forked from jjihwan/Voice-Cloning
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathvocal_remover.py
197 lines (155 loc) · 6.83 KB
/
vocal_remover.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
import argparse
import os
import librosa
import numpy as np
import soundfile as sf
import torch
from tqdm import tqdm
from lib import dataset
from lib import nets
from lib import spec_utils
from lib import utils
from glob import glob
class Separator(object):
def __init__(self, model, device=None, batchsize=1, cropsize=256, postprocess=False):
self.model = model
self.offset = model.offset
self.device = device
self.batchsize = batchsize
self.cropsize = cropsize
self.postprocess = postprocess
def _postprocess(self, X_spec, mask):
if self.postprocess:
mask_mag = np.abs(mask)
mask_mag = spec_utils.merge_artifacts(mask_mag)
mask = mask_mag * np.exp(1.j * np.angle(mask))
X_mag = np.abs(X_spec)
X_phase = np.angle(X_spec)
y_spec = mask * X_mag * np.exp(1.j * X_phase)
v_spec = (1 - mask) * X_mag * np.exp(1.j * X_phase)
# y_spec = X_spec * mask
# v_spec = X_spec - y_spec
return y_spec, v_spec
def _separate(self, X_spec_pad, roi_size):
X_dataset = []
patches = (X_spec_pad.shape[2] - 2 * self.offset) // roi_size
for i in range(patches):
start = i * roi_size
X_spec_crop = X_spec_pad[:, :, start:start + self.cropsize]
X_dataset.append(X_spec_crop)
X_dataset = np.asarray(X_dataset)
self.model.eval()
with torch.no_grad():
mask_list = []
# To reduce the overhead, dataloader is not used.
for i in tqdm(range(0, patches, self.batchsize)):
X_batch = X_dataset[i: i + self.batchsize]
X_batch = torch.from_numpy(X_batch).to(self.device)
mask = self.model.predict_mask(torch.abs(X_batch))
mask = mask.detach().cpu().numpy()
mask = np.concatenate(mask, axis=2)
mask_list.append(mask)
mask = np.concatenate(mask_list, axis=2)
return mask
def separate(self, X_spec):
n_frame = X_spec.shape[2]
pad_l, pad_r, roi_size = dataset.make_padding(n_frame, self.cropsize, self.offset)
X_spec_pad = np.pad(X_spec, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_spec_pad /= np.abs(X_spec).max()
mask = self._separate(X_spec_pad, roi_size)
mask = mask[:, :, :n_frame]
y_spec, v_spec = self._postprocess(X_spec, mask)
return y_spec, v_spec
def separate_tta(self, X_spec):
n_frame = X_spec.shape[2]
pad_l, pad_r, roi_size = dataset.make_padding(n_frame, self.cropsize, self.offset)
X_spec_pad = np.pad(X_spec, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_spec_pad /= X_spec_pad.max()
mask = self._separate(X_spec_pad, roi_size)
pad_l += roi_size // 2
pad_r += roi_size // 2
X_spec_pad = np.pad(X_spec, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_spec_pad /= X_spec_pad.max()
mask_tta = self._separate(X_spec_pad, roi_size)
mask_tta = mask_tta[:, :, roi_size // 2:]
mask = (mask[:, :, :n_frame] + mask_tta[:, :, :n_frame]) * 0.5
y_spec, v_spec = self._postprocess(X_spec, mask)
return y_spec, v_spec
MODEL_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'models')
DEFAULT_MODEL_PATH = os.path.join(MODEL_DIR, 'baseline.pth')
def main():
p = argparse.ArgumentParser()
p.add_argument('--gpu', '-g', type=int, default=-1)
p.add_argument('--pretrained_model', '-P', type=str, default=DEFAULT_MODEL_PATH)
p.add_argument('--input_dir', type=str, default=None)
p.add_argument('--input', '-i', required=False)
p.add_argument('--sr', '-r', type=int, default=44100)
p.add_argument('--n_fft', '-f', type=int, default=2048)
p.add_argument('--hop_length', '-H', type=int, default=1024)
p.add_argument('--batchsize', '-B', type=int, default=4)
p.add_argument('--cropsize', '-c', type=int, default=256)
p.add_argument('--output_image', '-I', action='store_true')
p.add_argument('--tta', '-t', action='store_true')
p.add_argument('--postprocess', '-p', action='store_true')
p.add_argument('--output_dir', '-o', type=str, default='output')
args = p.parse_args()
print('loading model...', end=' ')
device = torch.device('cpu')
if args.gpu >= 0:
if torch.cuda.is_available():
device = torch.device('cuda:{}'.format(args.gpu))
elif torch.backends.mps.is_available() and torch.backends.mps.is_built():
device = torch.device('mps')
model = nets.CascadedNet(args.n_fft, args.hop_length, 32, 128)
model.load_state_dict(torch.load(args.pretrained_model, map_location='cpu'))
model.to(device)
print('done')
if args.input_dir is None:
args.input = [args.input]
else:
args.input = sorted(glob(args.input_dir + '/*'))
for input in args.input:
print(f'loading wave source from {input}...')
X, sr = librosa.load(
input, sr=args.sr, mono=False, dtype=np.float32, res_type='kaiser_fast'
)
basename = os.path.splitext(os.path.basename(input))[0]
print('done')
if X.ndim == 1:
# mono to stereo
X = np.asarray([X, X])
print('stft of wave source...', end=' ')
X_spec = spec_utils.wave_to_spectrogram(X, args.hop_length, args.n_fft)
print('done')
sp = Separator(
model=model,
device=device,
batchsize=args.batchsize,
cropsize=args.cropsize,
postprocess=args.postprocess
)
if args.tta:
y_spec, v_spec = sp.separate_tta(X_spec)
else:
y_spec, v_spec = sp.separate(X_spec)
print('validating output directory...', end=' ')
vocal_dir = os.path.join(args.output_dir, 'vocals/')
instrument_dir = os.path.join(args.output_dir, 'instruments/')
os.makedirs(vocal_dir, exist_ok=True)
os.makedirs(instrument_dir, exist_ok=True)
print('done')
print('inverse stft of instruments...', end=' ')
wave = spec_utils.spectrogram_to_wave(y_spec, hop_length=args.hop_length)
print('done')
sf.write('{}{}_Instruments.wav'.format(instrument_dir, basename), wave.T, sr)
print('inverse stft of vocals...', end=' ')
wave = spec_utils.spectrogram_to_wave(v_spec, hop_length=args.hop_length)
print('done')
sf.write('{}{}_Vocals.wav'.format(vocal_dir, basename), wave.T, sr)
if args.output_image:
image = spec_utils.spectrogram_to_image(y_spec)
utils.imwrite('{}{}_Instruments.jpg'.format(vocal_dir, basename), image)
image = spec_utils.spectrogram_to_image(v_spec)
utils.imwrite('{}{}_Vocals.jpg'.format(instrument_dir, basename), image)
if __name__ == '__main__':
main()