/ReFlow-VAE-SVC
-
Notifications
You must be signed in to change notification settings - Fork 0
/
preprocess.py
211 lines (183 loc) · 7.79 KB
/
preprocess.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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
import os
import numpy as np
import random
import librosa
import torch
try:
import torch_musa
use_torch_musa = True
except ImportError:
use_torch_musa = False
import pyworld as pw
import parselmouth
import argparse
import shutil
from logger import utils
from tqdm import tqdm
from reflow.extractors import F0_Extractor, Volume_Extractor, Units_Encoder
from reflow.vocoder import Vocoder
from logger.utils import traverse_dir
import concurrent.futures
def parse_args(args=None, namespace=None):
"""Parse command-line arguments."""
parser = argparse.ArgumentParser()
parser.add_argument(
"-c",
"--config",
type=str,
required=True,
help="path to the config file")
parser.add_argument(
"-d",
"--device",
type=str,
default=None,
required=False,
help="cpu/cuda/musa, auto if not set")
return parser.parse_args(args=args, namespace=namespace)
def preprocess(path, f0_extractor, volume_extractor, mel_extractor, units_encoder, sample_rate, hop_size, device = 'cuda', use_pitch_aug = False, extensions = ['wav']):
path_srcdir = os.path.join(path, 'audio')
path_unitsdir = os.path.join(path, 'units')
path_f0dir = os.path.join(path, 'f0')
path_volumedir = os.path.join(path, 'volume')
path_augvoldir = os.path.join(path, 'aug_vol')
path_meldir = os.path.join(path, 'mel')
path_augmeldir = os.path.join(path, 'aug_mel')
path_skipdir = os.path.join(path, 'skip')
# list files
filelist = traverse_dir(
path_srcdir,
extensions=extensions,
is_pure=True,
is_sort=True,
is_ext=True)
# pitch augmentation dictionary
pitch_aug_dict = {}
# run
def process(file):
binfile = file+'.npy'
path_srcfile = os.path.join(path_srcdir, file)
path_unitsfile = os.path.join(path_unitsdir, binfile)
path_f0file = os.path.join(path_f0dir, binfile)
path_volumefile = os.path.join(path_volumedir, binfile)
path_augvolfile = os.path.join(path_augvoldir, binfile)
path_melfile = os.path.join(path_meldir, binfile)
path_augmelfile = os.path.join(path_augmeldir, binfile)
path_skipfile = os.path.join(path_skipdir, file)
# load audio
audio, _ = librosa.load(path_srcfile, sr=sample_rate)
if len(audio.shape) > 1:
audio = librosa.to_mono(audio)
audio_t = torch.from_numpy(audio).float().to(device)
audio_t = audio_t.unsqueeze(0)
# extract volume
volume = volume_extractor.extract(audio)
# extract mel and volume augmentaion
if mel_extractor is not None:
mel_t = mel_extractor.extract(audio_t, sample_rate)
mel = mel_t.squeeze().to('cpu').numpy()
max_amp = float(torch.max(torch.abs(audio_t))) + 1e-5
max_shift = min(1, np.log10(1/max_amp))
log10_vol_shift = random.uniform(-1, max_shift)
if use_pitch_aug:
keyshift = random.uniform(-5, 5)
else:
keyshift = 0
aug_mel_t = mel_extractor.extract(audio_t * (10 ** log10_vol_shift), sample_rate, keyshift = keyshift)
aug_mel = aug_mel_t.squeeze().to('cpu').numpy()
aug_vol = volume_extractor.extract(audio * (10 ** log10_vol_shift))
# units encode
units_t = units_encoder.encode(audio_t, sample_rate, hop_size)
units = units_t.squeeze().to('cpu').numpy()
# extract f0
f0 = f0_extractor.extract(audio, uv_interp = False)
uv = f0 == 0
if len(f0[~uv]) > 0:
# interpolate the unvoiced f0
f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
# save npy
os.makedirs(os.path.dirname(path_unitsfile), exist_ok=True)
np.save(path_unitsfile, units)
os.makedirs(os.path.dirname(path_f0file), exist_ok=True)
np.save(path_f0file, f0)
os.makedirs(os.path.dirname(path_volumefile), exist_ok=True)
np.save(path_volumefile, volume)
if mel_extractor is not None:
pitch_aug_dict[file] = keyshift
os.makedirs(os.path.dirname(path_melfile), exist_ok=True)
np.save(path_melfile, mel)
os.makedirs(os.path.dirname(path_augmelfile), exist_ok=True)
np.save(path_augmelfile, aug_mel)
os.makedirs(os.path.dirname(path_augvolfile), exist_ok=True)
np.save(path_augvolfile, aug_vol)
else:
print('\n[Error] F0 extraction failed: ' + path_srcfile)
os.makedirs(os.path.dirname(path_skipfile), exist_ok=True)
shutil.move(path_srcfile, os.path.dirname(path_skipfile))
print('This file has been moved to ' + path_skipfile)
print('Preprocess the audio clips in :', path_srcdir)
# single process
for file in tqdm(filelist, total=len(filelist)):
process(file)
if mel_extractor is not None:
path_pitchaugdict = os.path.join(path, 'pitch_aug_dict.npy')
np.save(path_pitchaugdict, pitch_aug_dict)
# multi-process (have bugs)
'''
with concurrent.futures.ProcessPoolExecutor(max_workers=2) as executor:
list(tqdm(executor.map(process, filelist), total=len(filelist)))
'''
if __name__ == '__main__':
# parse commands
cmd = parse_args()
device = cmd.device
if device is None:
if torch.cuda.is_available():
device = 'cuda'
elif use_torch_musa:
if torch.musa.is_available():
device = 'musa'
else:
device = 'cpu'
else:
device = 'cpu'
# load config
args = utils.load_config(cmd.config)
sample_rate = args.data.sampling_rate
hop_size = args.data.block_size
extensions = args.data.extensions
# initialize f0 extractor
f0_extractor = F0_Extractor(
args.data.f0_extractor,
args.data.sampling_rate,
args.data.block_size,
args.data.f0_min,
args.data.f0_max)
# initialize volume extractor
volume_extractor = Volume_Extractor(args.data.block_size)
# initialize mel extractor
mel_extractor = None
use_pitch_aug = False
if args.model.type in ['RectifiedFlow_VAE']:
mel_extractor = Vocoder(args.vocoder.type, args.vocoder.ckpt, device = device)
if mel_extractor.vocoder_sample_rate != sample_rate or mel_extractor.vocoder_hop_size != hop_size:
mel_extractor = None
print('Unmatch vocoder parameters, mel extraction is ignored!')
elif args.model.use_pitch_aug:
use_pitch_aug = True
# initialize units encoder
if args.data.encoder == 'cnhubertsoftfish':
cnhubertsoft_gate = args.data.cnhubertsoft_gate
else:
cnhubertsoft_gate = 10
units_encoder = Units_Encoder(
args.data.encoder,
args.data.encoder_ckpt,
args.data.encoder_sample_rate,
args.data.encoder_hop_size,
cnhubertsoft_gate=cnhubertsoft_gate,
device = device)
# preprocess training set
preprocess(args.data.train_path, f0_extractor, volume_extractor, mel_extractor, units_encoder, sample_rate, hop_size, device = device, use_pitch_aug = use_pitch_aug, extensions = extensions)
# preprocess validation set
preprocess(args.data.valid_path, f0_extractor, volume_extractor, mel_extractor, units_encoder, sample_rate, hop_size, device = device, use_pitch_aug = False, extensions = extensions)