Spaces:
Running
Running
Upload 2 files
Browse files- src/rmvpe.py +409 -0
- src/rvc.py +151 -0
src/rmvpe.py
ADDED
@@ -0,0 +1,409 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from librosa.filters import mel
|
6 |
+
|
7 |
+
|
8 |
+
class BiGRU(nn.Module):
|
9 |
+
def __init__(self, input_features, hidden_features, num_layers):
|
10 |
+
super(BiGRU, self).__init__()
|
11 |
+
self.gru = nn.GRU(
|
12 |
+
input_features,
|
13 |
+
hidden_features,
|
14 |
+
num_layers=num_layers,
|
15 |
+
batch_first=True,
|
16 |
+
bidirectional=True,
|
17 |
+
)
|
18 |
+
|
19 |
+
def forward(self, x):
|
20 |
+
return self.gru(x)[0]
|
21 |
+
|
22 |
+
|
23 |
+
class ConvBlockRes(nn.Module):
|
24 |
+
def __init__(self, in_channels, out_channels, momentum=0.01):
|
25 |
+
super(ConvBlockRes, self).__init__()
|
26 |
+
self.conv = nn.Sequential(
|
27 |
+
nn.Conv2d(
|
28 |
+
in_channels=in_channels,
|
29 |
+
out_channels=out_channels,
|
30 |
+
kernel_size=(3, 3),
|
31 |
+
stride=(1, 1),
|
32 |
+
padding=(1, 1),
|
33 |
+
bias=False,
|
34 |
+
),
|
35 |
+
nn.BatchNorm2d(out_channels, momentum=momentum),
|
36 |
+
nn.ReLU(),
|
37 |
+
nn.Conv2d(
|
38 |
+
in_channels=out_channels,
|
39 |
+
out_channels=out_channels,
|
40 |
+
kernel_size=(3, 3),
|
41 |
+
stride=(1, 1),
|
42 |
+
padding=(1, 1),
|
43 |
+
bias=False,
|
44 |
+
),
|
45 |
+
nn.BatchNorm2d(out_channels, momentum=momentum),
|
46 |
+
nn.ReLU(),
|
47 |
+
)
|
48 |
+
if in_channels != out_channels:
|
49 |
+
self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
|
50 |
+
self.is_shortcut = True
|
51 |
+
else:
|
52 |
+
self.is_shortcut = False
|
53 |
+
|
54 |
+
def forward(self, x):
|
55 |
+
if self.is_shortcut:
|
56 |
+
return self.conv(x) + self.shortcut(x)
|
57 |
+
else:
|
58 |
+
return self.conv(x) + x
|
59 |
+
|
60 |
+
|
61 |
+
class Encoder(nn.Module):
|
62 |
+
def __init__(
|
63 |
+
self,
|
64 |
+
in_channels,
|
65 |
+
in_size,
|
66 |
+
n_encoders,
|
67 |
+
kernel_size,
|
68 |
+
n_blocks,
|
69 |
+
out_channels=16,
|
70 |
+
momentum=0.01,
|
71 |
+
):
|
72 |
+
super(Encoder, self).__init__()
|
73 |
+
self.n_encoders = n_encoders
|
74 |
+
self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
|
75 |
+
self.layers = nn.ModuleList()
|
76 |
+
self.latent_channels = []
|
77 |
+
for i in range(self.n_encoders):
|
78 |
+
self.layers.append(
|
79 |
+
ResEncoderBlock(
|
80 |
+
in_channels, out_channels, kernel_size, n_blocks, momentum=momentum
|
81 |
+
)
|
82 |
+
)
|
83 |
+
self.latent_channels.append([out_channels, in_size])
|
84 |
+
in_channels = out_channels
|
85 |
+
out_channels *= 2
|
86 |
+
in_size //= 2
|
87 |
+
self.out_size = in_size
|
88 |
+
self.out_channel = out_channels
|
89 |
+
|
90 |
+
def forward(self, x):
|
91 |
+
concat_tensors = []
|
92 |
+
x = self.bn(x)
|
93 |
+
for i in range(self.n_encoders):
|
94 |
+
_, x = self.layers[i](x)
|
95 |
+
concat_tensors.append(_)
|
96 |
+
return x, concat_tensors
|
97 |
+
|
98 |
+
|
99 |
+
class ResEncoderBlock(nn.Module):
|
100 |
+
def __init__(
|
101 |
+
self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01
|
102 |
+
):
|
103 |
+
super(ResEncoderBlock, self).__init__()
|
104 |
+
self.n_blocks = n_blocks
|
105 |
+
self.conv = nn.ModuleList()
|
106 |
+
self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
|
107 |
+
for i in range(n_blocks - 1):
|
108 |
+
self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
|
109 |
+
self.kernel_size = kernel_size
|
110 |
+
if self.kernel_size is not None:
|
111 |
+
self.pool = nn.AvgPool2d(kernel_size=kernel_size)
|
112 |
+
|
113 |
+
def forward(self, x):
|
114 |
+
for i in range(self.n_blocks):
|
115 |
+
x = self.conv[i](x)
|
116 |
+
if self.kernel_size is not None:
|
117 |
+
return x, self.pool(x)
|
118 |
+
else:
|
119 |
+
return x
|
120 |
+
|
121 |
+
|
122 |
+
class Intermediate(nn.Module): #
|
123 |
+
def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
|
124 |
+
super(Intermediate, self).__init__()
|
125 |
+
self.n_inters = n_inters
|
126 |
+
self.layers = nn.ModuleList()
|
127 |
+
self.layers.append(
|
128 |
+
ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum)
|
129 |
+
)
|
130 |
+
for i in range(self.n_inters - 1):
|
131 |
+
self.layers.append(
|
132 |
+
ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum)
|
133 |
+
)
|
134 |
+
|
135 |
+
def forward(self, x):
|
136 |
+
for i in range(self.n_inters):
|
137 |
+
x = self.layers[i](x)
|
138 |
+
return x
|
139 |
+
|
140 |
+
|
141 |
+
class ResDecoderBlock(nn.Module):
|
142 |
+
def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
|
143 |
+
super(ResDecoderBlock, self).__init__()
|
144 |
+
out_padding = (0, 1) if stride == (1, 2) else (1, 1)
|
145 |
+
self.n_blocks = n_blocks
|
146 |
+
self.conv1 = nn.Sequential(
|
147 |
+
nn.ConvTranspose2d(
|
148 |
+
in_channels=in_channels,
|
149 |
+
out_channels=out_channels,
|
150 |
+
kernel_size=(3, 3),
|
151 |
+
stride=stride,
|
152 |
+
padding=(1, 1),
|
153 |
+
output_padding=out_padding,
|
154 |
+
bias=False,
|
155 |
+
),
|
156 |
+
nn.BatchNorm2d(out_channels, momentum=momentum),
|
157 |
+
nn.ReLU(),
|
158 |
+
)
|
159 |
+
self.conv2 = nn.ModuleList()
|
160 |
+
self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
|
161 |
+
for i in range(n_blocks - 1):
|
162 |
+
self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
|
163 |
+
|
164 |
+
def forward(self, x, concat_tensor):
|
165 |
+
x = self.conv1(x)
|
166 |
+
x = torch.cat((x, concat_tensor), dim=1)
|
167 |
+
for i in range(self.n_blocks):
|
168 |
+
x = self.conv2[i](x)
|
169 |
+
return x
|
170 |
+
|
171 |
+
|
172 |
+
class Decoder(nn.Module):
|
173 |
+
def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
|
174 |
+
super(Decoder, self).__init__()
|
175 |
+
self.layers = nn.ModuleList()
|
176 |
+
self.n_decoders = n_decoders
|
177 |
+
for i in range(self.n_decoders):
|
178 |
+
out_channels = in_channels // 2
|
179 |
+
self.layers.append(
|
180 |
+
ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum)
|
181 |
+
)
|
182 |
+
in_channels = out_channels
|
183 |
+
|
184 |
+
def forward(self, x, concat_tensors):
|
185 |
+
for i in range(self.n_decoders):
|
186 |
+
x = self.layers[i](x, concat_tensors[-1 - i])
|
187 |
+
return x
|
188 |
+
|
189 |
+
|
190 |
+
class DeepUnet(nn.Module):
|
191 |
+
def __init__(
|
192 |
+
self,
|
193 |
+
kernel_size,
|
194 |
+
n_blocks,
|
195 |
+
en_de_layers=5,
|
196 |
+
inter_layers=4,
|
197 |
+
in_channels=1,
|
198 |
+
en_out_channels=16,
|
199 |
+
):
|
200 |
+
super(DeepUnet, self).__init__()
|
201 |
+
self.encoder = Encoder(
|
202 |
+
in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels
|
203 |
+
)
|
204 |
+
self.intermediate = Intermediate(
|
205 |
+
self.encoder.out_channel // 2,
|
206 |
+
self.encoder.out_channel,
|
207 |
+
inter_layers,
|
208 |
+
n_blocks,
|
209 |
+
)
|
210 |
+
self.decoder = Decoder(
|
211 |
+
self.encoder.out_channel, en_de_layers, kernel_size, n_blocks
|
212 |
+
)
|
213 |
+
|
214 |
+
def forward(self, x):
|
215 |
+
x, concat_tensors = self.encoder(x)
|
216 |
+
x = self.intermediate(x)
|
217 |
+
x = self.decoder(x, concat_tensors)
|
218 |
+
return x
|
219 |
+
|
220 |
+
|
221 |
+
class E2E(nn.Module):
|
222 |
+
def __init__(
|
223 |
+
self,
|
224 |
+
n_blocks,
|
225 |
+
n_gru,
|
226 |
+
kernel_size,
|
227 |
+
en_de_layers=5,
|
228 |
+
inter_layers=4,
|
229 |
+
in_channels=1,
|
230 |
+
en_out_channels=16,
|
231 |
+
):
|
232 |
+
super(E2E, self).__init__()
|
233 |
+
self.unet = DeepUnet(
|
234 |
+
kernel_size,
|
235 |
+
n_blocks,
|
236 |
+
en_de_layers,
|
237 |
+
inter_layers,
|
238 |
+
in_channels,
|
239 |
+
en_out_channels,
|
240 |
+
)
|
241 |
+
self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
|
242 |
+
if n_gru:
|
243 |
+
self.fc = nn.Sequential(
|
244 |
+
BiGRU(3 * 128, 256, n_gru),
|
245 |
+
nn.Linear(512, 360),
|
246 |
+
nn.Dropout(0.25),
|
247 |
+
nn.Sigmoid(),
|
248 |
+
)
|
249 |
+
else:
|
250 |
+
self.fc = nn.Sequential(
|
251 |
+
nn.Linear(3 * N_MELS, N_CLASS), nn.Dropout(0.25), nn.Sigmoid()
|
252 |
+
)
|
253 |
+
|
254 |
+
def forward(self, mel):
|
255 |
+
mel = mel.transpose(-1, -2).unsqueeze(1)
|
256 |
+
x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
|
257 |
+
x = self.fc(x)
|
258 |
+
return x
|
259 |
+
|
260 |
+
|
261 |
+
class MelSpectrogram(torch.nn.Module):
|
262 |
+
def __init__(
|
263 |
+
self,
|
264 |
+
is_half,
|
265 |
+
n_mel_channels,
|
266 |
+
sampling_rate,
|
267 |
+
win_length,
|
268 |
+
hop_length,
|
269 |
+
n_fft=None,
|
270 |
+
mel_fmin=0,
|
271 |
+
mel_fmax=None,
|
272 |
+
clamp=1e-5,
|
273 |
+
):
|
274 |
+
super().__init__()
|
275 |
+
n_fft = win_length if n_fft is None else n_fft
|
276 |
+
self.hann_window = {}
|
277 |
+
mel_basis = mel(
|
278 |
+
sr=sampling_rate,
|
279 |
+
n_fft=n_fft,
|
280 |
+
n_mels=n_mel_channels,
|
281 |
+
fmin=mel_fmin,
|
282 |
+
fmax=mel_fmax,
|
283 |
+
htk=True,
|
284 |
+
)
|
285 |
+
mel_basis = torch.from_numpy(mel_basis).float()
|
286 |
+
self.register_buffer("mel_basis", mel_basis)
|
287 |
+
self.n_fft = win_length if n_fft is None else n_fft
|
288 |
+
self.hop_length = hop_length
|
289 |
+
self.win_length = win_length
|
290 |
+
self.sampling_rate = sampling_rate
|
291 |
+
self.n_mel_channels = n_mel_channels
|
292 |
+
self.clamp = clamp
|
293 |
+
self.is_half = is_half
|
294 |
+
|
295 |
+
def forward(self, audio, keyshift=0, speed=1, center=True):
|
296 |
+
factor = 2 ** (keyshift / 12)
|
297 |
+
n_fft_new = int(np.round(self.n_fft * factor))
|
298 |
+
win_length_new = int(np.round(self.win_length * factor))
|
299 |
+
hop_length_new = int(np.round(self.hop_length * speed))
|
300 |
+
keyshift_key = str(keyshift) + "_" + str(audio.device)
|
301 |
+
if keyshift_key not in self.hann_window:
|
302 |
+
self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(
|
303 |
+
audio.device
|
304 |
+
)
|
305 |
+
fft = torch.stft(
|
306 |
+
audio,
|
307 |
+
n_fft=n_fft_new,
|
308 |
+
hop_length=hop_length_new,
|
309 |
+
win_length=win_length_new,
|
310 |
+
window=self.hann_window[keyshift_key],
|
311 |
+
center=center,
|
312 |
+
return_complex=True,
|
313 |
+
)
|
314 |
+
magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
|
315 |
+
if keyshift != 0:
|
316 |
+
size = self.n_fft // 2 + 1
|
317 |
+
resize = magnitude.size(1)
|
318 |
+
if resize < size:
|
319 |
+
magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
|
320 |
+
magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
|
321 |
+
mel_output = torch.matmul(self.mel_basis, magnitude)
|
322 |
+
if self.is_half == True:
|
323 |
+
mel_output = mel_output.half()
|
324 |
+
log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
|
325 |
+
return log_mel_spec
|
326 |
+
|
327 |
+
|
328 |
+
class RMVPE:
|
329 |
+
def __init__(self, model_path, is_half, device=None):
|
330 |
+
self.resample_kernel = {}
|
331 |
+
model = E2E(4, 1, (2, 2))
|
332 |
+
ckpt = torch.load(model_path, map_location="cpu")
|
333 |
+
model.load_state_dict(ckpt)
|
334 |
+
model.eval()
|
335 |
+
if is_half == True:
|
336 |
+
model = model.half()
|
337 |
+
self.model = model
|
338 |
+
self.resample_kernel = {}
|
339 |
+
self.is_half = is_half
|
340 |
+
if device is None:
|
341 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
342 |
+
self.device = device
|
343 |
+
self.mel_extractor = MelSpectrogram(
|
344 |
+
is_half, 128, 16000, 1024, 160, None, 30, 8000
|
345 |
+
).to(device)
|
346 |
+
self.model = self.model.to(device)
|
347 |
+
cents_mapping = 20 * np.arange(360) + 1997.3794084376191
|
348 |
+
self.cents_mapping = np.pad(cents_mapping, (4, 4)) # 368
|
349 |
+
|
350 |
+
def mel2hidden(self, mel):
|
351 |
+
with torch.no_grad():
|
352 |
+
n_frames = mel.shape[-1]
|
353 |
+
mel = F.pad(
|
354 |
+
mel, (0, 32 * ((n_frames - 1) // 32 + 1) - n_frames), mode="reflect"
|
355 |
+
)
|
356 |
+
hidden = self.model(mel)
|
357 |
+
return hidden[:, :n_frames]
|
358 |
+
|
359 |
+
def decode(self, hidden, thred=0.03):
|
360 |
+
cents_pred = self.to_local_average_cents(hidden, thred=thred)
|
361 |
+
f0 = 10 * (2 ** (cents_pred / 1200))
|
362 |
+
f0[f0 == 10] = 0
|
363 |
+
# f0 = np.array([10 * (2 ** (cent_pred / 1200)) if cent_pred else 0 for cent_pred in cents_pred])
|
364 |
+
return f0
|
365 |
+
|
366 |
+
def infer_from_audio(self, audio, thred=0.03):
|
367 |
+
audio = torch.from_numpy(audio).float().to(self.device).unsqueeze(0)
|
368 |
+
# torch.cuda.synchronize()
|
369 |
+
# t0=ttime()
|
370 |
+
mel = self.mel_extractor(audio, center=True)
|
371 |
+
# torch.cuda.synchronize()
|
372 |
+
# t1=ttime()
|
373 |
+
hidden = self.mel2hidden(mel)
|
374 |
+
# torch.cuda.synchronize()
|
375 |
+
# t2=ttime()
|
376 |
+
hidden = hidden.squeeze(0).cpu().numpy()
|
377 |
+
if self.is_half == True:
|
378 |
+
hidden = hidden.astype("float32")
|
379 |
+
f0 = self.decode(hidden, thred=thred)
|
380 |
+
# torch.cuda.synchronize()
|
381 |
+
# t3=ttime()
|
382 |
+
# print("hmvpe:%s\t%s\t%s\t%s"%(t1-t0,t2-t1,t3-t2,t3-t0))
|
383 |
+
return f0
|
384 |
+
|
385 |
+
def to_local_average_cents(self, salience, thred=0.05):
|
386 |
+
# t0 = ttime()
|
387 |
+
center = np.argmax(salience, axis=1) # 帧长#index
|
388 |
+
salience = np.pad(salience, ((0, 0), (4, 4))) # 帧长,368
|
389 |
+
# t1 = ttime()
|
390 |
+
center += 4
|
391 |
+
todo_salience = []
|
392 |
+
todo_cents_mapping = []
|
393 |
+
starts = center - 4
|
394 |
+
ends = center + 5
|
395 |
+
for idx in range(salience.shape[0]):
|
396 |
+
todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
|
397 |
+
todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
|
398 |
+
# t2 = ttime()
|
399 |
+
todo_salience = np.array(todo_salience) # 帧长,9
|
400 |
+
todo_cents_mapping = np.array(todo_cents_mapping) # 帧长,9
|
401 |
+
product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
|
402 |
+
weight_sum = np.sum(todo_salience, 1) # 帧长
|
403 |
+
devided = product_sum / weight_sum # 帧长
|
404 |
+
# t3 = ttime()
|
405 |
+
maxx = np.max(salience, axis=1) # 帧长
|
406 |
+
devided[maxx <= thred] = 0
|
407 |
+
# t4 = ttime()
|
408 |
+
# print("decode:%s\t%s\t%s\t%s" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
|
409 |
+
return devided
|
src/rvc.py
ADDED
@@ -0,0 +1,151 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from multiprocessing import cpu_count
|
2 |
+
from pathlib import Path
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from fairseq import checkpoint_utils
|
6 |
+
from scipy.io import wavfile
|
7 |
+
|
8 |
+
from infer_pack.models import (
|
9 |
+
SynthesizerTrnMs256NSFsid,
|
10 |
+
SynthesizerTrnMs256NSFsid_nono,
|
11 |
+
SynthesizerTrnMs768NSFsid,
|
12 |
+
SynthesizerTrnMs768NSFsid_nono,
|
13 |
+
)
|
14 |
+
from my_utils import load_audio
|
15 |
+
from vc_infer_pipeline import VC
|
16 |
+
|
17 |
+
BASE_DIR = Path(__file__).resolve().parent.parent
|
18 |
+
|
19 |
+
|
20 |
+
class Config:
|
21 |
+
def __init__(self, device, is_half):
|
22 |
+
self.device = device
|
23 |
+
self.is_half = is_half
|
24 |
+
self.n_cpu = 0
|
25 |
+
self.gpu_name = None
|
26 |
+
self.gpu_mem = None
|
27 |
+
self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
|
28 |
+
|
29 |
+
def device_config(self) -> tuple:
|
30 |
+
if torch.cuda.is_available():
|
31 |
+
i_device = int(self.device.split(":")[-1])
|
32 |
+
self.gpu_name = torch.cuda.get_device_name(i_device)
|
33 |
+
if (
|
34 |
+
("16" in self.gpu_name and "V100" not in self.gpu_name.upper())
|
35 |
+
or "P40" in self.gpu_name.upper()
|
36 |
+
or "1060" in self.gpu_name
|
37 |
+
or "1070" in self.gpu_name
|
38 |
+
or "1080" in self.gpu_name
|
39 |
+
):
|
40 |
+
print("16 series/10 series P40 forced single precision")
|
41 |
+
self.is_half = False
|
42 |
+
for config_file in ["32k.json", "40k.json", "48k.json"]:
|
43 |
+
with open(BASE_DIR / "src" / "configs" / config_file, "r") as f:
|
44 |
+
strr = f.read().replace("true", "false")
|
45 |
+
with open(BASE_DIR / "src" / "configs" / config_file, "w") as f:
|
46 |
+
f.write(strr)
|
47 |
+
with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "r") as f:
|
48 |
+
strr = f.read().replace("3.7", "3.0")
|
49 |
+
with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "w") as f:
|
50 |
+
f.write(strr)
|
51 |
+
else:
|
52 |
+
self.gpu_name = None
|
53 |
+
self.gpu_mem = int(
|
54 |
+
torch.cuda.get_device_properties(i_device).total_memory
|
55 |
+
/ 1024
|
56 |
+
/ 1024
|
57 |
+
/ 1024
|
58 |
+
+ 0.4
|
59 |
+
)
|
60 |
+
if self.gpu_mem <= 4:
|
61 |
+
with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "r") as f:
|
62 |
+
strr = f.read().replace("3.7", "3.0")
|
63 |
+
with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "w") as f:
|
64 |
+
f.write(strr)
|
65 |
+
elif torch.backends.mps.is_available():
|
66 |
+
print("No supported N-card found, use MPS for inference")
|
67 |
+
self.device = "mps"
|
68 |
+
else:
|
69 |
+
print("No supported N-card found, use CPU for inference")
|
70 |
+
self.device = "cpu"
|
71 |
+
self.is_half = True
|
72 |
+
|
73 |
+
if self.n_cpu == 0:
|
74 |
+
self.n_cpu = cpu_count()
|
75 |
+
|
76 |
+
if self.is_half:
|
77 |
+
# 6G memory config
|
78 |
+
x_pad = 3
|
79 |
+
x_query = 10
|
80 |
+
x_center = 60
|
81 |
+
x_max = 65
|
82 |
+
else:
|
83 |
+
# 5G memory config
|
84 |
+
x_pad = 1
|
85 |
+
x_query = 6
|
86 |
+
x_center = 38
|
87 |
+
x_max = 41
|
88 |
+
|
89 |
+
if self.gpu_mem != None and self.gpu_mem <= 4:
|
90 |
+
x_pad = 1
|
91 |
+
x_query = 5
|
92 |
+
x_center = 30
|
93 |
+
x_max = 32
|
94 |
+
|
95 |
+
return x_pad, x_query, x_center, x_max
|
96 |
+
|
97 |
+
|
98 |
+
def load_hubert(device, is_half, model_path):
|
99 |
+
models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task([model_path], suffix='', )
|
100 |
+
hubert = models[0]
|
101 |
+
hubert = hubert.to(device)
|
102 |
+
|
103 |
+
if is_half:
|
104 |
+
hubert = hubert.half()
|
105 |
+
else:
|
106 |
+
hubert = hubert.float()
|
107 |
+
|
108 |
+
hubert.eval()
|
109 |
+
return hubert
|
110 |
+
|
111 |
+
|
112 |
+
def get_vc(device, is_half, config, model_path):
|
113 |
+
cpt = torch.load(model_path, map_location='cpu')
|
114 |
+
if "config" not in cpt or "weight" not in cpt:
|
115 |
+
raise ValueError(f'Incorrect format for {model_path}. Use a voice model trained using RVC v2 instead.')
|
116 |
+
|
117 |
+
tgt_sr = cpt["config"][-1]
|
118 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
|
119 |
+
if_f0 = cpt.get("f0", 1)
|
120 |
+
version = cpt.get("version", "v1")
|
121 |
+
|
122 |
+
if version == "v1":
|
123 |
+
if if_f0 == 1:
|
124 |
+
net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=is_half)
|
125 |
+
else:
|
126 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
127 |
+
elif version == "v2":
|
128 |
+
if if_f0 == 1:
|
129 |
+
net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=is_half)
|
130 |
+
else:
|
131 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
132 |
+
|
133 |
+
del net_g.enc_q
|
134 |
+
print(net_g.load_state_dict(cpt["weight"], strict=False))
|
135 |
+
net_g.eval().to(device)
|
136 |
+
|
137 |
+
if is_half:
|
138 |
+
net_g = net_g.half()
|
139 |
+
else:
|
140 |
+
net_g = net_g.float()
|
141 |
+
|
142 |
+
vc = VC(tgt_sr, config)
|
143 |
+
return cpt, version, net_g, tgt_sr, vc
|
144 |
+
|
145 |
+
|
146 |
+
def rvc_infer(index_path, index_rate, input_path, output_path, pitch_change, f0_method, cpt, version, net_g, filter_radius, tgt_sr, rms_mix_rate, protect, crepe_hop_length, vc, hubert_model):
|
147 |
+
audio = load_audio(input_path, 16000)
|
148 |
+
times = [0, 0, 0]
|
149 |
+
if_f0 = cpt.get('f0', 1)
|
150 |
+
audio_opt = vc.pipeline(hubert_model, net_g, 0, audio, input_path, times, pitch_change, f0_method, index_path, index_rate, if_f0, filter_radius, tgt_sr, 0, rms_mix_rate, version, protect, crepe_hop_length)
|
151 |
+
wavfile.write(output_path, tgt_sr, audio_opt)
|