Upload 38 files
Browse files- Dockerfile +13 -0
- MDXNet.py +274 -0
- RVC.py +1318 -0
- RVC_class.py +85 -0
- RVC_functions.py +76 -0
- Retrieval_based_Voice_Conversion_WebUI.ipynb +381 -0
- Retrieval_based_Voice_Conversion_WebUI_v2.ipynb +401 -0
- app.py +322 -0
- config.py +123 -0
- envfilescheck.bat +348 -0
- export_onnx.py +54 -0
- extract_f0_print.py +160 -0
- extract_feature_print.py +123 -0
- extract_locale.py +31 -0
- ffmpeg +0 -0
- go-realtime-gui.bat +2 -0
- go-web.bat +2 -0
- gui.py +698 -0
- host.json +15 -0
- hubert_base.pt +3 -0
- i18n.py +28 -0
- infer-web.py +1999 -0
- infer_batch_rvc.py +220 -0
- infer_uvr5.py +364 -0
- local.settings.json +8 -0
- main.py +41 -0
- my_utils.py +35 -0
- onnx_inference_demo.py +20 -0
- poetry.lock +0 -0
- pyproject.toml +61 -0
- requirements-win-for-realtime_vc_gui.txt +28 -0
- requirements.txt +55 -0
- slicer2.py +260 -0
- temp.py +1392 -0
- test.py +5 -0
- train_nsf_sim_cache_sid_load_pretrain.py +595 -0
- trainset_preprocess_pipeline_print.py +139 -0
- vc_infer_pipeline.py +449 -0
Dockerfile
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# syntax=docker/dockerfile:1
|
2 |
+
|
3 |
+
FROM python:3.10-bullseye
|
4 |
+
|
5 |
+
EXPOSE 7865
|
6 |
+
|
7 |
+
WORKDIR /app
|
8 |
+
|
9 |
+
COPY . .
|
10 |
+
|
11 |
+
RUN pip3 install -r requirements.txt
|
12 |
+
|
13 |
+
CMD ["python3", "infer-web.py"]
|
MDXNet.py
ADDED
@@ -0,0 +1,274 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import soundfile as sf
|
2 |
+
import torch, pdb, time, argparse, os, warnings, sys, librosa
|
3 |
+
import numpy as np
|
4 |
+
import onnxruntime as ort
|
5 |
+
from scipy.io.wavfile import write
|
6 |
+
from tqdm import tqdm
|
7 |
+
import torch
|
8 |
+
import torch.nn as nn
|
9 |
+
|
10 |
+
dim_c = 4
|
11 |
+
|
12 |
+
|
13 |
+
class Conv_TDF_net_trim:
|
14 |
+
def __init__(
|
15 |
+
self, device, model_name, target_name, L, dim_f, dim_t, n_fft, hop=1024
|
16 |
+
):
|
17 |
+
super(Conv_TDF_net_trim, self).__init__()
|
18 |
+
|
19 |
+
self.dim_f = dim_f
|
20 |
+
self.dim_t = 2**dim_t
|
21 |
+
self.n_fft = n_fft
|
22 |
+
self.hop = hop
|
23 |
+
self.n_bins = self.n_fft // 2 + 1
|
24 |
+
self.chunk_size = hop * (self.dim_t - 1)
|
25 |
+
self.window = torch.hann_window(window_length=self.n_fft, periodic=True).to(
|
26 |
+
device
|
27 |
+
)
|
28 |
+
self.target_name = target_name
|
29 |
+
self.blender = "blender" in model_name
|
30 |
+
|
31 |
+
out_c = dim_c * 4 if target_name == "*" else dim_c
|
32 |
+
self.freq_pad = torch.zeros(
|
33 |
+
[1, out_c, self.n_bins - self.dim_f, self.dim_t]
|
34 |
+
).to(device)
|
35 |
+
|
36 |
+
self.n = L // 2
|
37 |
+
|
38 |
+
def stft(self, x):
|
39 |
+
x = x.reshape([-1, self.chunk_size])
|
40 |
+
x = torch.stft(
|
41 |
+
x,
|
42 |
+
n_fft=self.n_fft,
|
43 |
+
hop_length=self.hop,
|
44 |
+
window=self.window,
|
45 |
+
center=True,
|
46 |
+
return_complex=True,
|
47 |
+
)
|
48 |
+
x = torch.view_as_real(x)
|
49 |
+
x = x.permute([0, 3, 1, 2])
|
50 |
+
x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape(
|
51 |
+
[-1, dim_c, self.n_bins, self.dim_t]
|
52 |
+
)
|
53 |
+
return x[:, :, : self.dim_f]
|
54 |
+
|
55 |
+
def istft(self, x, freq_pad=None):
|
56 |
+
freq_pad = (
|
57 |
+
self.freq_pad.repeat([x.shape[0], 1, 1, 1])
|
58 |
+
if freq_pad is None
|
59 |
+
else freq_pad
|
60 |
+
)
|
61 |
+
x = torch.cat([x, freq_pad], -2)
|
62 |
+
c = 4 * 2 if self.target_name == "*" else 2
|
63 |
+
x = x.reshape([-1, c, 2, self.n_bins, self.dim_t]).reshape(
|
64 |
+
[-1, 2, self.n_bins, self.dim_t]
|
65 |
+
)
|
66 |
+
x = x.permute([0, 2, 3, 1])
|
67 |
+
x = x.contiguous()
|
68 |
+
x = torch.view_as_complex(x)
|
69 |
+
x = torch.istft(
|
70 |
+
x, n_fft=self.n_fft, hop_length=self.hop, window=self.window, center=True
|
71 |
+
)
|
72 |
+
return x.reshape([-1, c, self.chunk_size])
|
73 |
+
|
74 |
+
|
75 |
+
def get_models(device, dim_f, dim_t, n_fft):
|
76 |
+
return Conv_TDF_net_trim(
|
77 |
+
device=device,
|
78 |
+
model_name="Conv-TDF",
|
79 |
+
target_name="vocals",
|
80 |
+
L=11,
|
81 |
+
dim_f=dim_f,
|
82 |
+
dim_t=dim_t,
|
83 |
+
n_fft=n_fft,
|
84 |
+
)
|
85 |
+
|
86 |
+
|
87 |
+
warnings.filterwarnings("ignore")
|
88 |
+
cpu = torch.device("cpu")
|
89 |
+
if torch.cuda.is_available():
|
90 |
+
device = torch.device("cuda:0")
|
91 |
+
elif torch.backends.mps.is_available():
|
92 |
+
device = torch.device("mps")
|
93 |
+
else:
|
94 |
+
device = torch.device("cpu")
|
95 |
+
|
96 |
+
|
97 |
+
class Predictor:
|
98 |
+
def __init__(self, args):
|
99 |
+
self.args = args
|
100 |
+
self.model_ = get_models(
|
101 |
+
device=cpu, dim_f=args.dim_f, dim_t=args.dim_t, n_fft=args.n_fft
|
102 |
+
)
|
103 |
+
self.model = ort.InferenceSession(
|
104 |
+
os.path.join(args.onnx, self.model_.target_name + ".onnx"),
|
105 |
+
providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
|
106 |
+
)
|
107 |
+
print("onnx load done")
|
108 |
+
|
109 |
+
def demix(self, mix):
|
110 |
+
samples = mix.shape[-1]
|
111 |
+
margin = self.args.margin
|
112 |
+
chunk_size = self.args.chunks * 44100
|
113 |
+
assert not margin == 0, "margin cannot be zero!"
|
114 |
+
if margin > chunk_size:
|
115 |
+
margin = chunk_size
|
116 |
+
|
117 |
+
segmented_mix = {}
|
118 |
+
|
119 |
+
if self.args.chunks == 0 or samples < chunk_size:
|
120 |
+
chunk_size = samples
|
121 |
+
|
122 |
+
counter = -1
|
123 |
+
for skip in range(0, samples, chunk_size):
|
124 |
+
counter += 1
|
125 |
+
|
126 |
+
s_margin = 0 if counter == 0 else margin
|
127 |
+
end = min(skip + chunk_size + margin, samples)
|
128 |
+
|
129 |
+
start = skip - s_margin
|
130 |
+
|
131 |
+
segmented_mix[skip] = mix[:, start:end].copy()
|
132 |
+
if end == samples:
|
133 |
+
break
|
134 |
+
|
135 |
+
sources = self.demix_base(segmented_mix, margin_size=margin)
|
136 |
+
"""
|
137 |
+
mix:(2,big_sample)
|
138 |
+
segmented_mix:offset->(2,small_sample)
|
139 |
+
sources:(1,2,big_sample)
|
140 |
+
"""
|
141 |
+
return sources
|
142 |
+
|
143 |
+
def demix_base(self, mixes, margin_size):
|
144 |
+
chunked_sources = []
|
145 |
+
progress_bar = tqdm(total=len(mixes))
|
146 |
+
progress_bar.set_description("Processing")
|
147 |
+
for mix in mixes:
|
148 |
+
cmix = mixes[mix]
|
149 |
+
sources = []
|
150 |
+
n_sample = cmix.shape[1]
|
151 |
+
model = self.model_
|
152 |
+
trim = model.n_fft // 2
|
153 |
+
gen_size = model.chunk_size - 2 * trim
|
154 |
+
pad = gen_size - n_sample % gen_size
|
155 |
+
mix_p = np.concatenate(
|
156 |
+
(np.zeros((2, trim)), cmix, np.zeros((2, pad)), np.zeros((2, trim))), 1
|
157 |
+
)
|
158 |
+
mix_waves = []
|
159 |
+
i = 0
|
160 |
+
while i < n_sample + pad:
|
161 |
+
waves = np.array(mix_p[:, i : i + model.chunk_size])
|
162 |
+
mix_waves.append(waves)
|
163 |
+
i += gen_size
|
164 |
+
mix_waves = torch.tensor(mix_waves, dtype=torch.float32).to(cpu)
|
165 |
+
with torch.no_grad():
|
166 |
+
_ort = self.model
|
167 |
+
spek = model.stft(mix_waves)
|
168 |
+
if self.args.denoise:
|
169 |
+
spec_pred = (
|
170 |
+
-_ort.run(None, {"input": -spek.cpu().numpy()})[0] * 0.5
|
171 |
+
+ _ort.run(None, {"input": spek.cpu().numpy()})[0] * 0.5
|
172 |
+
)
|
173 |
+
tar_waves = model.istft(torch.tensor(spec_pred))
|
174 |
+
else:
|
175 |
+
tar_waves = model.istft(
|
176 |
+
torch.tensor(_ort.run(None, {"input": spek.cpu().numpy()})[0])
|
177 |
+
)
|
178 |
+
tar_signal = (
|
179 |
+
tar_waves[:, :, trim:-trim]
|
180 |
+
.transpose(0, 1)
|
181 |
+
.reshape(2, -1)
|
182 |
+
.numpy()[:, :-pad]
|
183 |
+
)
|
184 |
+
|
185 |
+
start = 0 if mix == 0 else margin_size
|
186 |
+
end = None if mix == list(mixes.keys())[::-1][0] else -margin_size
|
187 |
+
if margin_size == 0:
|
188 |
+
end = None
|
189 |
+
sources.append(tar_signal[:, start:end])
|
190 |
+
|
191 |
+
progress_bar.update(1)
|
192 |
+
|
193 |
+
chunked_sources.append(sources)
|
194 |
+
_sources = np.concatenate(chunked_sources, axis=-1)
|
195 |
+
# del self.model
|
196 |
+
progress_bar.close()
|
197 |
+
return _sources
|
198 |
+
|
199 |
+
def prediction(self, m, vocal_root, others_root, format):
|
200 |
+
os.makedirs(vocal_root, exist_ok=True)
|
201 |
+
os.makedirs(others_root, exist_ok=True)
|
202 |
+
basename = os.path.basename(m)
|
203 |
+
mix, rate = librosa.load(m, mono=False, sr=44100)
|
204 |
+
if mix.ndim == 1:
|
205 |
+
mix = np.asfortranarray([mix, mix])
|
206 |
+
mix = mix.T
|
207 |
+
sources = self.demix(mix.T)
|
208 |
+
opt = sources[0].T
|
209 |
+
if format in ["wav", "flac"]:
|
210 |
+
sf.write(
|
211 |
+
"%s/%s_main_vocal.%s" % (vocal_root, basename, format), mix - opt, rate
|
212 |
+
)
|
213 |
+
sf.write("%s/%s_others.%s" % (others_root, basename, format), opt, rate)
|
214 |
+
else:
|
215 |
+
path_vocal = "%s/%s_main_vocal.wav" % (vocal_root, basename)
|
216 |
+
path_other = "%s/%s_others.wav" % (others_root, basename)
|
217 |
+
sf.write(path_vocal, mix - opt, rate)
|
218 |
+
sf.write(path_other, opt, rate)
|
219 |
+
if os.path.exists(path_vocal):
|
220 |
+
os.system(
|
221 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
222 |
+
% (path_vocal, path_vocal[:-4] + ".%s" % format)
|
223 |
+
)
|
224 |
+
if os.path.exists(path_other):
|
225 |
+
os.system(
|
226 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
227 |
+
% (path_other, path_other[:-4] + ".%s" % format)
|
228 |
+
)
|
229 |
+
|
230 |
+
|
231 |
+
class MDXNetDereverb:
|
232 |
+
def __init__(self, chunks):
|
233 |
+
self.onnx = "uvr5_weights/onnx_dereverb_By_FoxJoy"
|
234 |
+
self.shifts = 10 #'Predict with randomised equivariant stabilisation'
|
235 |
+
self.mixing = "min_mag" # ['default','min_mag','max_mag']
|
236 |
+
self.chunks = chunks
|
237 |
+
self.margin = 44100
|
238 |
+
self.dim_t = 9
|
239 |
+
self.dim_f = 3072
|
240 |
+
self.n_fft = 6144
|
241 |
+
self.denoise = True
|
242 |
+
self.pred = Predictor(self)
|
243 |
+
|
244 |
+
def _path_audio_(self, input, vocal_root, others_root, format):
|
245 |
+
self.pred.prediction(input, vocal_root, others_root, format)
|
246 |
+
|
247 |
+
|
248 |
+
if __name__ == "__main__":
|
249 |
+
dereverb = MDXNetDereverb(15)
|
250 |
+
from time import time as ttime
|
251 |
+
|
252 |
+
t0 = ttime()
|
253 |
+
dereverb._path_audio_(
|
254 |
+
"雪雪伴奏对消HP5.wav",
|
255 |
+
"vocal",
|
256 |
+
"others",
|
257 |
+
)
|
258 |
+
t1 = ttime()
|
259 |
+
print(t1 - t0)
|
260 |
+
|
261 |
+
|
262 |
+
"""
|
263 |
+
|
264 |
+
runtime\python.exe MDXNet.py
|
265 |
+
|
266 |
+
6G:
|
267 |
+
15/9:0.8G->6.8G
|
268 |
+
14:0.8G->6.5G
|
269 |
+
25:炸
|
270 |
+
|
271 |
+
half15:0.7G->6.6G,22.69s
|
272 |
+
fp32-15:0.7G->6.6G,20.85s
|
273 |
+
|
274 |
+
"""
|
RVC.py
ADDED
@@ -0,0 +1,1318 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import shutil
|
3 |
+
import requests
|
4 |
+
import sys
|
5 |
+
import tempfile
|
6 |
+
import soundfile as sf
|
7 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = ""
|
8 |
+
now_dir = os.getcwd()
|
9 |
+
sys.path.append(now_dir)
|
10 |
+
import traceback, pdb
|
11 |
+
import warnings
|
12 |
+
|
13 |
+
import numpy as np
|
14 |
+
import torch
|
15 |
+
from pydub import AudioSegment
|
16 |
+
os.environ['OPENBLAS_NUM_THREADS'] = '1'
|
17 |
+
os.environ["no_proxy"] = "localhost, 127.0.0.1, ::1"
|
18 |
+
import logging
|
19 |
+
import threading
|
20 |
+
from random import shuffle
|
21 |
+
from subprocess import Popen
|
22 |
+
from time import sleep
|
23 |
+
|
24 |
+
import faiss
|
25 |
+
import ffmpeg
|
26 |
+
import gradio as gr
|
27 |
+
import soundfile as sf
|
28 |
+
from config import Config
|
29 |
+
from fairseq import checkpoint_utils
|
30 |
+
from i18n import I18nAuto
|
31 |
+
from infer_pack.models import (
|
32 |
+
SynthesizerTrnMs256NSFsid,
|
33 |
+
SynthesizerTrnMs256NSFsid_nono,
|
34 |
+
SynthesizerTrnMs768NSFsid,
|
35 |
+
SynthesizerTrnMs768NSFsid_nono,
|
36 |
+
)
|
37 |
+
from infer_pack.models_onnx import SynthesizerTrnMsNSFsidM
|
38 |
+
from infer_uvr5 import _audio_pre_, _audio_pre_new
|
39 |
+
from MDXNet import MDXNetDereverb
|
40 |
+
from my_utils import load_audio
|
41 |
+
from train.process_ckpt import change_info, extract_small_model, merge, show_info
|
42 |
+
from vc_infer_pipeline import VC
|
43 |
+
from sklearn.cluster import MiniBatchKMeans
|
44 |
+
|
45 |
+
logging.getLogger("numba").setLevel(logging.WARNING)
|
46 |
+
|
47 |
+
|
48 |
+
tmp = os.path.join(now_dir, "TEMP")
|
49 |
+
shutil.rmtree(tmp, ignore_errors=True)
|
50 |
+
shutil.rmtree("%s/runtime/Lib/site-packages/infer_pack" % (now_dir), ignore_errors=True)
|
51 |
+
shutil.rmtree("%s/runtime/Lib/site-packages/uvr5_pack" % (now_dir), ignore_errors=True)
|
52 |
+
os.makedirs(tmp, exist_ok=True)
|
53 |
+
os.makedirs(os.path.join(now_dir, "logs"), exist_ok=True)
|
54 |
+
os.makedirs(os.path.join(now_dir, "weights"), exist_ok=True)
|
55 |
+
os.environ["TEMP"] = tmp
|
56 |
+
warnings.filterwarnings("ignore")
|
57 |
+
torch.manual_seed(114514)
|
58 |
+
|
59 |
+
|
60 |
+
config = Config()
|
61 |
+
i18n = I18nAuto()
|
62 |
+
i18n.print()
|
63 |
+
# 判断是否有能用来训练和加速推理的N卡
|
64 |
+
ngpu = torch.cuda.device_count()
|
65 |
+
gpu_infos = []
|
66 |
+
mem = []
|
67 |
+
if_gpu_ok = False
|
68 |
+
|
69 |
+
if torch.cuda.is_available() or ngpu != 0:
|
70 |
+
for i in range(ngpu):
|
71 |
+
gpu_name = torch.cuda.get_device_name(i)
|
72 |
+
if any(
|
73 |
+
value in gpu_name.upper()
|
74 |
+
for value in [
|
75 |
+
"10",
|
76 |
+
"16",
|
77 |
+
"20",
|
78 |
+
"30",
|
79 |
+
"40",
|
80 |
+
"A2",
|
81 |
+
"A3",
|
82 |
+
"A4",
|
83 |
+
"P4",
|
84 |
+
"A50",
|
85 |
+
"500",
|
86 |
+
"A60",
|
87 |
+
"70",
|
88 |
+
"80",
|
89 |
+
"90",
|
90 |
+
"M4",
|
91 |
+
"T4",
|
92 |
+
"TITAN",
|
93 |
+
]
|
94 |
+
):
|
95 |
+
# A10#A100#V100#A40#P40#M40#K80#A4500
|
96 |
+
if_gpu_ok = True # 至少有一张能用的N卡
|
97 |
+
gpu_infos.append("%s\t%s" % (i, gpu_name))
|
98 |
+
mem.append(
|
99 |
+
int(
|
100 |
+
torch.cuda.get_device_properties(i).total_memory
|
101 |
+
/ 1024
|
102 |
+
/ 1024
|
103 |
+
/ 1024
|
104 |
+
+ 0.4
|
105 |
+
)
|
106 |
+
)
|
107 |
+
if if_gpu_ok and len(gpu_infos) > 0:
|
108 |
+
gpu_info = "\n".join(gpu_infos)
|
109 |
+
default_batch_size = 1
|
110 |
+
else:
|
111 |
+
gpu_info = i18n("很遗憾您这没有能用的显卡来支持您训练")
|
112 |
+
default_batch_size = 1
|
113 |
+
gpus = "-".join([i[0] for i in gpu_infos])
|
114 |
+
|
115 |
+
|
116 |
+
class ToolButton(gr.Button, gr.components.FormComponent):
|
117 |
+
"""Small button with single emoji as text, fits inside gradio forms"""
|
118 |
+
|
119 |
+
def __init__(self, **kwargs):
|
120 |
+
super().__init__(variant="tool", **kwargs)
|
121 |
+
|
122 |
+
def get_block_name(self):
|
123 |
+
return "button"
|
124 |
+
|
125 |
+
|
126 |
+
hubert_model = None
|
127 |
+
|
128 |
+
|
129 |
+
def load_hubert():
|
130 |
+
global hubert_model
|
131 |
+
models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
|
132 |
+
["hubert_base.pt"],
|
133 |
+
suffix="",
|
134 |
+
)
|
135 |
+
hubert_model = models[0]
|
136 |
+
hubert_model = hubert_model.to(config.device)
|
137 |
+
if config.is_half:
|
138 |
+
hubert_model = hubert_model.half()
|
139 |
+
else:
|
140 |
+
hubert_model = hubert_model.float()
|
141 |
+
hubert_model.eval()
|
142 |
+
|
143 |
+
|
144 |
+
weight_root = "weights"
|
145 |
+
weight_uvr5_root = "uvr5_weights"
|
146 |
+
index_root = "logs"
|
147 |
+
names = []
|
148 |
+
for name in os.listdir(weight_root):
|
149 |
+
if name.endswith(".pth"):
|
150 |
+
names.append(name)
|
151 |
+
index_paths = []
|
152 |
+
for root, dirs, files in os.walk(index_root, topdown=False):
|
153 |
+
for name in files:
|
154 |
+
if name.endswith(".index") and "trained" not in name:
|
155 |
+
index_paths.append("%s/%s" % (root, name))
|
156 |
+
uvr5_names = []
|
157 |
+
for name in os.listdir(weight_uvr5_root):
|
158 |
+
if name.endswith(".pth") or "onnx" in name:
|
159 |
+
uvr5_names.append(name.replace(".pth", ""))
|
160 |
+
|
161 |
+
|
162 |
+
def vc_single(
|
163 |
+
sid,
|
164 |
+
input_audio_path,
|
165 |
+
f0_up_key,
|
166 |
+
f0_file,
|
167 |
+
f0_method,
|
168 |
+
file_index,
|
169 |
+
file_index2,
|
170 |
+
# file_big_npy,
|
171 |
+
index_rate,
|
172 |
+
filter_radius,
|
173 |
+
resample_sr,
|
174 |
+
rms_mix_rate,
|
175 |
+
protect,
|
176 |
+
song
|
177 |
+
): # spk_item, input_audio0, vc_transform0,f0_file,f0method0
|
178 |
+
global tgt_sr, net_g, vc, hubert_model, version
|
179 |
+
logging.info(f0_up_key)
|
180 |
+
|
181 |
+
if input_audio_path is None:
|
182 |
+
return "You need to upload an audio", None
|
183 |
+
|
184 |
+
f0_up_key = int(f0_up_key)
|
185 |
+
try:
|
186 |
+
# logging.info("before response")
|
187 |
+
# response = requests.get(input_audio_path, stream=True)
|
188 |
+
# response.raise_for_status()
|
189 |
+
# logging.info("after response")
|
190 |
+
# temp_file = tempfile.NamedTemporaryFile(suffix='.wav', delete=False)
|
191 |
+
# path = temp_file.name
|
192 |
+
# for chunk in response.iter_content(chunk_size=1024):
|
193 |
+
# temp_file.write(chunk)
|
194 |
+
|
195 |
+
# with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as temp_file:
|
196 |
+
# path = temp_file.name
|
197 |
+
# logging.info("Temporary file path: %s", path)
|
198 |
+
# temp_file.write(response.content)
|
199 |
+
|
200 |
+
# logging.info("tempfilewrite")
|
201 |
+
|
202 |
+
|
203 |
+
|
204 |
+
if not input_audio_path.lower().endswith(".wav"):
|
205 |
+
# Download the audio file
|
206 |
+
response = requests.get(input_audio_path, stream=True)
|
207 |
+
response.raise_for_status()
|
208 |
+
logging.info("after response")
|
209 |
+
|
210 |
+
# Create a temporary file and save the downloaded audio content
|
211 |
+
with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as temp_file:
|
212 |
+
path = temp_file.name
|
213 |
+
logging.info("Temporary file path: %s", path)
|
214 |
+
temp_file.write(response.content)
|
215 |
+
|
216 |
+
# Convert the audio file to WAV format using pydub
|
217 |
+
audio = AudioSegment.from_file(path)
|
218 |
+
output_file_path = os.path.splitext(path)[0] + ".wav"
|
219 |
+
audio.export(output_file_path, format="wav")
|
220 |
+
|
221 |
+
# Use the converted WAV file as the new temporary file
|
222 |
+
path = output_file_path
|
223 |
+
|
224 |
+
else:
|
225 |
+
# The input audio URL is already pointing to a WAV file, so use it as-is
|
226 |
+
response = requests.get(input_audio_path)
|
227 |
+
response.raise_for_status()
|
228 |
+
|
229 |
+
# Create a temporary file and save the downloaded audio content
|
230 |
+
with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as temp_file:
|
231 |
+
path = temp_file.name
|
232 |
+
logging.info("Temporary file path: %s", path)
|
233 |
+
temp_file.write(response.content)
|
234 |
+
|
235 |
+
if song == True:
|
236 |
+
uvr(path)
|
237 |
+
path = f"{os.getcwd()}/opt/vocal.wav"
|
238 |
+
|
239 |
+
|
240 |
+
|
241 |
+
audio = load_audio(path, 16000)
|
242 |
+
temp_file.close()
|
243 |
+
audio_max = np.abs(audio).max() / 0.95
|
244 |
+
if audio_max > 1:
|
245 |
+
audio = audio.astype(float)
|
246 |
+
audio /= audio_max
|
247 |
+
times = [0, 0, 0]
|
248 |
+
if not hubert_model:
|
249 |
+
load_hubert()
|
250 |
+
if_f0 = cpt.get("f0", 1)
|
251 |
+
file_index = (
|
252 |
+
(
|
253 |
+
file_index.strip(" ")
|
254 |
+
.strip('"')
|
255 |
+
.strip("\n")
|
256 |
+
.strip('"')
|
257 |
+
.strip(" ")
|
258 |
+
.replace("trained", "added")
|
259 |
+
)
|
260 |
+
if file_index != ""
|
261 |
+
else file_index2
|
262 |
+
) # 防止小白写错,自动帮他替换掉
|
263 |
+
# file_big_npy = (
|
264 |
+
# file_big_npy.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
265 |
+
# )
|
266 |
+
logging.info("before vc pipeline")
|
267 |
+
logging.info(vc)
|
268 |
+
audio_opt = vc.pipeline(
|
269 |
+
hubert_model,
|
270 |
+
net_g,
|
271 |
+
sid,
|
272 |
+
audio,
|
273 |
+
input_audio_path,
|
274 |
+
times,
|
275 |
+
f0_up_key,
|
276 |
+
f0_method,
|
277 |
+
file_index,
|
278 |
+
# file_big_npy,
|
279 |
+
index_rate,
|
280 |
+
if_f0,
|
281 |
+
filter_radius,
|
282 |
+
tgt_sr,
|
283 |
+
resample_sr,
|
284 |
+
rms_mix_rate,
|
285 |
+
version,
|
286 |
+
protect,
|
287 |
+
f0_file=f0_file,
|
288 |
+
)
|
289 |
+
logging.info("after vc pipeline")
|
290 |
+
|
291 |
+
logging.info(f0_up_key)
|
292 |
+
|
293 |
+
if tgt_sr != resample_sr >= 16000:
|
294 |
+
tgt_sr = resample_sr
|
295 |
+
index_info = (
|
296 |
+
"Using index:%s." % file_index
|
297 |
+
if os.path.exists(file_index)
|
298 |
+
else "Index not used."
|
299 |
+
)
|
300 |
+
return "Success.\n %s\nTime:\n npy:%ss, f0:%ss, infer:%ss" % (
|
301 |
+
index_info,
|
302 |
+
times[0],
|
303 |
+
times[1],
|
304 |
+
times[2],
|
305 |
+
), (tgt_sr, audio_opt)
|
306 |
+
except:
|
307 |
+
info = traceback.format_exc()
|
308 |
+
logging.info(info)
|
309 |
+
return info, (None, None)
|
310 |
+
|
311 |
+
|
312 |
+
def vc_multi(
|
313 |
+
sid,
|
314 |
+
dir_path,
|
315 |
+
opt_root,
|
316 |
+
paths,
|
317 |
+
f0_up_key,
|
318 |
+
f0_method,
|
319 |
+
file_index,
|
320 |
+
file_index2,
|
321 |
+
# file_big_npy,
|
322 |
+
index_rate,
|
323 |
+
filter_radius,
|
324 |
+
resample_sr,
|
325 |
+
rms_mix_rate,
|
326 |
+
protect,
|
327 |
+
format1,
|
328 |
+
):
|
329 |
+
|
330 |
+
try:
|
331 |
+
|
332 |
+
dir_path = (
|
333 |
+
dir_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
334 |
+
) # 防止小白拷路径头尾带了空格和"和回车
|
335 |
+
opt_root = opt_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
336 |
+
os.makedirs(opt_root, exist_ok=True)
|
337 |
+
try:
|
338 |
+
if dir_path != "":
|
339 |
+
paths = [os.path.join(dir_path, name) for name in os.listdir(dir_path)]
|
340 |
+
else:
|
341 |
+
paths = [path.name for path in paths]
|
342 |
+
except:
|
343 |
+
traceback.print_exc()
|
344 |
+
paths = [path.name for path in paths]
|
345 |
+
infos = []
|
346 |
+
for path in paths:
|
347 |
+
info, opt = vc_single(
|
348 |
+
sid,
|
349 |
+
path,
|
350 |
+
f0_up_key,
|
351 |
+
None,
|
352 |
+
f0_method,
|
353 |
+
file_index,
|
354 |
+
file_index2,
|
355 |
+
# file_big_npy,
|
356 |
+
index_rate,
|
357 |
+
filter_radius,
|
358 |
+
resample_sr,
|
359 |
+
rms_mix_rate,
|
360 |
+
protect,
|
361 |
+
)
|
362 |
+
if "Success" in info:
|
363 |
+
try:
|
364 |
+
tgt_sr, audio_opt = opt
|
365 |
+
if format1 in ["wav", "flac"]:
|
366 |
+
sf.write(
|
367 |
+
"%s/%s.%s" % (opt_root, os.path.basename(path), format1),
|
368 |
+
audio_opt,
|
369 |
+
tgt_sr,
|
370 |
+
)
|
371 |
+
else:
|
372 |
+
path = "%s/%s.wav" % (opt_root, os.path.basename(path))
|
373 |
+
sf.write(
|
374 |
+
path,
|
375 |
+
audio_opt,
|
376 |
+
tgt_sr,
|
377 |
+
)
|
378 |
+
if os.path.exists(path):
|
379 |
+
os.system(
|
380 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
381 |
+
% (path, path[:-4] + ".%s" % format1)
|
382 |
+
)
|
383 |
+
except:
|
384 |
+
info += traceback.format_exc()
|
385 |
+
infos.append("%s->%s" % (os.path.basename(path), info))
|
386 |
+
yield "\n".join(infos)
|
387 |
+
yield "\n".join(infos)
|
388 |
+
except:
|
389 |
+
yield traceback.format_exc()
|
390 |
+
|
391 |
+
|
392 |
+
def uvr(inp_root, model_name=uvr5_names[0], save_root_vocal='opt', paths=None, save_root_ins="opt", agg=10, format0="wav"):
|
393 |
+
try:
|
394 |
+
func = _audio_pre_
|
395 |
+
pre_fun = func(
|
396 |
+
agg=int(agg),
|
397 |
+
model_path=os.path.join(weight_uvr5_root, model_name + ".pth"),
|
398 |
+
device=config.device,
|
399 |
+
is_half=config.is_half,
|
400 |
+
)
|
401 |
+
|
402 |
+
pre_fun._path_audio_(
|
403 |
+
inp_root, save_root_ins, save_root_vocal, format0
|
404 |
+
)
|
405 |
+
|
406 |
+
|
407 |
+
finally:
|
408 |
+
try:
|
409 |
+
del pre_fun.model
|
410 |
+
del pre_fun
|
411 |
+
except:
|
412 |
+
traceback.print_exc()
|
413 |
+
logging.info("clean_empty_cache")
|
414 |
+
if torch.cuda.is_available():
|
415 |
+
torch.cuda.empty_cache()
|
416 |
+
|
417 |
+
|
418 |
+
|
419 |
+
# 一个选项卡全局只能有一个音色
|
420 |
+
def get_vc(sid, to_return_protect0, to_return_protect1):
|
421 |
+
global n_spk, tgt_sr, net_g, vc, cpt, version
|
422 |
+
if sid == "" or sid == []:
|
423 |
+
global hubert_model
|
424 |
+
if hubert_model is not None: # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
|
425 |
+
logging.info("clean_empty_cache")
|
426 |
+
del net_g, n_spk, vc, hubert_model, tgt_sr # ,cpt
|
427 |
+
hubert_model = net_g = n_spk = vc = hubert_model = tgt_sr = None
|
428 |
+
if torch.cuda.is_available():
|
429 |
+
torch.cuda.empty_cache()
|
430 |
+
###楼下不这么折腾清理不干净
|
431 |
+
if_f0 = cpt.get("f0", 1)
|
432 |
+
version = cpt.get("version", "v1")
|
433 |
+
if version == "v1":
|
434 |
+
if if_f0 == 1:
|
435 |
+
net_g = SynthesizerTrnMs256NSFsid(
|
436 |
+
*cpt["config"], is_half=config.is_half
|
437 |
+
)
|
438 |
+
else:
|
439 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
440 |
+
elif version == "v2":
|
441 |
+
if if_f0 == 1:
|
442 |
+
net_g = SynthesizerTrnMs768NSFsid(
|
443 |
+
*cpt["config"], is_half=config.is_half
|
444 |
+
)
|
445 |
+
else:
|
446 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
447 |
+
del net_g, cpt
|
448 |
+
if torch.cuda.is_available():
|
449 |
+
torch.cuda.empty_cache()
|
450 |
+
cpt = None
|
451 |
+
return {"visible": False, "__type__": "update"}
|
452 |
+
person = "%s/%s" % (weight_root, sid)
|
453 |
+
logging.info("loading %s" % person)
|
454 |
+
cpt = torch.load(person, map_location="cpu")
|
455 |
+
tgt_sr = cpt["config"][-1]
|
456 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
|
457 |
+
if_f0 = cpt.get("f0", 1)
|
458 |
+
if if_f0 == 0:
|
459 |
+
to_return_protect0 = to_return_protect1 = {
|
460 |
+
"visible": False,
|
461 |
+
"value": 0.5,
|
462 |
+
"__type__": "update",
|
463 |
+
}
|
464 |
+
else:
|
465 |
+
to_return_protect0 = {
|
466 |
+
"visible": True,
|
467 |
+
"value": to_return_protect0,
|
468 |
+
"__type__": "update",
|
469 |
+
}
|
470 |
+
to_return_protect1 = {
|
471 |
+
"visible": True,
|
472 |
+
"value": to_return_protect1,
|
473 |
+
"__type__": "update",
|
474 |
+
}
|
475 |
+
version = cpt.get("version", "v1")
|
476 |
+
if version == "v1":
|
477 |
+
if if_f0 == 1:
|
478 |
+
net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
|
479 |
+
else:
|
480 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
481 |
+
elif version == "v2":
|
482 |
+
if if_f0 == 1:
|
483 |
+
net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
|
484 |
+
else:
|
485 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
486 |
+
del net_g.enc_q
|
487 |
+
logging.info(net_g.load_state_dict(cpt["weight"], strict=False))
|
488 |
+
net_g.eval().to(config.device)
|
489 |
+
if config.is_half:
|
490 |
+
net_g = net_g.half()
|
491 |
+
else:
|
492 |
+
net_g = net_g.float()
|
493 |
+
vc = VC(tgt_sr, config)
|
494 |
+
n_spk = cpt["config"][-3]
|
495 |
+
return (
|
496 |
+
{"visible": True, "maximum": n_spk, "__type__": "update"},
|
497 |
+
to_return_protect0,
|
498 |
+
to_return_protect1,
|
499 |
+
)
|
500 |
+
|
501 |
+
|
502 |
+
def change_choices():
|
503 |
+
names = []
|
504 |
+
for name in os.listdir(weight_root):
|
505 |
+
if name.endswith(".pth"):
|
506 |
+
names.append(name)
|
507 |
+
index_paths = []
|
508 |
+
for root, dirs, files in os.walk(index_root, topdown=False):
|
509 |
+
for name in files:
|
510 |
+
if name.endswith(".index") and "trained" not in name:
|
511 |
+
index_paths.append("%s/%s" % (root, name))
|
512 |
+
return {"choices": sorted(names), "__type__": "update"}, {
|
513 |
+
"choices": sorted(index_paths),
|
514 |
+
"__type__": "update",
|
515 |
+
}
|
516 |
+
|
517 |
+
|
518 |
+
def clean():
|
519 |
+
return {"value": "", "__type__": "update"}
|
520 |
+
|
521 |
+
|
522 |
+
sr_dict = {
|
523 |
+
"32k": 32000,
|
524 |
+
"40k": 40000,
|
525 |
+
"48k": 48000,
|
526 |
+
}
|
527 |
+
|
528 |
+
|
529 |
+
def if_done(done, p):
|
530 |
+
while 1:
|
531 |
+
if p.poll() is None:
|
532 |
+
sleep(0.5)
|
533 |
+
else:
|
534 |
+
break
|
535 |
+
done[0] = True
|
536 |
+
|
537 |
+
|
538 |
+
def if_done_multi(done, ps):
|
539 |
+
while 1:
|
540 |
+
# poll==None代表进程未结束
|
541 |
+
# 只要有一个进程未结束都不停
|
542 |
+
flag = 1
|
543 |
+
for p in ps:
|
544 |
+
if p.poll() is None:
|
545 |
+
flag = 0
|
546 |
+
sleep(0.5)
|
547 |
+
break
|
548 |
+
if flag == 1:
|
549 |
+
break
|
550 |
+
done[0] = True
|
551 |
+
|
552 |
+
|
553 |
+
def preprocess_dataset(trainset_dir, exp_dir, sr, n_p):
|
554 |
+
sr = sr_dict[sr]
|
555 |
+
os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
|
556 |
+
f = open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "w")
|
557 |
+
f.close()
|
558 |
+
cmd = (
|
559 |
+
config.python_cmd
|
560 |
+
+ " trainset_preprocess_pipeline_print.py %s %s %s %s/logs/%s "
|
561 |
+
% (trainset_dir, sr, n_p, now_dir, exp_dir)
|
562 |
+
+ str(config.noparallel)
|
563 |
+
)
|
564 |
+
logging.info(cmd)
|
565 |
+
p = Popen(cmd, shell=True) # , stdin=PIPE, stdout=PIPE,stderr=PIPE,cwd=now_dir
|
566 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
567 |
+
done = [False]
|
568 |
+
threading.Thread(
|
569 |
+
target=if_done,
|
570 |
+
args=(
|
571 |
+
done,
|
572 |
+
p,
|
573 |
+
),
|
574 |
+
).start()
|
575 |
+
while 1:
|
576 |
+
with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
|
577 |
+
yield (f.read())
|
578 |
+
sleep(1)
|
579 |
+
if done[0]:
|
580 |
+
break
|
581 |
+
with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
|
582 |
+
log = f.read()
|
583 |
+
logging.info(log)
|
584 |
+
yield log
|
585 |
+
|
586 |
+
|
587 |
+
# but2.click(extract_f0,[gpus6,np7,f0method8,if_f0_3,trainset_dir4],[info2])
|
588 |
+
def extract_f0_feature(gpus, n_p, f0method, if_f0, exp_dir, version19):
|
589 |
+
gpus = gpus.split("-")
|
590 |
+
os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
|
591 |
+
f = open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "w")
|
592 |
+
f.close()
|
593 |
+
if if_f0:
|
594 |
+
cmd = config.python_cmd + " extract_f0_print.py %s/logs/%s %s %s" % (
|
595 |
+
now_dir,
|
596 |
+
exp_dir,
|
597 |
+
n_p,
|
598 |
+
f0method,
|
599 |
+
)
|
600 |
+
logging.info(cmd)
|
601 |
+
p = Popen(cmd, shell=True, cwd=now_dir) # , stdin=PIPE, stdout=PIPE,stderr=PIPE
|
602 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
603 |
+
done = [False]
|
604 |
+
threading.Thread(
|
605 |
+
target=if_done,
|
606 |
+
args=(
|
607 |
+
done,
|
608 |
+
p,
|
609 |
+
),
|
610 |
+
).start()
|
611 |
+
while 1:
|
612 |
+
with open(
|
613 |
+
"%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r"
|
614 |
+
) as f:
|
615 |
+
yield (f.read())
|
616 |
+
sleep(1)
|
617 |
+
if done[0]:
|
618 |
+
break
|
619 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
620 |
+
log = f.read()
|
621 |
+
logging.info(log)
|
622 |
+
yield log
|
623 |
+
####对不同part分别开多进程
|
624 |
+
"""
|
625 |
+
n_part=int(sys.argv[1])
|
626 |
+
i_part=int(sys.argv[2])
|
627 |
+
i_gpu=sys.argv[3]
|
628 |
+
exp_dir=sys.argv[4]
|
629 |
+
os.environ["CUDA_VISIBLE_DEVICES"]=str(i_gpu)
|
630 |
+
"""
|
631 |
+
leng = len(gpus)
|
632 |
+
ps = []
|
633 |
+
for idx, n_g in enumerate(gpus):
|
634 |
+
cmd = (
|
635 |
+
config.python_cmd
|
636 |
+
+ " extract_feature_print.py %s %s %s %s %s/logs/%s %s"
|
637 |
+
% (
|
638 |
+
config.device,
|
639 |
+
leng,
|
640 |
+
idx,
|
641 |
+
n_g,
|
642 |
+
now_dir,
|
643 |
+
exp_dir,
|
644 |
+
version19,
|
645 |
+
)
|
646 |
+
)
|
647 |
+
logging.info(cmd)
|
648 |
+
p = Popen(
|
649 |
+
cmd, shell=True, cwd=now_dir
|
650 |
+
) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
|
651 |
+
ps.append(p)
|
652 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
653 |
+
done = [False]
|
654 |
+
threading.Thread(
|
655 |
+
target=if_done_multi,
|
656 |
+
args=(
|
657 |
+
done,
|
658 |
+
ps,
|
659 |
+
),
|
660 |
+
).start()
|
661 |
+
while 1:
|
662 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
663 |
+
yield (f.read())
|
664 |
+
sleep(1)
|
665 |
+
if done[0]:
|
666 |
+
break
|
667 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
668 |
+
log = f.read()
|
669 |
+
logging.info(log)
|
670 |
+
yield log
|
671 |
+
|
672 |
+
|
673 |
+
def change_sr2(sr2, if_f0_3, version19):
|
674 |
+
path_str = "" if version19 == "v1" else "_v2"
|
675 |
+
f0_str = "f0" if if_f0_3 else ""
|
676 |
+
if_pretrained_generator_exist = os.access(
|
677 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
678 |
+
)
|
679 |
+
if_pretrained_discriminator_exist = os.access(
|
680 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
681 |
+
)
|
682 |
+
if not if_pretrained_generator_exist:
|
683 |
+
logging.info(
|
684 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2),
|
685 |
+
"not exist, will not use pretrained model",
|
686 |
+
)
|
687 |
+
if not if_pretrained_discriminator_exist:
|
688 |
+
logging.info(
|
689 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2),
|
690 |
+
"not exist, will not use pretrained model",
|
691 |
+
)
|
692 |
+
return (
|
693 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2)
|
694 |
+
if if_pretrained_generator_exist
|
695 |
+
else "",
|
696 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2)
|
697 |
+
if if_pretrained_discriminator_exist
|
698 |
+
else "",
|
699 |
+
)
|
700 |
+
|
701 |
+
|
702 |
+
def change_version19(sr2, if_f0_3, version19):
|
703 |
+
path_str = "" if version19 == "v1" else "_v2"
|
704 |
+
if sr2 == "32k" and version19 == "v1":
|
705 |
+
sr2 = "40k"
|
706 |
+
to_return_sr2 = (
|
707 |
+
{"choices": ["40k", "48k"], "__type__": "update", "value": sr2}
|
708 |
+
if version19 == "v1"
|
709 |
+
else {"choices": ["40k", "48k", "32k"], "__type__": "update", "value": sr2}
|
710 |
+
)
|
711 |
+
f0_str = "f0" if if_f0_3 else ""
|
712 |
+
if_pretrained_generator_exist = os.access(
|
713 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
714 |
+
)
|
715 |
+
if_pretrained_discriminator_exist = os.access(
|
716 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
717 |
+
)
|
718 |
+
if not if_pretrained_generator_exist:
|
719 |
+
logging.info(
|
720 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2),
|
721 |
+
"not exist, will not use pretrained model",
|
722 |
+
)
|
723 |
+
if not if_pretrained_discriminator_exist:
|
724 |
+
logging.info(
|
725 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2),
|
726 |
+
"not exist, will not use pretrained model",
|
727 |
+
)
|
728 |
+
return (
|
729 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2)
|
730 |
+
if if_pretrained_generator_exist
|
731 |
+
else "",
|
732 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2)
|
733 |
+
if if_pretrained_discriminator_exist
|
734 |
+
else "",
|
735 |
+
to_return_sr2,
|
736 |
+
)
|
737 |
+
|
738 |
+
|
739 |
+
def change_f0(if_f0_3, sr2, version19): # f0method8,pretrained_G14,pretrained_D15
|
740 |
+
path_str = "" if version19 == "v1" else "_v2"
|
741 |
+
if_pretrained_generator_exist = os.access(
|
742 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2), os.F_OK
|
743 |
+
)
|
744 |
+
if_pretrained_discriminator_exist = os.access(
|
745 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2), os.F_OK
|
746 |
+
)
|
747 |
+
if not if_pretrained_generator_exist:
|
748 |
+
logging.info(
|
749 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2),
|
750 |
+
"not exist, will not use pretrained model",
|
751 |
+
)
|
752 |
+
if not if_pretrained_discriminator_exist:
|
753 |
+
logging.info(
|
754 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2),
|
755 |
+
"not exist, will not use pretrained model",
|
756 |
+
)
|
757 |
+
if if_f0_3:
|
758 |
+
return (
|
759 |
+
{"visible": True, "__type__": "update"},
|
760 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2)
|
761 |
+
if if_pretrained_generator_exist
|
762 |
+
else "",
|
763 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2)
|
764 |
+
if if_pretrained_discriminator_exist
|
765 |
+
else "",
|
766 |
+
)
|
767 |
+
return (
|
768 |
+
{"visible": False, "__type__": "update"},
|
769 |
+
("pretrained%s/G%s.pth" % (path_str, sr2))
|
770 |
+
if if_pretrained_generator_exist
|
771 |
+
else "",
|
772 |
+
("pretrained%s/D%s.pth" % (path_str, sr2))
|
773 |
+
if if_pretrained_discriminator_exist
|
774 |
+
else "",
|
775 |
+
)
|
776 |
+
|
777 |
+
|
778 |
+
# but3.click(click_train,[exp_dir1,sr2,if_f0_3,save_epoch10,total_epoch11,batch_size12,if_save_latest13,pretrained_G14,pretrained_D15,gpus16])
|
779 |
+
def click_train(
|
780 |
+
exp_dir1,
|
781 |
+
sr2,
|
782 |
+
if_f0_3,
|
783 |
+
spk_id5,
|
784 |
+
save_epoch10,
|
785 |
+
total_epoch11,
|
786 |
+
batch_size12,
|
787 |
+
if_save_latest13,
|
788 |
+
pretrained_G14,
|
789 |
+
pretrained_D15,
|
790 |
+
gpus16,
|
791 |
+
if_cache_gpu17,
|
792 |
+
if_save_every_weights18,
|
793 |
+
version19,
|
794 |
+
):
|
795 |
+
# 生成filelist
|
796 |
+
exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
797 |
+
os.makedirs(exp_dir, exist_ok=True)
|
798 |
+
gt_wavs_dir = "%s/0_gt_wavs" % (exp_dir)
|
799 |
+
feature_dir = (
|
800 |
+
"%s/3_feature256" % (exp_dir)
|
801 |
+
if version19 == "v1"
|
802 |
+
else "%s/3_feature768" % (exp_dir)
|
803 |
+
)
|
804 |
+
if if_f0_3:
|
805 |
+
f0_dir = "%s/2a_f0" % (exp_dir)
|
806 |
+
f0nsf_dir = "%s/2b-f0nsf" % (exp_dir)
|
807 |
+
names = (
|
808 |
+
set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)])
|
809 |
+
& set([name.split(".")[0] for name in os.listdir(feature_dir)])
|
810 |
+
& set([name.split(".")[0] for name in os.listdir(f0_dir)])
|
811 |
+
& set([name.split(".")[0] for name in os.listdir(f0nsf_dir)])
|
812 |
+
)
|
813 |
+
else:
|
814 |
+
names = set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)]) & set(
|
815 |
+
[name.split(".")[0] for name in os.listdir(feature_dir)]
|
816 |
+
)
|
817 |
+
opt = []
|
818 |
+
for name in names:
|
819 |
+
if if_f0_3:
|
820 |
+
opt.append(
|
821 |
+
"%s/%s.wav|%s/%s.npy|%s/%s.wav.npy|%s/%s.wav.npy|%s"
|
822 |
+
% (
|
823 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
824 |
+
name,
|
825 |
+
feature_dir.replace("\\", "\\\\"),
|
826 |
+
name,
|
827 |
+
f0_dir.replace("\\", "\\\\"),
|
828 |
+
name,
|
829 |
+
f0nsf_dir.replace("\\", "\\\\"),
|
830 |
+
name,
|
831 |
+
spk_id5,
|
832 |
+
)
|
833 |
+
)
|
834 |
+
else:
|
835 |
+
opt.append(
|
836 |
+
"%s/%s.wav|%s/%s.npy|%s"
|
837 |
+
% (
|
838 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
839 |
+
name,
|
840 |
+
feature_dir.replace("\\", "\\\\"),
|
841 |
+
name,
|
842 |
+
spk_id5,
|
843 |
+
)
|
844 |
+
)
|
845 |
+
fea_dim = 256 if version19 == "v1" else 768
|
846 |
+
if if_f0_3:
|
847 |
+
for _ in range(2):
|
848 |
+
opt.append(
|
849 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s/logs/mute/2a_f0/mute.wav.npy|%s/logs/mute/2b-f0nsf/mute.wav.npy|%s"
|
850 |
+
% (now_dir, sr2, now_dir, fea_dim, now_dir, now_dir, spk_id5)
|
851 |
+
)
|
852 |
+
else:
|
853 |
+
for _ in range(2):
|
854 |
+
opt.append(
|
855 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s"
|
856 |
+
% (now_dir, sr2, now_dir, fea_dim, spk_id5)
|
857 |
+
)
|
858 |
+
shuffle(opt)
|
859 |
+
with open("%s/filelist.txt" % exp_dir, "w") as f:
|
860 |
+
f.write("\n".join(opt))
|
861 |
+
logging.info("write filelist done")
|
862 |
+
# 生成config#无需生成config
|
863 |
+
# cmd = python_cmd + " train_nsf_sim_cache_sid_load_pretrain.py -e mi-test -sr 40k -f0 1 -bs 4 -g 0 -te 10 -se 5 -pg pretrained/f0G40k.pth -pd pretrained/f0D40k.pth -l 1 -c 0"
|
864 |
+
logging.info("use gpus:", gpus16)
|
865 |
+
if pretrained_G14 == "":
|
866 |
+
logging.info("no pretrained Generator")
|
867 |
+
if pretrained_D15 == "":
|
868 |
+
logging.info("no pretrained Discriminator")
|
869 |
+
if gpus16:
|
870 |
+
cmd = (
|
871 |
+
config.python_cmd
|
872 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -g %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
873 |
+
% (
|
874 |
+
exp_dir1,
|
875 |
+
sr2,
|
876 |
+
1 if if_f0_3 else 0,
|
877 |
+
batch_size12,
|
878 |
+
gpus16,
|
879 |
+
total_epoch11,
|
880 |
+
save_epoch10,
|
881 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
882 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
883 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
884 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
885 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
886 |
+
version19,
|
887 |
+
)
|
888 |
+
)
|
889 |
+
else:
|
890 |
+
cmd = (
|
891 |
+
config.python_cmd
|
892 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
893 |
+
% (
|
894 |
+
exp_dir1,
|
895 |
+
sr2,
|
896 |
+
1 if if_f0_3 else 0,
|
897 |
+
batch_size12,
|
898 |
+
total_epoch11,
|
899 |
+
save_epoch10,
|
900 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "\b",
|
901 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "\b",
|
902 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
903 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
904 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
905 |
+
version19,
|
906 |
+
)
|
907 |
+
)
|
908 |
+
logging.info(cmd)
|
909 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
910 |
+
p.wait()
|
911 |
+
return "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log"
|
912 |
+
|
913 |
+
|
914 |
+
# but4.click(train_index, [exp_dir1], info3)
|
915 |
+
def train_index(exp_dir1, version19):
|
916 |
+
exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
917 |
+
os.makedirs(exp_dir, exist_ok=True)
|
918 |
+
feature_dir = (
|
919 |
+
"%s/3_feature256" % (exp_dir)
|
920 |
+
if version19 == "v1"
|
921 |
+
else "%s/3_feature768" % (exp_dir)
|
922 |
+
)
|
923 |
+
if not os.path.exists(feature_dir):
|
924 |
+
return "请先进行特征提取!"
|
925 |
+
listdir_res = list(os.listdir(feature_dir))
|
926 |
+
if len(listdir_res) == 0:
|
927 |
+
return "请先进行特征提取!"
|
928 |
+
infos = []
|
929 |
+
npys = []
|
930 |
+
for name in sorted(listdir_res):
|
931 |
+
phone = np.load("%s/%s" % (feature_dir, name))
|
932 |
+
npys.append(phone)
|
933 |
+
big_npy = np.concatenate(npys, 0)
|
934 |
+
big_npy_idx = np.arange(big_npy.shape[0])
|
935 |
+
np.random.shuffle(big_npy_idx)
|
936 |
+
big_npy = big_npy[big_npy_idx]
|
937 |
+
if big_npy.shape[0] > 2e5:
|
938 |
+
# if(1):
|
939 |
+
infos.append("Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0])
|
940 |
+
yield "\n".join(infos)
|
941 |
+
try:
|
942 |
+
big_npy = (
|
943 |
+
MiniBatchKMeans(
|
944 |
+
n_clusters=10000,
|
945 |
+
verbose=True,
|
946 |
+
batch_size=256 * config.n_cpu,
|
947 |
+
compute_labels=False,
|
948 |
+
init="random",
|
949 |
+
)
|
950 |
+
.fit(big_npy)
|
951 |
+
.cluster_centers_
|
952 |
+
)
|
953 |
+
except:
|
954 |
+
info = traceback.format_exc()
|
955 |
+
logging.info(info)
|
956 |
+
infos.append(info)
|
957 |
+
yield "\n".join(infos)
|
958 |
+
|
959 |
+
np.save("%s/total_fea.npy" % exp_dir, big_npy)
|
960 |
+
n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
|
961 |
+
infos.append("%s,%s" % (big_npy.shape, n_ivf))
|
962 |
+
yield "\n".join(infos)
|
963 |
+
index = faiss.index_factory(256 if version19 == "v1" else 768, "IVF%s,Flat" % n_ivf)
|
964 |
+
# index = faiss.index_factory(256if version19=="v1"else 768, "IVF%s,PQ128x4fs,RFlat"%n_ivf)
|
965 |
+
infos.append("training")
|
966 |
+
yield "\n".join(infos)
|
967 |
+
index_ivf = faiss.extract_index_ivf(index) #
|
968 |
+
index_ivf.nprobe = 1
|
969 |
+
index.train(big_npy)
|
970 |
+
faiss.write_index(
|
971 |
+
index,
|
972 |
+
"%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
973 |
+
% (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
974 |
+
)
|
975 |
+
# faiss.write_index(index, '%s/trained_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19))
|
976 |
+
infos.append("adding")
|
977 |
+
yield "\n".join(infos)
|
978 |
+
batch_size_add = 8192
|
979 |
+
for i in range(0, big_npy.shape[0], batch_size_add):
|
980 |
+
index.add(big_npy[i : i + batch_size_add])
|
981 |
+
faiss.write_index(
|
982 |
+
index,
|
983 |
+
"%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
984 |
+
% (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
985 |
+
)
|
986 |
+
infos.append(
|
987 |
+
"成功构建索引,added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
988 |
+
% (n_ivf, index_ivf.nprobe, exp_dir1, version19)
|
989 |
+
)
|
990 |
+
# faiss.write_index(index, '%s/added_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19))
|
991 |
+
# infos.append("成功构建索引,added_IVF%s_Flat_FastScan_%s.index"%(n_ivf,version19))
|
992 |
+
yield "\n".join(infos)
|
993 |
+
|
994 |
+
|
995 |
+
# but5.click(train1key, [exp_dir1, sr2, if_f0_3, trainset_dir4, spk_id5, gpus6, np7, f0method8, save_epoch10, total_epoch11, batch_size12, if_save_latest13, pretrained_G14, pretrained_D15, gpus16, if_cache_gpu17], info3)
|
996 |
+
def train1key(
|
997 |
+
exp_dir1,
|
998 |
+
sr2,
|
999 |
+
if_f0_3,
|
1000 |
+
trainset_dir4,
|
1001 |
+
spk_id5,
|
1002 |
+
np7,
|
1003 |
+
f0method8,
|
1004 |
+
save_epoch10,
|
1005 |
+
total_epoch11,
|
1006 |
+
batch_size12,
|
1007 |
+
if_save_latest13,
|
1008 |
+
pretrained_G14,
|
1009 |
+
pretrained_D15,
|
1010 |
+
gpus16,
|
1011 |
+
if_cache_gpu17,
|
1012 |
+
if_save_every_weights18,
|
1013 |
+
version19,
|
1014 |
+
):
|
1015 |
+
infos = []
|
1016 |
+
|
1017 |
+
def get_info_str(strr):
|
1018 |
+
infos.append(strr)
|
1019 |
+
return "\n".join(infos)
|
1020 |
+
|
1021 |
+
model_log_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
1022 |
+
preprocess_log_path = "%s/preprocess.log" % model_log_dir
|
1023 |
+
extract_f0_feature_log_path = "%s/extract_f0_feature.log" % model_log_dir
|
1024 |
+
gt_wavs_dir = "%s/0_gt_wavs" % model_log_dir
|
1025 |
+
feature_dir = (
|
1026 |
+
"%s/3_feature256" % model_log_dir
|
1027 |
+
if version19 == "v1"
|
1028 |
+
else "%s/3_feature768" % model_log_dir
|
1029 |
+
)
|
1030 |
+
|
1031 |
+
os.makedirs(model_log_dir, exist_ok=True)
|
1032 |
+
#########step1:处理数据
|
1033 |
+
open(preprocess_log_path, "w").close()
|
1034 |
+
cmd = (
|
1035 |
+
config.python_cmd
|
1036 |
+
+ " trainset_preprocess_pipeline_print.py %s %s %s %s "
|
1037 |
+
% (trainset_dir4, sr_dict[sr2], np7, model_log_dir)
|
1038 |
+
+ str(config.noparallel)
|
1039 |
+
)
|
1040 |
+
yield get_info_str(i18n("step1:正在处理数据"))
|
1041 |
+
yield get_info_str(cmd)
|
1042 |
+
p = Popen(cmd, shell=True)
|
1043 |
+
p.wait()
|
1044 |
+
with open(preprocess_log_path, "r") as f:
|
1045 |
+
logging.info(f.read())
|
1046 |
+
#########step2a:提取音高
|
1047 |
+
open(extract_f0_feature_log_path, "w")
|
1048 |
+
if if_f0_3:
|
1049 |
+
yield get_info_str("step2a:正在提取音高")
|
1050 |
+
cmd = config.python_cmd + " extract_f0_print.py %s %s %s" % (
|
1051 |
+
model_log_dir,
|
1052 |
+
np7,
|
1053 |
+
f0method8,
|
1054 |
+
)
|
1055 |
+
yield get_info_str(cmd)
|
1056 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
1057 |
+
p.wait()
|
1058 |
+
with open(extract_f0_feature_log_path, "r") as f:
|
1059 |
+
logging.info(f.read())
|
1060 |
+
else:
|
1061 |
+
yield get_info_str(i18n("step2a:无需提取音高"))
|
1062 |
+
#######step2b:提取特征
|
1063 |
+
yield get_info_str(i18n("step2b:正在提取特征"))
|
1064 |
+
gpus = gpus16.split("-")
|
1065 |
+
leng = len(gpus)
|
1066 |
+
ps = []
|
1067 |
+
for idx, n_g in enumerate(gpus):
|
1068 |
+
cmd = config.python_cmd + " extract_feature_print.py %s %s %s %s %s %s" % (
|
1069 |
+
config.device,
|
1070 |
+
leng,
|
1071 |
+
idx,
|
1072 |
+
n_g,
|
1073 |
+
model_log_dir,
|
1074 |
+
version19,
|
1075 |
+
)
|
1076 |
+
yield get_info_str(cmd)
|
1077 |
+
p = Popen(
|
1078 |
+
cmd, shell=True, cwd=now_dir
|
1079 |
+
) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
|
1080 |
+
ps.append(p)
|
1081 |
+
for p in ps:
|
1082 |
+
p.wait()
|
1083 |
+
with open(extract_f0_feature_log_path, "r") as f:
|
1084 |
+
logging.info(f.read())
|
1085 |
+
#######step3a:训练模型
|
1086 |
+
yield get_info_str(i18n("step3a:正在训练模型"))
|
1087 |
+
# 生成filelist
|
1088 |
+
if if_f0_3:
|
1089 |
+
f0_dir = "%s/2a_f0" % model_log_dir
|
1090 |
+
f0nsf_dir = "%s/2b-f0nsf" % model_log_dir
|
1091 |
+
names = (
|
1092 |
+
set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)])
|
1093 |
+
& set([name.split(".")[0] for name in os.listdir(feature_dir)])
|
1094 |
+
& set([name.split(".")[0] for name in os.listdir(f0_dir)])
|
1095 |
+
& set([name.split(".")[0] for name in os.listdir(f0nsf_dir)])
|
1096 |
+
)
|
1097 |
+
else:
|
1098 |
+
names = set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)]) & set(
|
1099 |
+
[name.split(".")[0] for name in os.listdir(feature_dir)]
|
1100 |
+
)
|
1101 |
+
opt = []
|
1102 |
+
for name in names:
|
1103 |
+
if if_f0_3:
|
1104 |
+
opt.append(
|
1105 |
+
"%s/%s.wav|%s/%s.npy|%s/%s.wav.npy|%s/%s.wav.npy|%s"
|
1106 |
+
% (
|
1107 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
1108 |
+
name,
|
1109 |
+
feature_dir.replace("\\", "\\\\"),
|
1110 |
+
name,
|
1111 |
+
f0_dir.replace("\\", "\\\\"),
|
1112 |
+
name,
|
1113 |
+
f0nsf_dir.replace("\\", "\\\\"),
|
1114 |
+
name,
|
1115 |
+
spk_id5,
|
1116 |
+
)
|
1117 |
+
)
|
1118 |
+
else:
|
1119 |
+
opt.append(
|
1120 |
+
"%s/%s.wav|%s/%s.npy|%s"
|
1121 |
+
% (
|
1122 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
1123 |
+
name,
|
1124 |
+
feature_dir.replace("\\", "\\\\"),
|
1125 |
+
name,
|
1126 |
+
spk_id5,
|
1127 |
+
)
|
1128 |
+
)
|
1129 |
+
fea_dim = 256 if version19 == "v1" else 768
|
1130 |
+
if if_f0_3:
|
1131 |
+
for _ in range(2):
|
1132 |
+
opt.append(
|
1133 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s/logs/mute/2a_f0/mute.wav.npy|%s/logs/mute/2b-f0nsf/mute.wav.npy|%s"
|
1134 |
+
% (now_dir, sr2, now_dir, fea_dim, now_dir, now_dir, spk_id5)
|
1135 |
+
)
|
1136 |
+
else:
|
1137 |
+
for _ in range(2):
|
1138 |
+
opt.append(
|
1139 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s"
|
1140 |
+
% (now_dir, sr2, now_dir, fea_dim, spk_id5)
|
1141 |
+
)
|
1142 |
+
shuffle(opt)
|
1143 |
+
with open("%s/filelist.txt" % model_log_dir, "w") as f:
|
1144 |
+
f.write("\n".join(opt))
|
1145 |
+
yield get_info_str("write filelist done")
|
1146 |
+
if gpus16:
|
1147 |
+
cmd = (
|
1148 |
+
config.python_cmd
|
1149 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -g %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
1150 |
+
% (
|
1151 |
+
exp_dir1,
|
1152 |
+
sr2,
|
1153 |
+
1 if if_f0_3 else 0,
|
1154 |
+
batch_size12,
|
1155 |
+
gpus16,
|
1156 |
+
total_epoch11,
|
1157 |
+
save_epoch10,
|
1158 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
1159 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
1160 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
1161 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
1162 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
1163 |
+
version19,
|
1164 |
+
)
|
1165 |
+
)
|
1166 |
+
else:
|
1167 |
+
cmd = (
|
1168 |
+
config.python_cmd
|
1169 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
1170 |
+
% (
|
1171 |
+
exp_dir1,
|
1172 |
+
sr2,
|
1173 |
+
1 if if_f0_3 else 0,
|
1174 |
+
batch_size12,
|
1175 |
+
total_epoch11,
|
1176 |
+
save_epoch10,
|
1177 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
1178 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
1179 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
1180 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
1181 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
1182 |
+
version19,
|
1183 |
+
)
|
1184 |
+
)
|
1185 |
+
yield get_info_str(cmd)
|
1186 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
1187 |
+
p.wait()
|
1188 |
+
yield get_info_str(i18n("训练结束, 您可查看控制台训练日志或实验文件夹下的train.log"))
|
1189 |
+
#######step3b:训练索引
|
1190 |
+
npys = []
|
1191 |
+
listdir_res = list(os.listdir(feature_dir))
|
1192 |
+
for name in sorted(listdir_res):
|
1193 |
+
phone = np.load("%s/%s" % (feature_dir, name))
|
1194 |
+
npys.append(phone)
|
1195 |
+
big_npy = np.concatenate(npys, 0)
|
1196 |
+
|
1197 |
+
big_npy_idx = np.arange(big_npy.shape[0])
|
1198 |
+
np.random.shuffle(big_npy_idx)
|
1199 |
+
big_npy = big_npy[big_npy_idx]
|
1200 |
+
|
1201 |
+
if big_npy.shape[0] > 2e5:
|
1202 |
+
# if(1):
|
1203 |
+
info = "Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0]
|
1204 |
+
logging.info(info)
|
1205 |
+
yield get_info_str(info)
|
1206 |
+
try:
|
1207 |
+
big_npy = (
|
1208 |
+
MiniBatchKMeans(
|
1209 |
+
n_clusters=10000,
|
1210 |
+
verbose=True,
|
1211 |
+
batch_size=256 * config.n_cpu,
|
1212 |
+
compute_labels=False,
|
1213 |
+
init="random",
|
1214 |
+
)
|
1215 |
+
.fit(big_npy)
|
1216 |
+
.cluster_centers_
|
1217 |
+
)
|
1218 |
+
except:
|
1219 |
+
info = traceback.format_exc()
|
1220 |
+
logging.info(info)
|
1221 |
+
yield get_info_str(info)
|
1222 |
+
|
1223 |
+
np.save("%s/total_fea.npy" % model_log_dir, big_npy)
|
1224 |
+
|
1225 |
+
# n_ivf = big_npy.shape[0] // 39
|
1226 |
+
n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
|
1227 |
+
yield get_info_str("%s,%s" % (big_npy.shape, n_ivf))
|
1228 |
+
index = faiss.index_factory(256 if version19 == "v1" else 768, "IVF%s,Flat" % n_ivf)
|
1229 |
+
yield get_info_str("training index")
|
1230 |
+
index_ivf = faiss.extract_index_ivf(index) #
|
1231 |
+
index_ivf.nprobe = 1
|
1232 |
+
index.train(big_npy)
|
1233 |
+
faiss.write_index(
|
1234 |
+
index,
|
1235 |
+
"%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1236 |
+
% (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
1237 |
+
)
|
1238 |
+
yield get_info_str("adding index")
|
1239 |
+
batch_size_add = 8192
|
1240 |
+
for i in range(0, big_npy.shape[0], batch_size_add):
|
1241 |
+
index.add(big_npy[i : i + batch_size_add])
|
1242 |
+
faiss.write_index(
|
1243 |
+
index,
|
1244 |
+
"%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1245 |
+
% (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
1246 |
+
)
|
1247 |
+
yield get_info_str(
|
1248 |
+
"成功构建索引, added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1249 |
+
% (n_ivf, index_ivf.nprobe, exp_dir1, version19)
|
1250 |
+
)
|
1251 |
+
yield get_info_str(i18n("全流程结束!"))
|
1252 |
+
|
1253 |
+
|
1254 |
+
# ckpt_path2.change(change_info_,[ckpt_path2],[sr__,if_f0__])
|
1255 |
+
def change_info_(ckpt_path):
|
1256 |
+
if not os.path.exists(ckpt_path.replace(os.path.basename(ckpt_path), "train.log")):
|
1257 |
+
return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
|
1258 |
+
try:
|
1259 |
+
with open(
|
1260 |
+
ckpt_path.replace(os.path.basename(ckpt_path), "train.log"), "r"
|
1261 |
+
) as f:
|
1262 |
+
info = eval(f.read().strip("\n").split("\n")[0].split("\t")[-1])
|
1263 |
+
sr, f0 = info["sample_rate"], info["if_f0"]
|
1264 |
+
version = "v2" if ("version" in info and info["version"] == "v2") else "v1"
|
1265 |
+
return sr, str(f0), version
|
1266 |
+
except:
|
1267 |
+
traceback.print_exc()
|
1268 |
+
return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
|
1269 |
+
|
1270 |
+
|
1271 |
+
def export_onnx(ModelPath, ExportedPath):
|
1272 |
+
cpt = torch.load(ModelPath, map_location="cpu")
|
1273 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
|
1274 |
+
vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768
|
1275 |
+
|
1276 |
+
test_phone = torch.rand(1, 200, vec_channels) # hidden unit
|
1277 |
+
test_phone_lengths = torch.tensor([200]).long() # hidden unit 长度(貌似没啥用)
|
1278 |
+
test_pitch = torch.randint(size=(1, 200), low=5, high=255) # 基频(单位赫兹)
|
1279 |
+
test_pitchf = torch.rand(1, 200) # nsf基频
|
1280 |
+
test_ds = torch.LongTensor([0]) # 说话人ID
|
1281 |
+
test_rnd = torch.rand(1, 192, 200) # 噪声(加入随机因子)
|
1282 |
+
|
1283 |
+
device = "cpu" # 导出时设备(不影响使用模型)
|
1284 |
+
|
1285 |
+
net_g = SynthesizerTrnMsNSFsidM(
|
1286 |
+
*cpt["config"], is_half=False, version=cpt.get("version", "v1")
|
1287 |
+
) # fp32导出(C++要支持fp16必须手动将内存重新排列所以暂时不用fp16)
|
1288 |
+
net_g.load_state_dict(cpt["weight"], strict=False)
|
1289 |
+
input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
|
1290 |
+
output_names = [
|
1291 |
+
"audio",
|
1292 |
+
]
|
1293 |
+
# net_g.construct_spkmixmap(n_speaker) 多角色混合轨道导出
|
1294 |
+
torch.onnx.export(
|
1295 |
+
net_g,
|
1296 |
+
(
|
1297 |
+
test_phone.to(device),
|
1298 |
+
test_phone_lengths.to(device),
|
1299 |
+
test_pitch.to(device),
|
1300 |
+
test_pitchf.to(device),
|
1301 |
+
test_ds.to(device),
|
1302 |
+
test_rnd.to(device),
|
1303 |
+
),
|
1304 |
+
ExportedPath,
|
1305 |
+
dynamic_axes={
|
1306 |
+
"phone": [1],
|
1307 |
+
"pitch": [1],
|
1308 |
+
"pitchf": [1],
|
1309 |
+
"rnd": [2],
|
1310 |
+
},
|
1311 |
+
do_constant_folding=False,
|
1312 |
+
opset_version=13,
|
1313 |
+
verbose=False,
|
1314 |
+
input_names=input_names,
|
1315 |
+
output_names=output_names,
|
1316 |
+
)
|
1317 |
+
return "Finished"
|
1318 |
+
|
RVC_class.py
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import RVC
|
2 |
+
from scipy.io import wavfile
|
3 |
+
import numpy as np
|
4 |
+
import os
|
5 |
+
import uuid
|
6 |
+
import io
|
7 |
+
import requests
|
8 |
+
|
9 |
+
class VoiceConverter:
|
10 |
+
def __init__(self):
|
11 |
+
self.models = RVC.names
|
12 |
+
|
13 |
+
def single_run(self, input_audio, model_name, vc_transform, song, opt_input="opt", f0_method="crepe",
|
14 |
+
filter_radius=3, file_index1="", file_index2="", index_rate1=0.75,
|
15 |
+
resample_sr=0, rms_mix_rate=0.25, protect=0.33, f0_file=None):
|
16 |
+
|
17 |
+
RVC.get_vc(model_name, protect, protect)
|
18 |
+
spk_item = 0
|
19 |
+
|
20 |
+
vc_output1, vc_output2 = RVC.vc_single(
|
21 |
+
spk_item,
|
22 |
+
input_audio,
|
23 |
+
vc_transform,
|
24 |
+
f0_file,
|
25 |
+
f0_method,
|
26 |
+
file_index1,
|
27 |
+
file_index2,
|
28 |
+
index_rate1,
|
29 |
+
filter_radius,
|
30 |
+
resample_sr,
|
31 |
+
rms_mix_rate,
|
32 |
+
protect,
|
33 |
+
song
|
34 |
+
)
|
35 |
+
|
36 |
+
random_string = str(uuid.uuid4())
|
37 |
+
filename = os.path.basename(input_audio)
|
38 |
+
name, extension = os.path.splitext(filename)
|
39 |
+
new_file_name = f"{name}_{random_string}{extension}"
|
40 |
+
|
41 |
+
sample_rate, audio_data = vc_output2
|
42 |
+
url = self.upload_audio(audio_data, sample_rate, new_file_name)
|
43 |
+
|
44 |
+
return url
|
45 |
+
|
46 |
+
def upload_audio(self, audio_data, sample_rate, filename):
|
47 |
+
try:
|
48 |
+
url = "https://tmpfiles.org/api/v1/upload"
|
49 |
+
# Convert audio data to WAV format in memory
|
50 |
+
wav_bytes = io.BytesIO()
|
51 |
+
scaled_audio_data = np.int16(audio_data)
|
52 |
+
wavfile.write(wav_bytes, sample_rate, scaled_audio_data)
|
53 |
+
wav_bytes.seek(0)
|
54 |
+
|
55 |
+
files = {'file': (filename, wav_bytes)}
|
56 |
+
|
57 |
+
response = requests.post(url, files=files)
|
58 |
+
response.raise_for_status()
|
59 |
+
|
60 |
+
return response.json()
|
61 |
+
except Exception as e:
|
62 |
+
raise RuntimeError(f"Failed to upload audio: {e}")
|
63 |
+
|
64 |
+
|
65 |
+
def uvr(self, dir_wav_input, wav_inputs=None, model_choose=RVC.uvr5_names[0],
|
66 |
+
opt_vocal_root="opt", opt_ins_root="opt", format0="wav"):
|
67 |
+
|
68 |
+
agg = 10
|
69 |
+
vc_output4 = RVC.uvr(
|
70 |
+
model_choose,
|
71 |
+
dir_wav_input,
|
72 |
+
opt_vocal_root,
|
73 |
+
wav_inputs,
|
74 |
+
opt_ins_root,
|
75 |
+
agg,
|
76 |
+
format0
|
77 |
+
)
|
78 |
+
|
79 |
+
for value in vc_output4:
|
80 |
+
print(value)
|
81 |
+
|
82 |
+
|
83 |
+
|
84 |
+
|
85 |
+
# converter.uvr('/home/teewhy/Downloads/around_the_world-atc.wav')
|
RVC_functions.py
ADDED
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import RVC
|
2 |
+
from scipy.io import wavfile
|
3 |
+
import numpy as np
|
4 |
+
import os
|
5 |
+
import uuid
|
6 |
+
import io
|
7 |
+
import requests
|
8 |
+
def get_models():
|
9 |
+
return RVC.names
|
10 |
+
|
11 |
+
def single_run(input_audio0, model_name, vc_transform, opt_input="opt", f0method0 = "pm", filter_radius0 = 3, file_index1 = "", file_index2 = "",
|
12 |
+
index_rate1 = 0.75, resample_sr0 = 0, rms_mix_rate0 = 0.25, protect0 = 0.33, f0_file = None):
|
13 |
+
|
14 |
+
RVC.get_vc(model_name, protect0, protect0)
|
15 |
+
spk_item = 0
|
16 |
+
# file_index2 = sorted(index_paths)
|
17 |
+
# f0_file = gr.File(label=i18n("F0曲线文件, 可选, 一行一个音高, 代替默认F0及升降调"))
|
18 |
+
|
19 |
+
vc_output1, vc_output2 = RVC.vc_single(
|
20 |
+
spk_item,
|
21 |
+
input_audio0,
|
22 |
+
vc_transform,
|
23 |
+
f0_file,
|
24 |
+
f0method0,
|
25 |
+
file_index1,
|
26 |
+
file_index2,
|
27 |
+
index_rate1,
|
28 |
+
filter_radius0,
|
29 |
+
resample_sr0,
|
30 |
+
rms_mix_rate0,
|
31 |
+
protect0
|
32 |
+
)
|
33 |
+
random_string = str(uuid.uuid4())
|
34 |
+
filename = os.path.basename(input_audio0)
|
35 |
+
name, extension = os.path.splitext(filename)
|
36 |
+
new_file_name = f"{name}_{random_string}{extension}"
|
37 |
+
|
38 |
+
sample_rate, audio_data = vc_output2
|
39 |
+
# scaled_audio_data = np.int16(audio_data)
|
40 |
+
# wavfile.write(new_file_name, sample_rate, scaled_audio_data)
|
41 |
+
# response = requests.post(f"https://filebin.net/dajeii61xk3c4oxi/{new_file_name}" , files=files)
|
42 |
+
url="https://filebin.net/dajeii61xk3c4oxi"
|
43 |
+
print(upload_audio(url, audio_data, sample_rate, new_file_name))
|
44 |
+
|
45 |
+
return f"{url}/{new_file_name}"
|
46 |
+
|
47 |
+
|
48 |
+
def upload_audio(url, audio_data, sample_rate, filename):
|
49 |
+
try:
|
50 |
+
# Convert audio data to WAV format in memory
|
51 |
+
wav_bytes = io.BytesIO()
|
52 |
+
scaled_audio_data = np.int16(audio_data)
|
53 |
+
wavfile.write(wav_bytes, sample_rate, scaled_audio_data)
|
54 |
+
wav_bytes.seek(0)
|
55 |
+
|
56 |
+
files = {'file': (filename, wav_bytes)}
|
57 |
+
|
58 |
+
response = requests.post(f"{url}/{filename}", files=files)
|
59 |
+
|
60 |
+
response.raise_for_status()
|
61 |
+
|
62 |
+
return response.json()
|
63 |
+
except Exception as e:
|
64 |
+
raise RuntimeError(f"Failed to upload audio: {e}")
|
65 |
+
|
66 |
+
|
67 |
+
def uvr(dir_wav_input, wav_inputs = None, model_choose = RVC.uvr5_names[0], opt_vocal_root ="opt", opt_ins_root = "opt", format0 = "wav"):
|
68 |
+
|
69 |
+
vc_output4 = RVC.uvr(
|
70 |
+
dir_wav_input,
|
71 |
+
)
|
72 |
+
for value in vc_output4:
|
73 |
+
print(value)
|
74 |
+
|
75 |
+
|
76 |
+
print(single_run("https://www.learningcontainer.com/wp-content/uploads/2020/02/Kalimba.mp3", "mymodelbilawal.pth", -2))
|
Retrieval_based_Voice_Conversion_WebUI.ipynb
ADDED
@@ -0,0 +1,381 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"nbformat": 4,
|
3 |
+
"nbformat_minor": 0,
|
4 |
+
"metadata": {
|
5 |
+
"colab": {
|
6 |
+
"private_outputs": true,
|
7 |
+
"provenance": []
|
8 |
+
},
|
9 |
+
"kernelspec": {
|
10 |
+
"name": "python3",
|
11 |
+
"display_name": "Python 3"
|
12 |
+
},
|
13 |
+
"language_info": {
|
14 |
+
"name": "python"
|
15 |
+
},
|
16 |
+
"accelerator": "GPU",
|
17 |
+
"gpuClass": "standard"
|
18 |
+
},
|
19 |
+
"cells": [
|
20 |
+
{
|
21 |
+
"cell_type": "markdown",
|
22 |
+
"source": [
|
23 |
+
"[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/Retrieval_based_Voice_Conversion_WebUI.ipynb)"
|
24 |
+
],
|
25 |
+
"metadata": {
|
26 |
+
"id": "ZFFCx5J80SGa"
|
27 |
+
}
|
28 |
+
},
|
29 |
+
{
|
30 |
+
"cell_type": "code",
|
31 |
+
"execution_count": null,
|
32 |
+
"metadata": {
|
33 |
+
"id": "GmFP6bN9dvOq"
|
34 |
+
},
|
35 |
+
"outputs": [],
|
36 |
+
"source": [
|
37 |
+
"#@title 查看显卡\n",
|
38 |
+
"!nvidia-smi"
|
39 |
+
]
|
40 |
+
},
|
41 |
+
{
|
42 |
+
"cell_type": "code",
|
43 |
+
"source": [
|
44 |
+
"#@title 安装依赖\n",
|
45 |
+
"!apt-get -y install build-essential python3-dev ffmpeg\n",
|
46 |
+
"!pip3 install --upgrade setuptools wheel\n",
|
47 |
+
"!pip3 install --upgrade pip\n",
|
48 |
+
"!pip3 install faiss-cpu==1.7.2 fairseq gradio==3.14.0 ffmpeg ffmpeg-python praat-parselmouth pyworld numpy==1.23.5 numba==0.56.4 librosa==0.9.2"
|
49 |
+
],
|
50 |
+
"metadata": {
|
51 |
+
"id": "wjddIFr1oS3W"
|
52 |
+
},
|
53 |
+
"execution_count": null,
|
54 |
+
"outputs": []
|
55 |
+
},
|
56 |
+
{
|
57 |
+
"cell_type": "code",
|
58 |
+
"source": [
|
59 |
+
"#@title 克隆仓库\n",
|
60 |
+
"\n",
|
61 |
+
"!git clone --depth=1 -b stable https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI\n",
|
62 |
+
"%cd /content/Retrieval-based-Voice-Conversion-WebUI\n",
|
63 |
+
"!mkdir -p pretrained uvr5_weights"
|
64 |
+
],
|
65 |
+
"metadata": {
|
66 |
+
"id": "ge_97mfpgqTm"
|
67 |
+
},
|
68 |
+
"execution_count": null,
|
69 |
+
"outputs": []
|
70 |
+
},
|
71 |
+
{
|
72 |
+
"cell_type": "code",
|
73 |
+
"source": [
|
74 |
+
"#@title 更新仓库(一般无需执行)\n",
|
75 |
+
"!git pull"
|
76 |
+
],
|
77 |
+
"metadata": {
|
78 |
+
"id": "BLDEZADkvlw1"
|
79 |
+
},
|
80 |
+
"execution_count": null,
|
81 |
+
"outputs": []
|
82 |
+
},
|
83 |
+
{
|
84 |
+
"cell_type": "code",
|
85 |
+
"source": [
|
86 |
+
"#@title 安装aria2\n",
|
87 |
+
"!apt -y install -qq aria2"
|
88 |
+
],
|
89 |
+
"metadata": {
|
90 |
+
"id": "pqE0PrnuRqI2"
|
91 |
+
},
|
92 |
+
"execution_count": null,
|
93 |
+
"outputs": []
|
94 |
+
},
|
95 |
+
{
|
96 |
+
"cell_type": "code",
|
97 |
+
"source": [
|
98 |
+
"#@title 下载底模\n",
|
99 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o D32k.pth\n",
|
100 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o D40k.pth\n",
|
101 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o D48k.pth\n",
|
102 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o G32k.pth\n",
|
103 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o G40k.pth\n",
|
104 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o G48k.pth\n",
|
105 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0D32k.pth\n",
|
106 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0D40k.pth\n",
|
107 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0D48k.pth\n",
|
108 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0G32k.pth\n",
|
109 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0G40k.pth\n",
|
110 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0G48k.pth"
|
111 |
+
],
|
112 |
+
"metadata": {
|
113 |
+
"id": "UG3XpUwEomUz"
|
114 |
+
},
|
115 |
+
"execution_count": null,
|
116 |
+
"outputs": []
|
117 |
+
},
|
118 |
+
{
|
119 |
+
"cell_type": "code",
|
120 |
+
"source": [
|
121 |
+
"#@title 下载人声分离模型\n",
|
122 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP2-人声vocals+非人声instrumentals.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/uvr5_weights -o HP2-人声vocals+非人声instrumentals.pth\n",
|
123 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP5-主旋律人声vocals+其他instrumentals.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/uvr5_weights -o HP5-主旋律人声vocals+其他instrumentals.pth"
|
124 |
+
],
|
125 |
+
"metadata": {
|
126 |
+
"id": "HugjmZqZRuiF"
|
127 |
+
},
|
128 |
+
"execution_count": null,
|
129 |
+
"outputs": []
|
130 |
+
},
|
131 |
+
{
|
132 |
+
"cell_type": "code",
|
133 |
+
"source": [
|
134 |
+
"#@title 下载hubert_base\n",
|
135 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt -d /content/Retrieval-based-Voice-Conversion-WebUI -o hubert_base.pt"
|
136 |
+
],
|
137 |
+
"metadata": {
|
138 |
+
"id": "2RCaT9FTR0ej"
|
139 |
+
},
|
140 |
+
"execution_count": null,
|
141 |
+
"outputs": []
|
142 |
+
},
|
143 |
+
{
|
144 |
+
"cell_type": "code",
|
145 |
+
"source": [
|
146 |
+
"#@title 挂载谷歌云盘\n",
|
147 |
+
"\n",
|
148 |
+
"from google.colab import drive\n",
|
149 |
+
"drive.mount('/content/drive')"
|
150 |
+
],
|
151 |
+
"metadata": {
|
152 |
+
"id": "jwu07JgqoFON"
|
153 |
+
},
|
154 |
+
"execution_count": null,
|
155 |
+
"outputs": []
|
156 |
+
},
|
157 |
+
{
|
158 |
+
"cell_type": "code",
|
159 |
+
"source": [
|
160 |
+
"#@title 从谷歌云盘加载打包好的数据集到/content/dataset\n",
|
161 |
+
"\n",
|
162 |
+
"#@markdown 数据集位置\n",
|
163 |
+
"DATASET = \"/content/drive/MyDrive/dataset/lulu20230327_32k.zip\" #@param {type:\"string\"}\n",
|
164 |
+
"\n",
|
165 |
+
"!mkdir -p /content/dataset\n",
|
166 |
+
"!unzip -d /content/dataset -B {DATASET}"
|
167 |
+
],
|
168 |
+
"metadata": {
|
169 |
+
"id": "Mwk7Q0Loqzjx"
|
170 |
+
},
|
171 |
+
"execution_count": null,
|
172 |
+
"outputs": []
|
173 |
+
},
|
174 |
+
{
|
175 |
+
"cell_type": "code",
|
176 |
+
"source": [
|
177 |
+
"#@title 重命名数据集中的重名文件\n",
|
178 |
+
"!ls -a /content/dataset/\n",
|
179 |
+
"!rename 's/(\\w+)\\.(\\w+)~(\\d*)/$1_$3.$2/' /content/dataset/*.*~*"
|
180 |
+
],
|
181 |
+
"metadata": {
|
182 |
+
"id": "PDlFxWHWEynD"
|
183 |
+
},
|
184 |
+
"execution_count": null,
|
185 |
+
"outputs": []
|
186 |
+
},
|
187 |
+
{
|
188 |
+
"cell_type": "code",
|
189 |
+
"source": [
|
190 |
+
"#@title 启动web\n",
|
191 |
+
"%cd /content/Retrieval-based-Voice-Conversion-WebUI\n",
|
192 |
+
"# %load_ext tensorboard\n",
|
193 |
+
"# %tensorboard --logdir /content/Retrieval-based-Voice-Conversion-WebUI/logs\n",
|
194 |
+
"!python3 infer-web.py --colab --pycmd python3"
|
195 |
+
],
|
196 |
+
"metadata": {
|
197 |
+
"id": "7vh6vphDwO0b"
|
198 |
+
},
|
199 |
+
"execution_count": null,
|
200 |
+
"outputs": []
|
201 |
+
},
|
202 |
+
{
|
203 |
+
"cell_type": "code",
|
204 |
+
"source": [
|
205 |
+
"#@title 手动将训练后的模型文件备份到谷歌云盘\n",
|
206 |
+
"#@markdown 需要自己查看logs文件夹下模型的文件名,手动修改下方命令末尾的文件名\n",
|
207 |
+
"\n",
|
208 |
+
"#@markdown 模型名\n",
|
209 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
210 |
+
"#@markdown 模型epoch\n",
|
211 |
+
"MODELEPOCH = 9600 #@param {type:\"integer\"}\n",
|
212 |
+
"\n",
|
213 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth /content/drive/MyDrive/{MODELNAME}_D_{MODELEPOCH}.pth\n",
|
214 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth /content/drive/MyDrive/{MODELNAME}_G_{MODELEPOCH}.pth\n",
|
215 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/added_*.index /content/drive/MyDrive/\n",
|
216 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/total_*.npy /content/drive/MyDrive/\n",
|
217 |
+
"\n",
|
218 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/weights/{MODELNAME}.pth /content/drive/MyDrive/{MODELNAME}{MODELEPOCH}.pth"
|
219 |
+
],
|
220 |
+
"metadata": {
|
221 |
+
"id": "FgJuNeAwx5Y_"
|
222 |
+
},
|
223 |
+
"execution_count": null,
|
224 |
+
"outputs": []
|
225 |
+
},
|
226 |
+
{
|
227 |
+
"cell_type": "code",
|
228 |
+
"source": [
|
229 |
+
"#@title 从谷歌云盘恢复pth\n",
|
230 |
+
"#@markdown 需要自己查看logs文件夹下模型的文件名,手动修改下方命令末尾的文件名\n",
|
231 |
+
"\n",
|
232 |
+
"#@markdown 模型名\n",
|
233 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
234 |
+
"#@markdown 模型epoch\n",
|
235 |
+
"MODELEPOCH = 7500 #@param {type:\"integer\"}\n",
|
236 |
+
"\n",
|
237 |
+
"!mkdir -p /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}\n",
|
238 |
+
"\n",
|
239 |
+
"!cp /content/drive/MyDrive/{MODELNAME}_D_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth\n",
|
240 |
+
"!cp /content/drive/MyDrive/{MODELNAME}_G_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth\n",
|
241 |
+
"!cp /content/drive/MyDrive/*.index /content/\n",
|
242 |
+
"!cp /content/drive/MyDrive/*.npy /content/\n",
|
243 |
+
"!cp /content/drive/MyDrive/{MODELNAME}{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/weights/{MODELNAME}.pth"
|
244 |
+
],
|
245 |
+
"metadata": {
|
246 |
+
"id": "OVQoLQJXS7WX"
|
247 |
+
},
|
248 |
+
"execution_count": null,
|
249 |
+
"outputs": []
|
250 |
+
},
|
251 |
+
{
|
252 |
+
"cell_type": "code",
|
253 |
+
"source": [
|
254 |
+
"#@title 手动预处理(不推荐)\n",
|
255 |
+
"#@markdown 模型名\n",
|
256 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
257 |
+
"#@markdown 采样率\n",
|
258 |
+
"BITRATE = 48000 #@param {type:\"integer\"}\n",
|
259 |
+
"#@markdown 使用的进程数\n",
|
260 |
+
"THREADCOUNT = 8 #@param {type:\"integer\"}\n",
|
261 |
+
"\n",
|
262 |
+
"!python3 trainset_preprocess_pipeline_print.py /content/dataset {BITRATE} {THREADCOUNT} logs/{MODELNAME} True\n"
|
263 |
+
],
|
264 |
+
"metadata": {
|
265 |
+
"id": "ZKAyuKb9J6dz"
|
266 |
+
},
|
267 |
+
"execution_count": null,
|
268 |
+
"outputs": []
|
269 |
+
},
|
270 |
+
{
|
271 |
+
"cell_type": "code",
|
272 |
+
"source": [
|
273 |
+
"#@title 手动提取特征(不推荐)\n",
|
274 |
+
"#@markdown 模型名\n",
|
275 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
276 |
+
"#@markdown 使用的进程数\n",
|
277 |
+
"THREADCOUNT = 8 #@param {type:\"integer\"}\n",
|
278 |
+
"#@markdown 音高提取算法\n",
|
279 |
+
"ALGO = \"harvest\" #@param {type:\"string\"}\n",
|
280 |
+
"\n",
|
281 |
+
"!python3 extract_f0_print.py logs/{MODELNAME} {THREADCOUNT} {ALGO}\n",
|
282 |
+
"\n",
|
283 |
+
"!python3 extract_feature_print.py cpu 1 0 0 logs/{MODELNAME}\n"
|
284 |
+
],
|
285 |
+
"metadata": {
|
286 |
+
"id": "CrxJqzAUKmPJ"
|
287 |
+
},
|
288 |
+
"execution_count": null,
|
289 |
+
"outputs": []
|
290 |
+
},
|
291 |
+
{
|
292 |
+
"cell_type": "code",
|
293 |
+
"source": [
|
294 |
+
"#@title 手动训练(不推荐)\n",
|
295 |
+
"#@markdown 模型名\n",
|
296 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
297 |
+
"#@markdown 使用的GPU\n",
|
298 |
+
"USEGPU = \"0\" #@param {type:\"string\"}\n",
|
299 |
+
"#@markdown 批大小\n",
|
300 |
+
"BATCHSIZE = 32 #@param {type:\"integer\"}\n",
|
301 |
+
"#@markdown 停止的epoch\n",
|
302 |
+
"MODELEPOCH = 3200 #@param {type:\"integer\"}\n",
|
303 |
+
"#@markdown 保存epoch间隔\n",
|
304 |
+
"EPOCHSAVE = 100 #@param {type:\"integer\"}\n",
|
305 |
+
"#@markdown 采样率\n",
|
306 |
+
"MODELSAMPLE = \"48k\" #@param {type:\"string\"}\n",
|
307 |
+
"#@markdown 是否缓存训练集\n",
|
308 |
+
"CACHEDATA = 1 #@param {type:\"integer\"}\n",
|
309 |
+
"#@markdown 是否仅保存最新的ckpt文件\n",
|
310 |
+
"ONLYLATEST = 0 #@param {type:\"integer\"}\n",
|
311 |
+
"\n",
|
312 |
+
"!python3 train_nsf_sim_cache_sid_load_pretrain.py -e lulu -sr {MODELSAMPLE} -f0 1 -bs {BATCHSIZE} -g {USEGPU} -te {MODELEPOCH} -se {EPOCHSAVE} -pg pretrained/f0G{MODELSAMPLE}.pth -pd pretrained/f0D{MODELSAMPLE}.pth -l {ONLYLATEST} -c {CACHEDATA}\n"
|
313 |
+
],
|
314 |
+
"metadata": {
|
315 |
+
"id": "IMLPLKOaKj58"
|
316 |
+
},
|
317 |
+
"execution_count": null,
|
318 |
+
"outputs": []
|
319 |
+
},
|
320 |
+
{
|
321 |
+
"cell_type": "code",
|
322 |
+
"source": [
|
323 |
+
"#@title 删除其它pth,只留选中的(慎点,仔细看代码)\n",
|
324 |
+
"#@markdown 模型名\n",
|
325 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
326 |
+
"#@markdown 选中模型epoch\n",
|
327 |
+
"MODELEPOCH = 9600 #@param {type:\"integer\"}\n",
|
328 |
+
"\n",
|
329 |
+
"!echo \"备份选中的模型。。。\"\n",
|
330 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth /content/{MODELNAME}_D_{MODELEPOCH}.pth\n",
|
331 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth /content/{MODELNAME}_G_{MODELEPOCH}.pth\n",
|
332 |
+
"\n",
|
333 |
+
"!echo \"正在删除。。。\"\n",
|
334 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}\n",
|
335 |
+
"!rm /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/*.pth\n",
|
336 |
+
"\n",
|
337 |
+
"!echo \"恢复选中的模型。。。\"\n",
|
338 |
+
"!mv /content/{MODELNAME}_D_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth \n",
|
339 |
+
"!mv /content/{MODELNAME}_G_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth\n",
|
340 |
+
"\n",
|
341 |
+
"!echo \"删除完成\"\n",
|
342 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}"
|
343 |
+
],
|
344 |
+
"metadata": {
|
345 |
+
"id": "haYA81hySuDl"
|
346 |
+
},
|
347 |
+
"execution_count": null,
|
348 |
+
"outputs": []
|
349 |
+
},
|
350 |
+
{
|
351 |
+
"cell_type": "code",
|
352 |
+
"source": [
|
353 |
+
"#@title 清除项目下所有文件,只留选中的模型(慎点,仔细看代码)\n",
|
354 |
+
"#@markdown 模型名\n",
|
355 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
356 |
+
"#@markdown 选中模型epoch\n",
|
357 |
+
"MODELEPOCH = 9600 #@param {type:\"integer\"}\n",
|
358 |
+
"\n",
|
359 |
+
"!echo \"备份选中的模型。。。\"\n",
|
360 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth /content/{MODELNAME}_D_{MODELEPOCH}.pth\n",
|
361 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth /content/{MODELNAME}_G_{MODELEPOCH}.pth\n",
|
362 |
+
"\n",
|
363 |
+
"!echo \"正在删除。。。\"\n",
|
364 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}\n",
|
365 |
+
"!rm -rf /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/*\n",
|
366 |
+
"\n",
|
367 |
+
"!echo \"恢复选中的模型。。。\"\n",
|
368 |
+
"!mv /content/{MODELNAME}_D_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth \n",
|
369 |
+
"!mv /content/{MODELNAME}_G_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth\n",
|
370 |
+
"\n",
|
371 |
+
"!echo \"删除完成\"\n",
|
372 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}"
|
373 |
+
],
|
374 |
+
"metadata": {
|
375 |
+
"id": "QhSiPTVPoIRh"
|
376 |
+
},
|
377 |
+
"execution_count": null,
|
378 |
+
"outputs": []
|
379 |
+
}
|
380 |
+
]
|
381 |
+
}
|
Retrieval_based_Voice_Conversion_WebUI_v2.ipynb
ADDED
@@ -0,0 +1,401 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"attachments": {},
|
5 |
+
"cell_type": "markdown",
|
6 |
+
"metadata": {
|
7 |
+
"id": "ZFFCx5J80SGa"
|
8 |
+
},
|
9 |
+
"source": [
|
10 |
+
"[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/Retrieval_based_Voice_Conversion_WebUI_v2.ipynb)"
|
11 |
+
]
|
12 |
+
},
|
13 |
+
{
|
14 |
+
"cell_type": "code",
|
15 |
+
"execution_count": null,
|
16 |
+
"metadata": {
|
17 |
+
"id": "GmFP6bN9dvOq"
|
18 |
+
},
|
19 |
+
"outputs": [],
|
20 |
+
"source": [
|
21 |
+
"#@title 查看显卡\n",
|
22 |
+
"!nvidia-smi"
|
23 |
+
]
|
24 |
+
},
|
25 |
+
{
|
26 |
+
"cell_type": "code",
|
27 |
+
"execution_count": null,
|
28 |
+
"metadata": {
|
29 |
+
"id": "wjddIFr1oS3W"
|
30 |
+
},
|
31 |
+
"outputs": [],
|
32 |
+
"source": [
|
33 |
+
"#@title 安装依赖\n",
|
34 |
+
"!apt-get -y install build-essential python3-dev ffmpeg\n",
|
35 |
+
"!pip3 install --upgrade setuptools wheel\n",
|
36 |
+
"!pip3 install --upgrade pip\n",
|
37 |
+
"!pip3 install faiss-cpu==1.7.2 fairseq gradio==3.14.0 ffmpeg ffmpeg-python praat-parselmouth pyworld numpy==1.23.5 numba==0.56.4 librosa==0.9.2"
|
38 |
+
]
|
39 |
+
},
|
40 |
+
{
|
41 |
+
"cell_type": "code",
|
42 |
+
"execution_count": null,
|
43 |
+
"metadata": {
|
44 |
+
"id": "ge_97mfpgqTm"
|
45 |
+
},
|
46 |
+
"outputs": [],
|
47 |
+
"source": [
|
48 |
+
"#@title 克隆仓库\n",
|
49 |
+
"\n",
|
50 |
+
"!mkdir Retrieval-based-Voice-Conversion-WebUI\n",
|
51 |
+
"%cd /content/Retrieval-based-Voice-Conversion-WebUI\n",
|
52 |
+
"!git init\n",
|
53 |
+
"!git remote add origin https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI.git\n",
|
54 |
+
"!git fetch origin cfd984812804ddc9247d65b14c82cd32e56c1133 --depth=1 \n",
|
55 |
+
"!git reset --hard FETCH_HEAD"
|
56 |
+
]
|
57 |
+
},
|
58 |
+
{
|
59 |
+
"cell_type": "code",
|
60 |
+
"execution_count": null,
|
61 |
+
"metadata": {
|
62 |
+
"id": "BLDEZADkvlw1"
|
63 |
+
},
|
64 |
+
"outputs": [],
|
65 |
+
"source": [
|
66 |
+
"#@title 更新仓库(一般无需执行)\n",
|
67 |
+
"!git pull"
|
68 |
+
]
|
69 |
+
},
|
70 |
+
{
|
71 |
+
"cell_type": "code",
|
72 |
+
"execution_count": null,
|
73 |
+
"metadata": {
|
74 |
+
"id": "pqE0PrnuRqI2"
|
75 |
+
},
|
76 |
+
"outputs": [],
|
77 |
+
"source": [
|
78 |
+
"#@title 安装aria2\n",
|
79 |
+
"!apt -y install -qq aria2"
|
80 |
+
]
|
81 |
+
},
|
82 |
+
{
|
83 |
+
"cell_type": "code",
|
84 |
+
"execution_count": null,
|
85 |
+
"metadata": {
|
86 |
+
"id": "UG3XpUwEomUz"
|
87 |
+
},
|
88 |
+
"outputs": [],
|
89 |
+
"source": [
|
90 |
+
"#@title 下载底模\n",
|
91 |
+
"\n",
|
92 |
+
"# v1\n",
|
93 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o D32k.pth\n",
|
94 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o D40k.pth\n",
|
95 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o D48k.pth\n",
|
96 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o G32k.pth\n",
|
97 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o G40k.pth\n",
|
98 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o G48k.pth\n",
|
99 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0D32k.pth\n",
|
100 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0D40k.pth\n",
|
101 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0D48k.pth\n",
|
102 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0G32k.pth\n",
|
103 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0G40k.pth\n",
|
104 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained -o f0G48k.pth\n",
|
105 |
+
"\n",
|
106 |
+
"#v2\n",
|
107 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/D32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o D32k.pth\n",
|
108 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/D40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o D40k.pth\n",
|
109 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/D48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o D48k.pth\n",
|
110 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/G32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o G32k.pth\n",
|
111 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/G40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o G40k.pth\n",
|
112 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/G48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o G48k.pth\n",
|
113 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0D32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o f0D32k.pth\n",
|
114 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0D40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o f0D40k.pth\n",
|
115 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0D48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o f0D48k.pth\n",
|
116 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0G32k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o f0G32k.pth\n",
|
117 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0G40k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o f0G40k.pth\n",
|
118 |
+
"# !aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0G48k.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/pretrained_v2 -o f0G48k.pth"
|
119 |
+
]
|
120 |
+
},
|
121 |
+
{
|
122 |
+
"cell_type": "code",
|
123 |
+
"execution_count": null,
|
124 |
+
"metadata": {
|
125 |
+
"id": "HugjmZqZRuiF"
|
126 |
+
},
|
127 |
+
"outputs": [],
|
128 |
+
"source": [
|
129 |
+
"#@title 下载人声分离模型\n",
|
130 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP2-人声vocals+非人声instrumentals.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/uvr5_weights -o HP2-人声vocals+非人声instrumentals.pth\n",
|
131 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP5-主旋律人声vocals+其他instrumentals.pth -d /content/Retrieval-based-Voice-Conversion-WebUI/uvr5_weights -o HP5-主旋律人声vocals+其他instrumentals.pth"
|
132 |
+
]
|
133 |
+
},
|
134 |
+
{
|
135 |
+
"cell_type": "code",
|
136 |
+
"execution_count": null,
|
137 |
+
"metadata": {
|
138 |
+
"id": "2RCaT9FTR0ej"
|
139 |
+
},
|
140 |
+
"outputs": [],
|
141 |
+
"source": [
|
142 |
+
"#@title 下载hubert_base\n",
|
143 |
+
"!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt -d /content/Retrieval-based-Voice-Conversion-WebUI -o hubert_base.pt"
|
144 |
+
]
|
145 |
+
},
|
146 |
+
{
|
147 |
+
"cell_type": "code",
|
148 |
+
"execution_count": null,
|
149 |
+
"metadata": {
|
150 |
+
"id": "jwu07JgqoFON"
|
151 |
+
},
|
152 |
+
"outputs": [],
|
153 |
+
"source": [
|
154 |
+
"#@title 挂载谷歌云盘\n",
|
155 |
+
"\n",
|
156 |
+
"from google.colab import drive\n",
|
157 |
+
"drive.mount('/content/drive')"
|
158 |
+
]
|
159 |
+
},
|
160 |
+
{
|
161 |
+
"cell_type": "code",
|
162 |
+
"execution_count": null,
|
163 |
+
"metadata": {
|
164 |
+
"id": "Mwk7Q0Loqzjx"
|
165 |
+
},
|
166 |
+
"outputs": [],
|
167 |
+
"source": [
|
168 |
+
"#@title 从谷歌云盘加载打包好的数据集到/content/dataset\n",
|
169 |
+
"\n",
|
170 |
+
"#@markdown 数据集位置\n",
|
171 |
+
"DATASET = \"/content/drive/MyDrive/dataset/lulu20230327_32k.zip\" #@param {type:\"string\"}\n",
|
172 |
+
"\n",
|
173 |
+
"!mkdir -p /content/dataset\n",
|
174 |
+
"!unzip -d /content/dataset -B {DATASET}"
|
175 |
+
]
|
176 |
+
},
|
177 |
+
{
|
178 |
+
"cell_type": "code",
|
179 |
+
"execution_count": null,
|
180 |
+
"metadata": {
|
181 |
+
"id": "PDlFxWHWEynD"
|
182 |
+
},
|
183 |
+
"outputs": [],
|
184 |
+
"source": [
|
185 |
+
"#@title 重命名数据集中的重名文件\n",
|
186 |
+
"!ls -a /content/dataset/\n",
|
187 |
+
"!rename 's/(\\w+)\\.(\\w+)~(\\d*)/$1_$3.$2/' /content/dataset/*.*~*"
|
188 |
+
]
|
189 |
+
},
|
190 |
+
{
|
191 |
+
"cell_type": "code",
|
192 |
+
"execution_count": null,
|
193 |
+
"metadata": {
|
194 |
+
"id": "7vh6vphDwO0b"
|
195 |
+
},
|
196 |
+
"outputs": [],
|
197 |
+
"source": [
|
198 |
+
"#@title 启动web\n",
|
199 |
+
"%cd /content/Retrieval-based-Voice-Conversion-WebUI\n",
|
200 |
+
"# %load_ext tensorboard\n",
|
201 |
+
"# %tensorboard --logdir /content/Retrieval-based-Voice-Conversion-WebUI/logs\n",
|
202 |
+
"!python3 infer-web.py --colab --pycmd python3"
|
203 |
+
]
|
204 |
+
},
|
205 |
+
{
|
206 |
+
"cell_type": "code",
|
207 |
+
"execution_count": null,
|
208 |
+
"metadata": {
|
209 |
+
"id": "FgJuNeAwx5Y_"
|
210 |
+
},
|
211 |
+
"outputs": [],
|
212 |
+
"source": [
|
213 |
+
"#@title 手动将训练后的模型文件备份到谷歌云盘\n",
|
214 |
+
"#@markdown 需要自己查看logs文件夹下模型的文件名,手动修改下方命令末尾的文件名\n",
|
215 |
+
"\n",
|
216 |
+
"#@markdown 模型名\n",
|
217 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
218 |
+
"#@markdown 模型epoch\n",
|
219 |
+
"MODELEPOCH = 9600 #@param {type:\"integer\"}\n",
|
220 |
+
"\n",
|
221 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth /content/drive/MyDrive/{MODELNAME}_D_{MODELEPOCH}.pth\n",
|
222 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth /content/drive/MyDrive/{MODELNAME}_G_{MODELEPOCH}.pth\n",
|
223 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/added_*.index /content/drive/MyDrive/\n",
|
224 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/total_*.npy /content/drive/MyDrive/\n",
|
225 |
+
"\n",
|
226 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/weights/{MODELNAME}.pth /content/drive/MyDrive/{MODELNAME}{MODELEPOCH}.pth"
|
227 |
+
]
|
228 |
+
},
|
229 |
+
{
|
230 |
+
"cell_type": "code",
|
231 |
+
"execution_count": null,
|
232 |
+
"metadata": {
|
233 |
+
"id": "OVQoLQJXS7WX"
|
234 |
+
},
|
235 |
+
"outputs": [],
|
236 |
+
"source": [
|
237 |
+
"#@title 从谷歌云盘恢复pth\n",
|
238 |
+
"#@markdown 需要自己查看logs文件夹下模型的文件名,手动修改下方命令末尾的文件名\n",
|
239 |
+
"\n",
|
240 |
+
"#@markdown 模型名\n",
|
241 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
242 |
+
"#@markdown 模型epoch\n",
|
243 |
+
"MODELEPOCH = 7500 #@param {type:\"integer\"}\n",
|
244 |
+
"\n",
|
245 |
+
"!mkdir -p /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}\n",
|
246 |
+
"\n",
|
247 |
+
"!cp /content/drive/MyDrive/{MODELNAME}_D_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth\n",
|
248 |
+
"!cp /content/drive/MyDrive/{MODELNAME}_G_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth\n",
|
249 |
+
"!cp /content/drive/MyDrive/*.index /content/\n",
|
250 |
+
"!cp /content/drive/MyDrive/*.npy /content/\n",
|
251 |
+
"!cp /content/drive/MyDrive/{MODELNAME}{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/weights/{MODELNAME}.pth"
|
252 |
+
]
|
253 |
+
},
|
254 |
+
{
|
255 |
+
"cell_type": "code",
|
256 |
+
"execution_count": null,
|
257 |
+
"metadata": {
|
258 |
+
"id": "ZKAyuKb9J6dz"
|
259 |
+
},
|
260 |
+
"outputs": [],
|
261 |
+
"source": [
|
262 |
+
"#@title 手动预处理(不推荐)\n",
|
263 |
+
"#@markdown 模型名\n",
|
264 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
265 |
+
"#@markdown 采样率\n",
|
266 |
+
"BITRATE = 48000 #@param {type:\"integer\"}\n",
|
267 |
+
"#@markdown 使用的进程数\n",
|
268 |
+
"THREADCOUNT = 8 #@param {type:\"integer\"}\n",
|
269 |
+
"\n",
|
270 |
+
"!python3 trainset_preprocess_pipeline_print.py /content/dataset {BITRATE} {THREADCOUNT} logs/{MODELNAME} True\n"
|
271 |
+
]
|
272 |
+
},
|
273 |
+
{
|
274 |
+
"cell_type": "code",
|
275 |
+
"execution_count": null,
|
276 |
+
"metadata": {
|
277 |
+
"id": "CrxJqzAUKmPJ"
|
278 |
+
},
|
279 |
+
"outputs": [],
|
280 |
+
"source": [
|
281 |
+
"#@title 手动提取特征(不推荐)\n",
|
282 |
+
"#@markdown 模型名\n",
|
283 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
284 |
+
"#@markdown 使用的进程数\n",
|
285 |
+
"THREADCOUNT = 8 #@param {type:\"integer\"}\n",
|
286 |
+
"#@markdown 音高提取算法\n",
|
287 |
+
"ALGO = \"harvest\" #@param {type:\"string\"}\n",
|
288 |
+
"\n",
|
289 |
+
"!python3 extract_f0_print.py logs/{MODELNAME} {THREADCOUNT} {ALGO}\n",
|
290 |
+
"\n",
|
291 |
+
"!python3 extract_feature_print.py cpu 1 0 0 logs/{MODELNAME}\n"
|
292 |
+
]
|
293 |
+
},
|
294 |
+
{
|
295 |
+
"cell_type": "code",
|
296 |
+
"execution_count": null,
|
297 |
+
"metadata": {
|
298 |
+
"id": "IMLPLKOaKj58"
|
299 |
+
},
|
300 |
+
"outputs": [],
|
301 |
+
"source": [
|
302 |
+
"#@title 手动训练(不推荐)\n",
|
303 |
+
"#@markdown 模型名\n",
|
304 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
305 |
+
"#@markdown 使用的GPU\n",
|
306 |
+
"USEGPU = \"0\" #@param {type:\"string\"}\n",
|
307 |
+
"#@markdown 批大小\n",
|
308 |
+
"BATCHSIZE = 32 #@param {type:\"integer\"}\n",
|
309 |
+
"#@markdown 停止的epoch\n",
|
310 |
+
"MODELEPOCH = 3200 #@param {type:\"integer\"}\n",
|
311 |
+
"#@markdown 保存epoch间隔\n",
|
312 |
+
"EPOCHSAVE = 100 #@param {type:\"integer\"}\n",
|
313 |
+
"#@markdown 采样率\n",
|
314 |
+
"MODELSAMPLE = \"48k\" #@param {type:\"string\"}\n",
|
315 |
+
"#@markdown 是否缓存训练集\n",
|
316 |
+
"CACHEDATA = 1 #@param {type:\"integer\"}\n",
|
317 |
+
"#@markdown 是否仅保存最新的ckpt文件\n",
|
318 |
+
"ONLYLATEST = 0 #@param {type:\"integer\"}\n",
|
319 |
+
"\n",
|
320 |
+
"!python3 train_nsf_sim_cache_sid_load_pretrain.py -e lulu -sr {MODELSAMPLE} -f0 1 -bs {BATCHSIZE} -g {USEGPU} -te {MODELEPOCH} -se {EPOCHSAVE} -pg pretrained/f0G{MODELSAMPLE}.pth -pd pretrained/f0D{MODELSAMPLE}.pth -l {ONLYLATEST} -c {CACHEDATA}\n"
|
321 |
+
]
|
322 |
+
},
|
323 |
+
{
|
324 |
+
"cell_type": "code",
|
325 |
+
"execution_count": null,
|
326 |
+
"metadata": {
|
327 |
+
"id": "haYA81hySuDl"
|
328 |
+
},
|
329 |
+
"outputs": [],
|
330 |
+
"source": [
|
331 |
+
"#@title 删除其它pth,只留选中的(慎点,仔细看代码)\n",
|
332 |
+
"#@markdown 模型名\n",
|
333 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
334 |
+
"#@markdown 选中模型epoch\n",
|
335 |
+
"MODELEPOCH = 9600 #@param {type:\"integer\"}\n",
|
336 |
+
"\n",
|
337 |
+
"!echo \"备份选中的模型。。。\"\n",
|
338 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth /content/{MODELNAME}_D_{MODELEPOCH}.pth\n",
|
339 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth /content/{MODELNAME}_G_{MODELEPOCH}.pth\n",
|
340 |
+
"\n",
|
341 |
+
"!echo \"正在删除。。。\"\n",
|
342 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}\n",
|
343 |
+
"!rm /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/*.pth\n",
|
344 |
+
"\n",
|
345 |
+
"!echo \"恢复选中的模型。。。\"\n",
|
346 |
+
"!mv /content/{MODELNAME}_D_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth \n",
|
347 |
+
"!mv /content/{MODELNAME}_G_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth\n",
|
348 |
+
"\n",
|
349 |
+
"!echo \"删除完成\"\n",
|
350 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}"
|
351 |
+
]
|
352 |
+
},
|
353 |
+
{
|
354 |
+
"cell_type": "code",
|
355 |
+
"execution_count": null,
|
356 |
+
"metadata": {
|
357 |
+
"id": "QhSiPTVPoIRh"
|
358 |
+
},
|
359 |
+
"outputs": [],
|
360 |
+
"source": [
|
361 |
+
"#@title 清除项目下所有文件,只留选中的模型(慎点,仔细看代码)\n",
|
362 |
+
"#@markdown 模型名\n",
|
363 |
+
"MODELNAME = \"lulu\" #@param {type:\"string\"}\n",
|
364 |
+
"#@markdown 选中模型epoch\n",
|
365 |
+
"MODELEPOCH = 9600 #@param {type:\"integer\"}\n",
|
366 |
+
"\n",
|
367 |
+
"!echo \"备份选中的模型。。。\"\n",
|
368 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth /content/{MODELNAME}_D_{MODELEPOCH}.pth\n",
|
369 |
+
"!cp /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth /content/{MODELNAME}_G_{MODELEPOCH}.pth\n",
|
370 |
+
"\n",
|
371 |
+
"!echo \"正在删除。。。\"\n",
|
372 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}\n",
|
373 |
+
"!rm -rf /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/*\n",
|
374 |
+
"\n",
|
375 |
+
"!echo \"恢复选中的模型。。。\"\n",
|
376 |
+
"!mv /content/{MODELNAME}_D_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/G_{MODELEPOCH}.pth \n",
|
377 |
+
"!mv /content/{MODELNAME}_G_{MODELEPOCH}.pth /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}/D_{MODELEPOCH}.pth\n",
|
378 |
+
"\n",
|
379 |
+
"!echo \"删除完成\"\n",
|
380 |
+
"!ls /content/Retrieval-based-Voice-Conversion-WebUI/logs/{MODELNAME}"
|
381 |
+
]
|
382 |
+
}
|
383 |
+
],
|
384 |
+
"metadata": {
|
385 |
+
"accelerator": "GPU",
|
386 |
+
"colab": {
|
387 |
+
"private_outputs": true,
|
388 |
+
"provenance": []
|
389 |
+
},
|
390 |
+
"gpuClass": "standard",
|
391 |
+
"kernelspec": {
|
392 |
+
"display_name": "Python 3",
|
393 |
+
"name": "python3"
|
394 |
+
},
|
395 |
+
"language_info": {
|
396 |
+
"name": "python"
|
397 |
+
}
|
398 |
+
},
|
399 |
+
"nbformat": 4,
|
400 |
+
"nbformat_minor": 0
|
401 |
+
}
|
app.py
ADDED
@@ -0,0 +1,322 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import io
|
2 |
+
import os
|
3 |
+
import torch
|
4 |
+
|
5 |
+
# os.system("wget -P cvec/ https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt")
|
6 |
+
import gradio as gr
|
7 |
+
import librosa
|
8 |
+
import numpy as np
|
9 |
+
import soundfile
|
10 |
+
import logging
|
11 |
+
from fairseq import checkpoint_utils
|
12 |
+
from my_utils import load_audio
|
13 |
+
from vc_infer_pipeline import VC
|
14 |
+
import traceback
|
15 |
+
from config import Config
|
16 |
+
from infer_pack.models import (
|
17 |
+
SynthesizerTrnMs256NSFsid,
|
18 |
+
SynthesizerTrnMs256NSFsid_nono,
|
19 |
+
SynthesizerTrnMs768NSFsid,
|
20 |
+
SynthesizerTrnMs768NSFsid_nono,
|
21 |
+
)
|
22 |
+
from i18n import I18nAuto
|
23 |
+
|
24 |
+
logging.getLogger("numba").setLevel(logging.WARNING)
|
25 |
+
logging.getLogger("markdown_it").setLevel(logging.WARNING)
|
26 |
+
logging.getLogger("urllib3").setLevel(logging.WARNING)
|
27 |
+
logging.getLogger("matplotlib").setLevel(logging.WARNING)
|
28 |
+
|
29 |
+
i18n = I18nAuto()
|
30 |
+
i18n.print()
|
31 |
+
|
32 |
+
config = Config()
|
33 |
+
|
34 |
+
weight_root = "weights"
|
35 |
+
weight_uvr5_root = "uvr5_weights"
|
36 |
+
index_root = "logs"
|
37 |
+
names = []
|
38 |
+
hubert_model = None
|
39 |
+
for name in os.listdir(weight_root):
|
40 |
+
if name.endswith(".pth"):
|
41 |
+
names.append(name)
|
42 |
+
index_paths = []
|
43 |
+
for root, dirs, files in os.walk(index_root, topdown=False):
|
44 |
+
for name in files:
|
45 |
+
if name.endswith(".index") and "trained" not in name:
|
46 |
+
index_paths.append("%s/%s" % (root, name))
|
47 |
+
|
48 |
+
|
49 |
+
def get_vc(sid):
|
50 |
+
global n_spk, tgt_sr, net_g, vc, cpt, version
|
51 |
+
if sid == "" or sid == []:
|
52 |
+
global hubert_model
|
53 |
+
if hubert_model != None: # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
|
54 |
+
print("clean_empty_cache")
|
55 |
+
del net_g, n_spk, vc, hubert_model, tgt_sr # ,cpt
|
56 |
+
hubert_model = net_g = n_spk = vc = hubert_model = tgt_sr = None
|
57 |
+
if torch.cuda.is_available():
|
58 |
+
torch.cuda.empty_cache()
|
59 |
+
###楼下不这么折腾清理不干净
|
60 |
+
if_f0 = cpt.get("f0", 1)
|
61 |
+
version = cpt.get("version", "v1")
|
62 |
+
if version == "v1":
|
63 |
+
if if_f0 == 1:
|
64 |
+
net_g = SynthesizerTrnMs256NSFsid(
|
65 |
+
*cpt["config"], is_half=config.is_half
|
66 |
+
)
|
67 |
+
else:
|
68 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
69 |
+
elif version == "v2":
|
70 |
+
if if_f0 == 1:
|
71 |
+
net_g = SynthesizerTrnMs768NSFsid(
|
72 |
+
*cpt["config"], is_half=config.is_half
|
73 |
+
)
|
74 |
+
else:
|
75 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
76 |
+
del net_g, cpt
|
77 |
+
if torch.cuda.is_available():
|
78 |
+
torch.cuda.empty_cache()
|
79 |
+
cpt = None
|
80 |
+
return {"visible": False, "__type__": "update"}
|
81 |
+
person = "%s/%s" % (weight_root, sid)
|
82 |
+
print("loading %s" % person)
|
83 |
+
cpt = torch.load(person, map_location="cpu")
|
84 |
+
tgt_sr = cpt["config"][-1]
|
85 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
|
86 |
+
if_f0 = cpt.get("f0", 1)
|
87 |
+
version = cpt.get("version", "v1")
|
88 |
+
if version == "v1":
|
89 |
+
if if_f0 == 1:
|
90 |
+
net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
|
91 |
+
else:
|
92 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
93 |
+
elif version == "v2":
|
94 |
+
if if_f0 == 1:
|
95 |
+
net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
|
96 |
+
else:
|
97 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
98 |
+
del net_g.enc_q
|
99 |
+
print(net_g.load_state_dict(cpt["weight"], strict=False))
|
100 |
+
net_g.eval().to(config.device)
|
101 |
+
if config.is_half:
|
102 |
+
net_g = net_g.half()
|
103 |
+
else:
|
104 |
+
net_g = net_g.float()
|
105 |
+
vc = VC(tgt_sr, config)
|
106 |
+
n_spk = cpt["config"][-3]
|
107 |
+
return {"visible": True, "maximum": n_spk, "__type__": "update"}
|
108 |
+
|
109 |
+
|
110 |
+
def load_hubert():
|
111 |
+
global hubert_model
|
112 |
+
models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
|
113 |
+
["hubert_base.pt"],
|
114 |
+
suffix="",
|
115 |
+
)
|
116 |
+
hubert_model = models[0]
|
117 |
+
hubert_model = hubert_model.to(config.device)
|
118 |
+
if config.is_half:
|
119 |
+
hubert_model = hubert_model.half()
|
120 |
+
else:
|
121 |
+
hubert_model = hubert_model.float()
|
122 |
+
hubert_model.eval()
|
123 |
+
|
124 |
+
|
125 |
+
def vc_single(
|
126 |
+
sid,
|
127 |
+
input_audio_path,
|
128 |
+
f0_up_key,
|
129 |
+
f0_file,
|
130 |
+
f0_method,
|
131 |
+
file_index,
|
132 |
+
file_index2,
|
133 |
+
# file_big_npy,
|
134 |
+
index_rate,
|
135 |
+
filter_radius,
|
136 |
+
resample_sr,
|
137 |
+
rms_mix_rate,
|
138 |
+
protect,
|
139 |
+
): # spk_item, input_audio0, vc_transform0,f0_file,f0method0
|
140 |
+
global tgt_sr, net_g, vc, hubert_model, version
|
141 |
+
if input_audio_path is None:
|
142 |
+
return "You need to upload an audio", None
|
143 |
+
f0_up_key = int(f0_up_key)
|
144 |
+
try:
|
145 |
+
audio = input_audio_path[1] / 32768.0
|
146 |
+
if len(audio.shape) == 2:
|
147 |
+
audio = np.mean(audio, -1)
|
148 |
+
audio = librosa.resample(audio, orig_sr=input_audio_path[0], target_sr=16000)
|
149 |
+
audio_max = np.abs(audio).max() / 0.95
|
150 |
+
if audio_max > 1:
|
151 |
+
audio /= audio_max
|
152 |
+
times = [0, 0, 0]
|
153 |
+
if hubert_model == None:
|
154 |
+
load_hubert()
|
155 |
+
if_f0 = cpt.get("f0", 1)
|
156 |
+
file_index = (
|
157 |
+
(
|
158 |
+
file_index.strip(" ")
|
159 |
+
.strip('"')
|
160 |
+
.strip("\n")
|
161 |
+
.strip('"')
|
162 |
+
.strip(" ")
|
163 |
+
.replace("trained", "added")
|
164 |
+
)
|
165 |
+
if file_index != ""
|
166 |
+
else file_index2
|
167 |
+
) # 防止小白写错,自动帮他替换掉
|
168 |
+
# file_big_npy = (
|
169 |
+
# file_big_npy.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
170 |
+
# )
|
171 |
+
audio_opt = vc.pipeline(
|
172 |
+
hubert_model,
|
173 |
+
net_g,
|
174 |
+
sid,
|
175 |
+
audio,
|
176 |
+
input_audio_path,
|
177 |
+
times,
|
178 |
+
f0_up_key,
|
179 |
+
f0_method,
|
180 |
+
file_index,
|
181 |
+
# file_big_npy,
|
182 |
+
index_rate,
|
183 |
+
if_f0,
|
184 |
+
filter_radius,
|
185 |
+
tgt_sr,
|
186 |
+
resample_sr,
|
187 |
+
rms_mix_rate,
|
188 |
+
version,
|
189 |
+
protect,
|
190 |
+
f0_file=f0_file,
|
191 |
+
)
|
192 |
+
if resample_sr >= 16000 and tgt_sr != resample_sr:
|
193 |
+
tgt_sr = resample_sr
|
194 |
+
index_info = (
|
195 |
+
"Using index:%s." % file_index
|
196 |
+
if os.path.exists(file_index)
|
197 |
+
else "Index not used."
|
198 |
+
)
|
199 |
+
return "Success.\n %s\nTime:\n npy:%ss, f0:%ss, infer:%ss" % (
|
200 |
+
index_info,
|
201 |
+
times[0],
|
202 |
+
times[1],
|
203 |
+
times[2],
|
204 |
+
), (tgt_sr, audio_opt)
|
205 |
+
except:
|
206 |
+
info = traceback.format_exc()
|
207 |
+
print(info)
|
208 |
+
return info, (None, None)
|
209 |
+
|
210 |
+
|
211 |
+
app = gr.Blocks()
|
212 |
+
with app:
|
213 |
+
with gr.Tabs():
|
214 |
+
with gr.TabItem("在线demo"):
|
215 |
+
gr.Markdown(
|
216 |
+
value="""
|
217 |
+
RVC 在线demo
|
218 |
+
"""
|
219 |
+
)
|
220 |
+
sid = gr.Dropdown(label=i18n("推理音色"), choices=sorted(names))
|
221 |
+
with gr.Column():
|
222 |
+
spk_item = gr.Slider(
|
223 |
+
minimum=0,
|
224 |
+
maximum=2333,
|
225 |
+
step=1,
|
226 |
+
label=i18n("请选择说话人id"),
|
227 |
+
value=0,
|
228 |
+
visible=False,
|
229 |
+
interactive=True,
|
230 |
+
)
|
231 |
+
sid.change(
|
232 |
+
fn=get_vc,
|
233 |
+
inputs=[sid],
|
234 |
+
outputs=[spk_item],
|
235 |
+
)
|
236 |
+
gr.Markdown(
|
237 |
+
value=i18n("男转女推荐+12key, 女转男推荐-12key, 如果音域爆炸导致音色失真也可以自己调整到合适音域. ")
|
238 |
+
)
|
239 |
+
vc_input3 = gr.Audio(label="上传音频(长度小于90秒)")
|
240 |
+
vc_transform0 = gr.Number(label=i18n("变调(整数, 半音数量, 升八度12降八度-12)"), value=0)
|
241 |
+
f0method0 = gr.Radio(
|
242 |
+
label=i18n("选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比,crepe效果好但吃GPU"),
|
243 |
+
choices=["pm", "harvest", "crepe"],
|
244 |
+
value="pm",
|
245 |
+
interactive=True,
|
246 |
+
)
|
247 |
+
filter_radius0 = gr.Slider(
|
248 |
+
minimum=0,
|
249 |
+
maximum=7,
|
250 |
+
label=i18n(">=3则使用对harvest音高识别的结果使用中值滤波,数值为滤波半径,使用可以削弱哑音"),
|
251 |
+
value=3,
|
252 |
+
step=1,
|
253 |
+
interactive=True,
|
254 |
+
)
|
255 |
+
with gr.Column():
|
256 |
+
file_index1 = gr.Textbox(
|
257 |
+
label=i18n("特征检索库文件路径,为空则使用下拉的选择结果"),
|
258 |
+
value="",
|
259 |
+
interactive=False,
|
260 |
+
visible=False,
|
261 |
+
)
|
262 |
+
file_index2 = gr.Dropdown(
|
263 |
+
label=i18n("自动检测index路径,下拉式选择(dropdown)"),
|
264 |
+
choices=sorted(index_paths),
|
265 |
+
interactive=True,
|
266 |
+
)
|
267 |
+
index_rate1 = gr.Slider(
|
268 |
+
minimum=0,
|
269 |
+
maximum=1,
|
270 |
+
label=i18n("检索特征占比"),
|
271 |
+
value=0.88,
|
272 |
+
interactive=True,
|
273 |
+
)
|
274 |
+
resample_sr0 = gr.Slider(
|
275 |
+
minimum=0,
|
276 |
+
maximum=48000,
|
277 |
+
label=i18n("后处理重采样至最终采样率,0为不进行重采样"),
|
278 |
+
value=0,
|
279 |
+
step=1,
|
280 |
+
interactive=True,
|
281 |
+
)
|
282 |
+
rms_mix_rate0 = gr.Slider(
|
283 |
+
minimum=0,
|
284 |
+
maximum=1,
|
285 |
+
label=i18n("输入源音量包络替换输出音量包络融合比例,越靠近1越使用输出包络"),
|
286 |
+
value=1,
|
287 |
+
interactive=True,
|
288 |
+
)
|
289 |
+
protect0 = gr.Slider(
|
290 |
+
minimum=0,
|
291 |
+
maximum=0.5,
|
292 |
+
label=i18n("保护清辅音和呼吸声,防止电音撕裂等artifact,拉满0.5不开启,调低加大保护力度但可能降低索引效果"),
|
293 |
+
value=0.33,
|
294 |
+
step=0.01,
|
295 |
+
interactive=True,
|
296 |
+
)
|
297 |
+
f0_file = gr.File(label=i18n("F0曲线文件, 可选, 一行一个音高, 代替默认F0及升降调"))
|
298 |
+
but0 = gr.Button(i18n("转换"), variant="primary")
|
299 |
+
vc_output1 = gr.Textbox(label=i18n("输出信息"))
|
300 |
+
vc_output2 = gr.Audio(label=i18n("输出音频(右下角三个点,点了可以下载)"))
|
301 |
+
but0.click(
|
302 |
+
vc_single,
|
303 |
+
[
|
304 |
+
spk_item,
|
305 |
+
vc_input3,
|
306 |
+
vc_transform0,
|
307 |
+
f0_file,
|
308 |
+
f0method0,
|
309 |
+
file_index1,
|
310 |
+
file_index2,
|
311 |
+
# file_big_npy1,
|
312 |
+
index_rate1,
|
313 |
+
filter_radius0,
|
314 |
+
resample_sr0,
|
315 |
+
rms_mix_rate0,
|
316 |
+
protect0,
|
317 |
+
],
|
318 |
+
[vc_output1, vc_output2],
|
319 |
+
)
|
320 |
+
|
321 |
+
|
322 |
+
app.launch()
|
config.py
ADDED
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import torch
|
3 |
+
from multiprocessing import cpu_count
|
4 |
+
|
5 |
+
|
6 |
+
def config_file_change_fp32():
|
7 |
+
for config_file in ["32k.json", "40k.json", "48k.json"]:
|
8 |
+
with open(f"configs/{config_file}", "r") as f:
|
9 |
+
strr = f.read().replace("true", "false")
|
10 |
+
with open(f"configs/{config_file}", "w") as f:
|
11 |
+
f.write(strr)
|
12 |
+
with open("trainset_preprocess_pipeline_print.py", "r") as f:
|
13 |
+
strr = f.read().replace("3.7", "3.0")
|
14 |
+
with open("trainset_preprocess_pipeline_print.py", "w") as f:
|
15 |
+
f.write(strr)
|
16 |
+
|
17 |
+
|
18 |
+
class Config:
|
19 |
+
def __init__(self):
|
20 |
+
self.device = "cuda:0"
|
21 |
+
self.is_half = True
|
22 |
+
self.n_cpu = 0
|
23 |
+
self.gpu_name = None
|
24 |
+
self.gpu_mem = None
|
25 |
+
|
26 |
+
self.python_cmd = "python3"
|
27 |
+
self.listen_port = 7865
|
28 |
+
self.iscolab = False
|
29 |
+
self.noparallel = False
|
30 |
+
self.noautoopen = False
|
31 |
+
|
32 |
+
self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
|
33 |
+
|
34 |
+
@staticmethod
|
35 |
+
def arg_parse() -> tuple:
|
36 |
+
parser = argparse.ArgumentParser()
|
37 |
+
parser.add_argument("--port", type=int, default=7865, help="Listen port")
|
38 |
+
parser.add_argument(
|
39 |
+
"--pycmd", type=str, default="python3", help="Python command"
|
40 |
+
)
|
41 |
+
parser.add_argument("--colab", action="store_true", help="Launch in colab")
|
42 |
+
parser.add_argument(
|
43 |
+
"--noparallel", action="store_true", help="Disable parallel processing"
|
44 |
+
)
|
45 |
+
parser.add_argument(
|
46 |
+
"--noautoopen",
|
47 |
+
action="store_true",
|
48 |
+
help="Do not open in browser automatically",
|
49 |
+
)
|
50 |
+
cmd_opts = parser.parse_args()
|
51 |
+
|
52 |
+
cmd_opts.port = cmd_opts.port if 0 <= cmd_opts.port <= 65535 else 7865
|
53 |
+
|
54 |
+
return (
|
55 |
+
cmd_opts.pycmd,
|
56 |
+
cmd_opts.port,
|
57 |
+
cmd_opts.colab,
|
58 |
+
cmd_opts.noparallel,
|
59 |
+
cmd_opts.noautoopen,
|
60 |
+
)
|
61 |
+
|
62 |
+
def device_config(self) -> tuple:
|
63 |
+
if torch.cuda.is_available():
|
64 |
+
i_device = int(self.device.split(":")[-1])
|
65 |
+
self.gpu_name = torch.cuda.get_device_name(i_device)
|
66 |
+
if (
|
67 |
+
("16" in self.gpu_name and "V100" not in self.gpu_name.upper())
|
68 |
+
or "P40" in self.gpu_name.upper()
|
69 |
+
or "1060" in self.gpu_name
|
70 |
+
or "1070" in self.gpu_name
|
71 |
+
or "1080" in self.gpu_name
|
72 |
+
):
|
73 |
+
print("16系/10系显卡和P40强制单精度")
|
74 |
+
self.is_half = False
|
75 |
+
config_file_change_fp32()
|
76 |
+
else:
|
77 |
+
self.gpu_name = None
|
78 |
+
self.gpu_mem = int(
|
79 |
+
torch.cuda.get_device_properties(i_device).total_memory
|
80 |
+
/ 1024
|
81 |
+
/ 1024
|
82 |
+
/ 1024
|
83 |
+
+ 0.4
|
84 |
+
)
|
85 |
+
if self.gpu_mem <= 4:
|
86 |
+
with open("trainset_preprocess_pipeline_print.py", "r") as f:
|
87 |
+
strr = f.read().replace("3.7", "3.0")
|
88 |
+
with open("trainset_preprocess_pipeline_print.py", "w") as f:
|
89 |
+
f.write(strr)
|
90 |
+
elif torch.backends.mps.is_available():
|
91 |
+
print("没有发现支持的N卡, 使用MPS进行推理")
|
92 |
+
self.device = "mps"
|
93 |
+
self.is_half = False
|
94 |
+
config_file_change_fp32()
|
95 |
+
else:
|
96 |
+
print("没有发现支持的N卡, 使用CPU进行推理")
|
97 |
+
self.device = "cpu"
|
98 |
+
self.is_half = False
|
99 |
+
config_file_change_fp32()
|
100 |
+
|
101 |
+
if self.n_cpu == 0:
|
102 |
+
self.n_cpu = cpu_count()
|
103 |
+
|
104 |
+
if self.is_half:
|
105 |
+
# 6G显存配置
|
106 |
+
x_pad = 3
|
107 |
+
x_query = 10
|
108 |
+
x_center = 60
|
109 |
+
x_max = 65
|
110 |
+
else:
|
111 |
+
# 5G显存配置
|
112 |
+
x_pad = 1
|
113 |
+
x_query = 6
|
114 |
+
x_center = 38
|
115 |
+
x_max = 41
|
116 |
+
|
117 |
+
if self.gpu_mem != None and self.gpu_mem <= 4:
|
118 |
+
x_pad = 1
|
119 |
+
x_query = 5
|
120 |
+
x_center = 30
|
121 |
+
x_max = 32
|
122 |
+
|
123 |
+
return x_pad, x_query, x_center, x_max
|
envfilescheck.bat
ADDED
@@ -0,0 +1,348 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
@echo off && chcp 65001
|
2 |
+
|
3 |
+
echo working dir is %cd%
|
4 |
+
echo downloading requirement aria2 check.
|
5 |
+
echo=
|
6 |
+
dir /a:d/b | findstr "aria2" > flag.txt
|
7 |
+
findstr "aria2" flag.txt >nul
|
8 |
+
if %errorlevel% ==0 (
|
9 |
+
echo aria2 checked.
|
10 |
+
echo=
|
11 |
+
) else (
|
12 |
+
echo failed. please downloading aria2 from webpage!
|
13 |
+
echo unzip it and put in this directory!
|
14 |
+
timeout /T 5
|
15 |
+
start https://github.com/aria2/aria2/releases/tag/release-1.36.0
|
16 |
+
echo=
|
17 |
+
goto end
|
18 |
+
)
|
19 |
+
|
20 |
+
echo envfiles checking start.
|
21 |
+
echo=
|
22 |
+
|
23 |
+
for /f %%x in ('findstr /i /c:"aria2" "flag.txt"') do (set aria2=%%x)&goto endSch
|
24 |
+
:endSch
|
25 |
+
|
26 |
+
set d32=f0D32k.pth
|
27 |
+
set d40=f0D40k.pth
|
28 |
+
set d48=f0D48k.pth
|
29 |
+
set g32=f0G32k.pth
|
30 |
+
set g40=f0G40k.pth
|
31 |
+
set g48=f0G48k.pth
|
32 |
+
|
33 |
+
set d40v2=f0D40k.pth
|
34 |
+
set g40v2=f0G40k.pth
|
35 |
+
|
36 |
+
set dld32=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D32k.pth
|
37 |
+
set dld40=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D40k.pth
|
38 |
+
set dld48=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0D48k.pth
|
39 |
+
set dlg32=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G32k.pth
|
40 |
+
set dlg40=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G40k.pth
|
41 |
+
set dlg48=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/f0G48k.pth
|
42 |
+
|
43 |
+
set dld40v2=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0D40k.pth
|
44 |
+
set dlg40v2=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/f0G40k.pth
|
45 |
+
|
46 |
+
set hp2_all=HP2_all_vocals.pth
|
47 |
+
set hp3_all=HP3_all_vocals.pth
|
48 |
+
set hp5_only=HP5_only_main_vocal.pth
|
49 |
+
set VR_DeEchoAggressive=VR-DeEchoAggressive.pth
|
50 |
+
set VR_DeEchoDeReverb=VR-DeEchoDeReverb.pth
|
51 |
+
set VR_DeEchoNormal=VR-DeEchoNormal.pth
|
52 |
+
set onnx_dereverb=vocals.onnx
|
53 |
+
|
54 |
+
set dlhp2_all=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP2_all_vocals.pth
|
55 |
+
set dlhp3_all=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP3_all_vocals.pth
|
56 |
+
set dlhp5_only=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/HP5_only_main_vocal.pth
|
57 |
+
set dlVR_DeEchoAggressive=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/VR-DeEchoAggressive.pth
|
58 |
+
set dlVR_DeEchoDeReverb=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/VR-DeEchoDeReverb.pth
|
59 |
+
set dlVR_DeEchoNormal=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/VR-DeEchoNormal.pth
|
60 |
+
set dlonnx_dereverb=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/uvr5_weights/onnx_dereverb_By_FoxJoy/vocals.onnx
|
61 |
+
|
62 |
+
set hb=hubert_base.pt
|
63 |
+
|
64 |
+
set dlhb=https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt
|
65 |
+
|
66 |
+
echo dir check start.
|
67 |
+
echo=
|
68 |
+
|
69 |
+
if exist "%~dp0pretrained" (
|
70 |
+
echo dir .\pretrained checked.
|
71 |
+
) else (
|
72 |
+
echo failed. generating dir .\pretrained.
|
73 |
+
mkdir pretrained
|
74 |
+
)
|
75 |
+
if exist "%~dp0pretrained_v2" (
|
76 |
+
echo dir .\pretrained_v2 checked.
|
77 |
+
) else (
|
78 |
+
echo failed. generating dir .\pretrained_v2.
|
79 |
+
mkdir pretrained_v2
|
80 |
+
)
|
81 |
+
if exist "%~dp0uvr5_weights" (
|
82 |
+
echo dir .\uvr5_weights checked.
|
83 |
+
) else (
|
84 |
+
echo failed. generating dir .\uvr5_weights.
|
85 |
+
mkdir uvr5_weights
|
86 |
+
)
|
87 |
+
if exist "%~dp0uvr5_weights\onnx_dereverb_By_FoxJoy" (
|
88 |
+
echo dir .\uvr5_weights\onnx_dereverb_By_FoxJoy checked.
|
89 |
+
) else (
|
90 |
+
echo failed. generating dir .\uvr5_weights\onnx_dereverb_By_FoxJoy.
|
91 |
+
mkdir uvr5_weights\onnx_dereverb_By_FoxJoy
|
92 |
+
)
|
93 |
+
|
94 |
+
echo=
|
95 |
+
echo dir check finished.
|
96 |
+
|
97 |
+
echo=
|
98 |
+
echo required files check start.
|
99 |
+
|
100 |
+
echo checking D32k.pth
|
101 |
+
if exist "%~dp0pretrained\D32k.pth" (
|
102 |
+
echo D32k.pth in .\pretrained checked.
|
103 |
+
echo=
|
104 |
+
) else (
|
105 |
+
echo failed. starting download from huggingface.
|
106 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D32k.pth -d %~dp0pretrained -o D32k.pth
|
107 |
+
if exist "%~dp0pretrained\D32k.pth" (echo download successful.) else (echo please try again!
|
108 |
+
echo=)
|
109 |
+
)
|
110 |
+
echo checking D40k.pth
|
111 |
+
if exist "%~dp0pretrained\D40k.pth" (
|
112 |
+
echo D40k.pth in .\pretrained checked.
|
113 |
+
echo=
|
114 |
+
) else (
|
115 |
+
echo failed. starting download from huggingface.
|
116 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D40k.pth -d %~dp0pretrained -o D40k.pth
|
117 |
+
if exist "%~dp0pretrained\D40k.pth" (echo download successful.) else (echo please try again!
|
118 |
+
echo=)
|
119 |
+
)
|
120 |
+
echo checking D40k.pth
|
121 |
+
if exist "%~dp0pretrained_v2\D40k.pth" (
|
122 |
+
echo D40k.pth in .\pretrained_v2 checked.
|
123 |
+
echo=
|
124 |
+
) else (
|
125 |
+
echo failed. starting download from huggingface.
|
126 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/D40k.pth -d %~dp0pretrained_v2 -o D40k.pth
|
127 |
+
if exist "%~dp0pretrained_v2\D40k.pth" (echo download successful.) else (echo please try again!
|
128 |
+
echo=)
|
129 |
+
)
|
130 |
+
echo checking D48k.pth
|
131 |
+
if exist "%~dp0pretrained\D48k.pth" (
|
132 |
+
echo D48k.pth in .\pretrained checked.
|
133 |
+
echo=
|
134 |
+
) else (
|
135 |
+
echo failed. starting download from huggingface.
|
136 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/D48k.pth -d %~dp0pretrained -o D48k.pth
|
137 |
+
if exist "%~dp0pretrained\D48k.pth" (echo download successful.) else (echo please try again!
|
138 |
+
echo=)
|
139 |
+
)
|
140 |
+
echo checking G32k.pth
|
141 |
+
if exist "%~dp0pretrained\G32k.pth" (
|
142 |
+
echo G32k.pth in .\pretrained checked.
|
143 |
+
echo=
|
144 |
+
) else (
|
145 |
+
echo failed. starting download from huggingface.
|
146 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G32k.pth -d %~dp0pretrained -o G32k.pth
|
147 |
+
if exist "%~dp0pretrained\G32k.pth" (echo download successful.) else (echo please try again!
|
148 |
+
echo=)
|
149 |
+
)
|
150 |
+
echo checking G40k.pth
|
151 |
+
if exist "%~dp0pretrained\G40k.pth" (
|
152 |
+
echo G40k.pth in .\pretrained checked.
|
153 |
+
echo=
|
154 |
+
) else (
|
155 |
+
echo failed. starting download from huggingface.
|
156 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G40k.pth -d %~dp0pretrained -o G40k.pth
|
157 |
+
if exist "%~dp0pretrained\G40k.pth" (echo download successful.) else (echo please try again!
|
158 |
+
echo=)
|
159 |
+
)
|
160 |
+
echo checking G40k.pth
|
161 |
+
if exist "%~dp0pretrained_v2\G40k.pth" (
|
162 |
+
echo G40k.pth in .\pretrained_v2 checked.
|
163 |
+
echo=
|
164 |
+
) else (
|
165 |
+
echo failed. starting download from huggingface.
|
166 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained_v2/G40k.pth -d %~dp0pretrained_v2 -o G40k.pth
|
167 |
+
if exist "%~dp0pretrained_v2\G40k.pth" (echo download successful.) else (echo please try again!
|
168 |
+
echo=)
|
169 |
+
)
|
170 |
+
echo checking G48k.pth
|
171 |
+
if exist "%~dp0pretrained\G48k.pth" (
|
172 |
+
echo G48k.pth in .\pretrained checked.
|
173 |
+
echo=
|
174 |
+
) else (
|
175 |
+
echo failed. starting download from huggingface.
|
176 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/pretrained/G48k.pth -d %~dp0pretrained -o G48k.pth
|
177 |
+
if exist "%~dp0pretrained\G48k.pth" (echo download successful.) else (echo please try again!
|
178 |
+
echo=)
|
179 |
+
)
|
180 |
+
|
181 |
+
echo checking %d32%
|
182 |
+
if exist "%~dp0pretrained\%d32%" (
|
183 |
+
echo %d32% in .\pretrained checked.
|
184 |
+
echo=
|
185 |
+
) else (
|
186 |
+
echo failed. starting download from huggingface.
|
187 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dld32% -d %~dp0pretrained -o %d32%
|
188 |
+
if exist "%~dp0pretrained\%d32%" (echo download successful.) else (echo please try again!
|
189 |
+
echo=)
|
190 |
+
)
|
191 |
+
echo checking %d40%
|
192 |
+
if exist "%~dp0pretrained\%d40%" (
|
193 |
+
echo %d40% in .\pretrained checked.
|
194 |
+
echo=
|
195 |
+
) else (
|
196 |
+
echo failed. starting download from huggingface.
|
197 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dld40% -d %~dp0pretrained -o %d40%
|
198 |
+
if exist "%~dp0pretrained\%d40%" (echo download successful.) else (echo please try again!
|
199 |
+
echo=)
|
200 |
+
)
|
201 |
+
echo checking %d40v2%
|
202 |
+
if exist "%~dp0pretrained_v2\%d40v2%" (
|
203 |
+
echo %d40v2% in .\pretrained_v2 checked.
|
204 |
+
echo=
|
205 |
+
) else (
|
206 |
+
echo failed. starting download from huggingface.
|
207 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dld40v2% -d %~dp0pretrained_v2 -o %d40v2%
|
208 |
+
if exist "%~dp0pretrained_v2\%d40v2%" (echo download successful.) else (echo please try again!
|
209 |
+
echo=)
|
210 |
+
)
|
211 |
+
echo checking %d48%
|
212 |
+
if exist "%~dp0pretrained\%d48%" (
|
213 |
+
echo %d48% in .\pretrained checked.
|
214 |
+
echo=
|
215 |
+
) else (
|
216 |
+
echo failed. starting download from huggingface.
|
217 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dld48% -d %~dp0pretrained -o %d48%
|
218 |
+
if exist "%~dp0pretrained\%d48%" (echo download successful.) else (echo please try again!
|
219 |
+
echo=)
|
220 |
+
)
|
221 |
+
echo checking %g32%
|
222 |
+
if exist "%~dp0pretrained\%g32%" (
|
223 |
+
echo %g32% in .\pretrained checked.
|
224 |
+
echo=
|
225 |
+
) else (
|
226 |
+
echo failed. starting download from huggingface.
|
227 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlg32% -d %~dp0pretrained -o %g32%
|
228 |
+
if exist "%~dp0pretrained\%g32%" (echo download successful.) else (echo please try again!
|
229 |
+
echo=)
|
230 |
+
)
|
231 |
+
echo checking %g40%
|
232 |
+
if exist "%~dp0pretrained\%g40%" (
|
233 |
+
echo %g40% in .\pretrained checked.
|
234 |
+
echo=
|
235 |
+
) else (
|
236 |
+
echo failed. starting download from huggingface.
|
237 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlg40% -d %~dp0pretrained -o %g40%
|
238 |
+
if exist "%~dp0pretrained\%g40%" (echo download successful.) else (echo please try again!
|
239 |
+
echo=)
|
240 |
+
)
|
241 |
+
echo checking %g40v2%
|
242 |
+
if exist "%~dp0pretrained_v2\%g40v2%" (
|
243 |
+
echo %g40v2% in .\pretrained_v2 checked.
|
244 |
+
echo=
|
245 |
+
) else (
|
246 |
+
echo failed. starting download from huggingface.
|
247 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlg40v2% -d %~dp0pretrained_v2 -o %g40v2%
|
248 |
+
if exist "%~dp0pretrained_v2\%g40v2%" (echo download successful.) else (echo please try again!
|
249 |
+
echo=)
|
250 |
+
)
|
251 |
+
echo checking %g48%
|
252 |
+
if exist "%~dp0pretrained\%g48%" (
|
253 |
+
echo %g48% in .\pretrained checked.
|
254 |
+
echo=
|
255 |
+
) else (
|
256 |
+
echo failed. starting download from huggingface.
|
257 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlg48% -d %~dp0\pretrained -o %g48%
|
258 |
+
if exist "%~dp0pretrained\%g48%" (echo download successful.) else (echo please try again!
|
259 |
+
echo=)
|
260 |
+
)
|
261 |
+
|
262 |
+
echo checking %hp2_all%
|
263 |
+
if exist "%~dp0uvr5_weights\%hp2_all%" (
|
264 |
+
echo %hp2_all% in .\uvr5_weights checked.
|
265 |
+
echo=
|
266 |
+
) else (
|
267 |
+
echo failed. starting download from huggingface.
|
268 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlhp2_all% -d %~dp0\uvr5_weights -o %hp2_all%
|
269 |
+
if exist "%~dp0uvr5_weights\%hp2_all%" (echo download successful.) else (echo please try again!
|
270 |
+
echo=)
|
271 |
+
)
|
272 |
+
echo checking %hp3_all%
|
273 |
+
if exist "%~dp0uvr5_weights\%hp3_all%" (
|
274 |
+
echo %hp3_all% in .\uvr5_weights checked.
|
275 |
+
echo=
|
276 |
+
) else (
|
277 |
+
echo failed. starting download from huggingface.
|
278 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlhp3_all% -d %~dp0\uvr5_weights -o %hp3_all%
|
279 |
+
if exist "%~dp0uvr5_weights\%hp3_all%" (echo download successful.) else (echo please try again!
|
280 |
+
echo=)
|
281 |
+
)
|
282 |
+
echo checking %hp5_only%
|
283 |
+
if exist "%~dp0uvr5_weights\%hp5_only%" (
|
284 |
+
echo %hp5_only% in .\uvr5_weights checked.
|
285 |
+
echo=
|
286 |
+
) else (
|
287 |
+
echo failed. starting download from huggingface.
|
288 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlhp5_only% -d %~dp0\uvr5_weights -o %hp5_only%
|
289 |
+
if exist "%~dp0uvr5_weights\%hp5_only%" (echo download successful.) else (echo please try again!
|
290 |
+
echo=)
|
291 |
+
)
|
292 |
+
echo checking %VR_DeEchoAggressive%
|
293 |
+
if exist "%~dp0uvr5_weights\%VR_DeEchoAggressive%" (
|
294 |
+
echo %VR_DeEchoAggressive% in .\uvr5_weights checked.
|
295 |
+
echo=
|
296 |
+
) else (
|
297 |
+
echo failed. starting download from huggingface.
|
298 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlVR_DeEchoAggressive% -d %~dp0\uvr5_weights -o %VR_DeEchoAggressive%
|
299 |
+
if exist "%~dp0uvr5_weights\%VR_DeEchoAggressive%" (echo download successful.) else (echo please try again!
|
300 |
+
echo=)
|
301 |
+
)
|
302 |
+
echo checking %VR_DeEchoDeReverb%
|
303 |
+
if exist "%~dp0uvr5_weights\%VR_DeEchoDeReverb%" (
|
304 |
+
echo %VR_DeEchoDeReverb% in .\uvr5_weights checked.
|
305 |
+
echo=
|
306 |
+
) else (
|
307 |
+
echo failed. starting download from huggingface.
|
308 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlVR_DeEchoDeReverb% -d %~dp0\uvr5_weights -o %VR_DeEchoDeReverb%
|
309 |
+
if exist "%~dp0uvr5_weights\%VR_DeEchoDeReverb%" (echo download successful.) else (echo please try again!
|
310 |
+
echo=)
|
311 |
+
)
|
312 |
+
echo checking %VR_DeEchoNormal%
|
313 |
+
if exist "%~dp0uvr5_weights\%VR_DeEchoNormal%" (
|
314 |
+
echo %VR_DeEchoNormal% in .\uvr5_weights checked.
|
315 |
+
echo=
|
316 |
+
) else (
|
317 |
+
echo failed. starting download from huggingface.
|
318 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlVR_DeEchoNormal% -d %~dp0\uvr5_weights -o %VR_DeEchoNormal%
|
319 |
+
if exist "%~dp0uvr5_weights\%VR_DeEchoNormal%" (echo download successful.) else (echo please try again!
|
320 |
+
echo=)
|
321 |
+
)
|
322 |
+
echo checking %onnx_dereverb%
|
323 |
+
if exist "%~dp0uvr5_weights\onnx_dereverb_By_FoxJoy\%onnx_dereverb%" (
|
324 |
+
echo %onnx_dereverb% in .\uvr5_weights\onnx_dereverb_By_FoxJoy checked.
|
325 |
+
echo=
|
326 |
+
) else (
|
327 |
+
echo failed. starting download from huggingface.
|
328 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlonnx_dereverb% -d %~dp0\uvr5_weights\onnx_dereverb_By_FoxJoy -o %onnx_dereverb%
|
329 |
+
if exist "%~dp0uvr5_weights\onnx_dereverb_By_FoxJoy\%onnx_dereverb%" (echo download successful.) else (echo please try again!
|
330 |
+
echo=)
|
331 |
+
)
|
332 |
+
|
333 |
+
echo checking %hb%
|
334 |
+
if exist "%~dp0%hb%" (
|
335 |
+
echo %hb% in .\pretrained checked.
|
336 |
+
echo=
|
337 |
+
) else (
|
338 |
+
echo failed. starting download from huggingface.
|
339 |
+
%~dp0%aria2%\aria2c --console-log-level=error -c -x 16 -s 16 -k 1M %dlhb% -d %~dp0 -o %hb%
|
340 |
+
if exist "%~dp0%hb%" (echo download successful.) else (echo please try again!
|
341 |
+
echo=)
|
342 |
+
)
|
343 |
+
|
344 |
+
echo required files check finished.
|
345 |
+
echo envfiles check complete.
|
346 |
+
pause
|
347 |
+
:end
|
348 |
+
del flag.txt
|
export_onnx.py
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from infer_pack.models_onnx import SynthesizerTrnMsNSFsidM
|
2 |
+
import torch
|
3 |
+
|
4 |
+
if __name__ == "__main__":
|
5 |
+
MoeVS = True # 模型是否为MoeVoiceStudio(原MoeSS)使用
|
6 |
+
|
7 |
+
ModelPath = "Shiroha/shiroha.pth" # 模型路径
|
8 |
+
ExportedPath = "model.onnx" # 输出路径
|
9 |
+
hidden_channels = 256 # hidden_channels,为768Vec做准备
|
10 |
+
cpt = torch.load(ModelPath, map_location="cpu")
|
11 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
|
12 |
+
print(*cpt["config"])
|
13 |
+
|
14 |
+
test_phone = torch.rand(1, 200, hidden_channels) # hidden unit
|
15 |
+
test_phone_lengths = torch.tensor([200]).long() # hidden unit 长度(貌似没啥用)
|
16 |
+
test_pitch = torch.randint(size=(1, 200), low=5, high=255) # 基频(单位赫兹)
|
17 |
+
test_pitchf = torch.rand(1, 200) # nsf基频
|
18 |
+
test_ds = torch.LongTensor([0]) # 说话人ID
|
19 |
+
test_rnd = torch.rand(1, 192, 200) # 噪声(加入随机因子)
|
20 |
+
|
21 |
+
device = "cpu" # 导出时设备(不影响使用模型)
|
22 |
+
|
23 |
+
net_g = SynthesizerTrnMsNSFsidM(
|
24 |
+
*cpt["config"], is_half=False
|
25 |
+
) # fp32导出(C++要支持fp16必须手动将内存重新排列所以暂时不用fp16)
|
26 |
+
net_g.load_state_dict(cpt["weight"], strict=False)
|
27 |
+
input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
|
28 |
+
output_names = [
|
29 |
+
"audio",
|
30 |
+
]
|
31 |
+
# net_g.construct_spkmixmap(n_speaker) 多角色混合轨道导出
|
32 |
+
torch.onnx.export(
|
33 |
+
net_g,
|
34 |
+
(
|
35 |
+
test_phone.to(device),
|
36 |
+
test_phone_lengths.to(device),
|
37 |
+
test_pitch.to(device),
|
38 |
+
test_pitchf.to(device),
|
39 |
+
test_ds.to(device),
|
40 |
+
test_rnd.to(device),
|
41 |
+
),
|
42 |
+
ExportedPath,
|
43 |
+
dynamic_axes={
|
44 |
+
"phone": [1],
|
45 |
+
"pitch": [1],
|
46 |
+
"pitchf": [1],
|
47 |
+
"rnd": [2],
|
48 |
+
},
|
49 |
+
do_constant_folding=False,
|
50 |
+
opset_version=16,
|
51 |
+
verbose=False,
|
52 |
+
input_names=input_names,
|
53 |
+
output_names=output_names,
|
54 |
+
)
|
extract_f0_print.py
ADDED
@@ -0,0 +1,160 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os, traceback, sys, parselmouth
|
2 |
+
|
3 |
+
now_dir = os.getcwd()
|
4 |
+
sys.path.append(now_dir)
|
5 |
+
from my_utils import load_audio
|
6 |
+
import pyworld
|
7 |
+
from scipy.io import wavfile
|
8 |
+
import numpy as np, logging
|
9 |
+
|
10 |
+
logging.getLogger("numba").setLevel(logging.WARNING)
|
11 |
+
from multiprocessing import Process
|
12 |
+
|
13 |
+
exp_dir = sys.argv[1]
|
14 |
+
f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
|
15 |
+
|
16 |
+
|
17 |
+
def printt(strr):
|
18 |
+
print(strr)
|
19 |
+
f.write("%s\n" % strr)
|
20 |
+
f.flush()
|
21 |
+
|
22 |
+
|
23 |
+
n_p = int(sys.argv[2])
|
24 |
+
f0method = sys.argv[3]
|
25 |
+
|
26 |
+
|
27 |
+
class FeatureInput(object):
|
28 |
+
def __init__(self, samplerate=16000, hop_size=160):
|
29 |
+
self.fs = samplerate
|
30 |
+
self.hop = hop_size
|
31 |
+
|
32 |
+
self.f0_bin = 256
|
33 |
+
self.f0_max = 1100.0
|
34 |
+
self.f0_min = 50.0
|
35 |
+
self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
|
36 |
+
self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
|
37 |
+
|
38 |
+
def compute_f0(self, path, f0_method):
|
39 |
+
x = load_audio(path, self.fs)
|
40 |
+
p_len = x.shape[0] // self.hop
|
41 |
+
if f0_method == "pm":
|
42 |
+
time_step = 160 / 16000 * 1000
|
43 |
+
f0_min = 50
|
44 |
+
f0_max = 1100
|
45 |
+
f0 = (
|
46 |
+
parselmouth.Sound(x, self.fs)
|
47 |
+
.to_pitch_ac(
|
48 |
+
time_step=time_step / 1000,
|
49 |
+
voicing_threshold=0.6,
|
50 |
+
pitch_floor=f0_min,
|
51 |
+
pitch_ceiling=f0_max,
|
52 |
+
)
|
53 |
+
.selected_array["frequency"]
|
54 |
+
)
|
55 |
+
pad_size = (p_len - len(f0) + 1) // 2
|
56 |
+
if pad_size > 0 or p_len - len(f0) - pad_size > 0:
|
57 |
+
f0 = np.pad(
|
58 |
+
f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
|
59 |
+
)
|
60 |
+
elif f0_method == "harvest":
|
61 |
+
f0, t = pyworld.harvest(
|
62 |
+
x.astype(np.double),
|
63 |
+
fs=self.fs,
|
64 |
+
f0_ceil=self.f0_max,
|
65 |
+
f0_floor=self.f0_min,
|
66 |
+
frame_period=1000 * self.hop / self.fs,
|
67 |
+
)
|
68 |
+
f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
|
69 |
+
elif f0_method == "dio":
|
70 |
+
f0, t = pyworld.dio(
|
71 |
+
x.astype(np.double),
|
72 |
+
fs=self.fs,
|
73 |
+
f0_ceil=self.f0_max,
|
74 |
+
f0_floor=self.f0_min,
|
75 |
+
frame_period=1000 * self.hop / self.fs,
|
76 |
+
)
|
77 |
+
f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
|
78 |
+
return f0
|
79 |
+
|
80 |
+
def coarse_f0(self, f0):
|
81 |
+
f0_mel = 1127 * np.log(1 + f0 / 700)
|
82 |
+
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
|
83 |
+
self.f0_bin - 2
|
84 |
+
) / (self.f0_mel_max - self.f0_mel_min) + 1
|
85 |
+
|
86 |
+
# use 0 or 1
|
87 |
+
f0_mel[f0_mel <= 1] = 1
|
88 |
+
f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
|
89 |
+
f0_coarse = np.rint(f0_mel).astype(int)
|
90 |
+
assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
|
91 |
+
f0_coarse.max(),
|
92 |
+
f0_coarse.min(),
|
93 |
+
)
|
94 |
+
return f0_coarse
|
95 |
+
|
96 |
+
def go(self, paths, f0_method):
|
97 |
+
if len(paths) == 0:
|
98 |
+
printt("no-f0-todo")
|
99 |
+
else:
|
100 |
+
printt("todo-f0-%s" % len(paths))
|
101 |
+
n = max(len(paths) // 5, 1) # 每个进程最多打印5条
|
102 |
+
for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
|
103 |
+
try:
|
104 |
+
if idx % n == 0:
|
105 |
+
printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
|
106 |
+
if (
|
107 |
+
os.path.exists(opt_path1 + ".npy") == True
|
108 |
+
and os.path.exists(opt_path2 + ".npy") == True
|
109 |
+
):
|
110 |
+
continue
|
111 |
+
featur_pit = self.compute_f0(inp_path, f0_method)
|
112 |
+
np.save(
|
113 |
+
opt_path2,
|
114 |
+
featur_pit,
|
115 |
+
allow_pickle=False,
|
116 |
+
) # nsf
|
117 |
+
coarse_pit = self.coarse_f0(featur_pit)
|
118 |
+
np.save(
|
119 |
+
opt_path1,
|
120 |
+
coarse_pit,
|
121 |
+
allow_pickle=False,
|
122 |
+
) # ori
|
123 |
+
except:
|
124 |
+
printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
|
125 |
+
|
126 |
+
|
127 |
+
if __name__ == "__main__":
|
128 |
+
# exp_dir=r"E:\codes\py39\dataset\mi-test"
|
129 |
+
# n_p=16
|
130 |
+
# f = open("%s/log_extract_f0.log"%exp_dir, "w")
|
131 |
+
printt(sys.argv)
|
132 |
+
featureInput = FeatureInput()
|
133 |
+
paths = []
|
134 |
+
inp_root = "%s/1_16k_wavs" % (exp_dir)
|
135 |
+
opt_root1 = "%s/2a_f0" % (exp_dir)
|
136 |
+
opt_root2 = "%s/2b-f0nsf" % (exp_dir)
|
137 |
+
|
138 |
+
os.makedirs(opt_root1, exist_ok=True)
|
139 |
+
os.makedirs(opt_root2, exist_ok=True)
|
140 |
+
for name in sorted(list(os.listdir(inp_root))):
|
141 |
+
inp_path = "%s/%s" % (inp_root, name)
|
142 |
+
if "spec" in inp_path:
|
143 |
+
continue
|
144 |
+
opt_path1 = "%s/%s" % (opt_root1, name)
|
145 |
+
opt_path2 = "%s/%s" % (opt_root2, name)
|
146 |
+
paths.append([inp_path, opt_path1, opt_path2])
|
147 |
+
|
148 |
+
ps = []
|
149 |
+
for i in range(n_p):
|
150 |
+
p = Process(
|
151 |
+
target=featureInput.go,
|
152 |
+
args=(
|
153 |
+
paths[i::n_p],
|
154 |
+
f0method,
|
155 |
+
),
|
156 |
+
)
|
157 |
+
ps.append(p)
|
158 |
+
p.start()
|
159 |
+
for i in range(n_p):
|
160 |
+
ps[i].join()
|
extract_feature_print.py
ADDED
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os, sys, traceback
|
2 |
+
|
3 |
+
# device=sys.argv[1]
|
4 |
+
n_part = int(sys.argv[2])
|
5 |
+
i_part = int(sys.argv[3])
|
6 |
+
if len(sys.argv) == 5:
|
7 |
+
exp_dir = sys.argv[4]
|
8 |
+
version = sys.argv[5]
|
9 |
+
else:
|
10 |
+
i_gpu = sys.argv[4]
|
11 |
+
exp_dir = sys.argv[5]
|
12 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
|
13 |
+
version = sys.argv[6]
|
14 |
+
import torch
|
15 |
+
import torch.nn.functional as F
|
16 |
+
import soundfile as sf
|
17 |
+
import numpy as np
|
18 |
+
from fairseq import checkpoint_utils
|
19 |
+
|
20 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
21 |
+
|
22 |
+
if torch.cuda.is_available():
|
23 |
+
device = "cuda"
|
24 |
+
elif torch.backends.mps.is_available():
|
25 |
+
device = "mps"
|
26 |
+
else:
|
27 |
+
device = "cpu"
|
28 |
+
|
29 |
+
f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
|
30 |
+
|
31 |
+
|
32 |
+
def printt(strr):
|
33 |
+
print(strr)
|
34 |
+
f.write("%s\n" % strr)
|
35 |
+
f.flush()
|
36 |
+
|
37 |
+
|
38 |
+
printt(sys.argv)
|
39 |
+
model_path = "hubert_base.pt"
|
40 |
+
|
41 |
+
printt(exp_dir)
|
42 |
+
wavPath = "%s/1_16k_wavs" % exp_dir
|
43 |
+
outPath = (
|
44 |
+
"%s/3_feature256" % exp_dir if version == "v1" else "%s/3_feature768" % exp_dir
|
45 |
+
)
|
46 |
+
os.makedirs(outPath, exist_ok=True)
|
47 |
+
|
48 |
+
|
49 |
+
# wave must be 16k, hop_size=320
|
50 |
+
def readwave(wav_path, normalize=False):
|
51 |
+
wav, sr = sf.read(wav_path)
|
52 |
+
assert sr == 16000
|
53 |
+
feats = torch.from_numpy(wav).float()
|
54 |
+
if feats.dim() == 2: # double channels
|
55 |
+
feats = feats.mean(-1)
|
56 |
+
assert feats.dim() == 1, feats.dim()
|
57 |
+
if normalize:
|
58 |
+
with torch.no_grad():
|
59 |
+
feats = F.layer_norm(feats, feats.shape)
|
60 |
+
feats = feats.view(1, -1)
|
61 |
+
return feats
|
62 |
+
|
63 |
+
|
64 |
+
# HuBERT model
|
65 |
+
printt("load model(s) from {}".format(model_path))
|
66 |
+
# if hubert model is exist
|
67 |
+
if os.access(model_path, os.F_OK) == False:
|
68 |
+
printt(
|
69 |
+
"Error: Extracting is shut down because %s does not exist, you may download it from https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main"
|
70 |
+
% model_path
|
71 |
+
)
|
72 |
+
exit(0)
|
73 |
+
models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
|
74 |
+
[model_path],
|
75 |
+
suffix="",
|
76 |
+
)
|
77 |
+
model = models[0]
|
78 |
+
model = model.to(device)
|
79 |
+
printt("move model to %s" % device)
|
80 |
+
if device not in ["mps", "cpu"]:
|
81 |
+
model = model.half()
|
82 |
+
model.eval()
|
83 |
+
|
84 |
+
todo = sorted(list(os.listdir(wavPath)))[i_part::n_part]
|
85 |
+
n = max(1, len(todo) // 10) # 最多打印十条
|
86 |
+
if len(todo) == 0:
|
87 |
+
printt("no-feature-todo")
|
88 |
+
else:
|
89 |
+
printt("all-feature-%s" % len(todo))
|
90 |
+
for idx, file in enumerate(todo):
|
91 |
+
try:
|
92 |
+
if file.endswith(".wav"):
|
93 |
+
wav_path = "%s/%s" % (wavPath, file)
|
94 |
+
out_path = "%s/%s" % (outPath, file.replace("wav", "npy"))
|
95 |
+
|
96 |
+
if os.path.exists(out_path):
|
97 |
+
continue
|
98 |
+
|
99 |
+
feats = readwave(wav_path, normalize=saved_cfg.task.normalize)
|
100 |
+
padding_mask = torch.BoolTensor(feats.shape).fill_(False)
|
101 |
+
inputs = {
|
102 |
+
"source": feats.half().to(device)
|
103 |
+
if device not in ["mps", "cpu"]
|
104 |
+
else feats.to(device),
|
105 |
+
"padding_mask": padding_mask.to(device),
|
106 |
+
"output_layer": 9 if version == "v1" else 12, # layer 9
|
107 |
+
}
|
108 |
+
with torch.no_grad():
|
109 |
+
logits = model.extract_features(**inputs)
|
110 |
+
feats = (
|
111 |
+
model.final_proj(logits[0]) if version == "v1" else logits[0]
|
112 |
+
)
|
113 |
+
|
114 |
+
feats = feats.squeeze(0).float().cpu().numpy()
|
115 |
+
if np.isnan(feats).sum() == 0:
|
116 |
+
np.save(out_path, feats, allow_pickle=False)
|
117 |
+
else:
|
118 |
+
printt("%s-contains nan" % file)
|
119 |
+
if idx % n == 0:
|
120 |
+
printt("now-%s,all-%s,%s,%s" % (len(todo), idx, file, feats.shape))
|
121 |
+
except:
|
122 |
+
printt(traceback.format_exc())
|
123 |
+
printt("all-feature-done")
|
extract_locale.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import re
|
3 |
+
|
4 |
+
# Define regular expression patterns
|
5 |
+
pattern = r"""i18n\([\s\n\t]*(["'][^"']+["'])[\s\n\t]*\)"""
|
6 |
+
|
7 |
+
# Initialize the dictionary to store key-value pairs
|
8 |
+
data = {}
|
9 |
+
|
10 |
+
|
11 |
+
def process(fn: str):
|
12 |
+
global data
|
13 |
+
with open(fn, "r", encoding="utf-8") as f:
|
14 |
+
contents = f.read()
|
15 |
+
matches = re.findall(pattern, contents)
|
16 |
+
for key in matches:
|
17 |
+
key = eval(key)
|
18 |
+
print("extract:", key)
|
19 |
+
data[key] = key
|
20 |
+
|
21 |
+
|
22 |
+
print("processing infer-web.py")
|
23 |
+
process("infer-web.py")
|
24 |
+
|
25 |
+
print("processing gui.py")
|
26 |
+
process("gui.py")
|
27 |
+
|
28 |
+
# Save as a JSON file
|
29 |
+
with open("./i18n/zh_CN.json", "w", encoding="utf-8") as f:
|
30 |
+
json.dump(data, f, ensure_ascii=False, indent=4)
|
31 |
+
f.write("\n")
|
ffmpeg
ADDED
Binary file (302 kB). View file
|
|
go-realtime-gui.bat
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
runtime\python.exe gui.py
|
2 |
+
pause
|
go-web.bat
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
runtime\python.exe infer-web.py --pycmd runtime\python.exe --port 7897
|
2 |
+
pause
|
gui.py
ADDED
@@ -0,0 +1,698 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
0416后的更新:
|
3 |
+
引入config中half
|
4 |
+
重建npy而不用填写
|
5 |
+
v2支持
|
6 |
+
无f0模型支持
|
7 |
+
修复
|
8 |
+
|
9 |
+
int16:
|
10 |
+
增加无索引支持
|
11 |
+
f0算法改harvest(怎么看就只有这个会影响CPU占用),但是不这么改效果不好
|
12 |
+
"""
|
13 |
+
import os, sys, traceback, re
|
14 |
+
|
15 |
+
import json
|
16 |
+
|
17 |
+
now_dir = os.getcwd()
|
18 |
+
sys.path.append(now_dir)
|
19 |
+
from config import Config
|
20 |
+
|
21 |
+
Config = Config()
|
22 |
+
import PySimpleGUI as sg
|
23 |
+
import sounddevice as sd
|
24 |
+
import noisereduce as nr
|
25 |
+
import numpy as np
|
26 |
+
from fairseq import checkpoint_utils
|
27 |
+
import librosa, torch, pyworld, faiss, time, threading
|
28 |
+
import torch.nn.functional as F
|
29 |
+
import torchaudio.transforms as tat
|
30 |
+
import scipy.signal as signal
|
31 |
+
|
32 |
+
|
33 |
+
# import matplotlib.pyplot as plt
|
34 |
+
from infer_pack.models import (
|
35 |
+
SynthesizerTrnMs256NSFsid,
|
36 |
+
SynthesizerTrnMs256NSFsid_nono,
|
37 |
+
SynthesizerTrnMs768NSFsid,
|
38 |
+
SynthesizerTrnMs768NSFsid_nono,
|
39 |
+
)
|
40 |
+
from i18n import I18nAuto
|
41 |
+
|
42 |
+
i18n = I18nAuto()
|
43 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
44 |
+
current_dir = os.getcwd()
|
45 |
+
|
46 |
+
|
47 |
+
class RVC:
|
48 |
+
def __init__(
|
49 |
+
self, key, hubert_path, pth_path, index_path, npy_path, index_rate
|
50 |
+
) -> None:
|
51 |
+
"""
|
52 |
+
初始化
|
53 |
+
"""
|
54 |
+
try:
|
55 |
+
self.f0_up_key = key
|
56 |
+
self.time_step = 160 / 16000 * 1000
|
57 |
+
self.f0_min = 50
|
58 |
+
self.f0_max = 1100
|
59 |
+
self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
|
60 |
+
self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
|
61 |
+
self.sr = 16000
|
62 |
+
self.window = 160
|
63 |
+
if index_rate != 0:
|
64 |
+
self.index = faiss.read_index(index_path)
|
65 |
+
# self.big_npy = np.load(npy_path)
|
66 |
+
self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
|
67 |
+
print("index search enabled")
|
68 |
+
self.index_rate = index_rate
|
69 |
+
model_path = hubert_path
|
70 |
+
print("load model(s) from {}".format(model_path))
|
71 |
+
models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
|
72 |
+
[model_path],
|
73 |
+
suffix="",
|
74 |
+
)
|
75 |
+
self.model = models[0]
|
76 |
+
self.model = self.model.to(device)
|
77 |
+
if Config.is_half:
|
78 |
+
self.model = self.model.half()
|
79 |
+
else:
|
80 |
+
self.model = self.model.float()
|
81 |
+
self.model.eval()
|
82 |
+
cpt = torch.load(pth_path, map_location="cpu")
|
83 |
+
self.tgt_sr = cpt["config"][-1]
|
84 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
|
85 |
+
self.if_f0 = cpt.get("f0", 1)
|
86 |
+
self.version = cpt.get("version", "v1")
|
87 |
+
if self.version == "v1":
|
88 |
+
if self.if_f0 == 1:
|
89 |
+
self.net_g = SynthesizerTrnMs256NSFsid(
|
90 |
+
*cpt["config"], is_half=Config.is_half
|
91 |
+
)
|
92 |
+
else:
|
93 |
+
self.net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
94 |
+
elif self.version == "v2":
|
95 |
+
if self.if_f0 == 1:
|
96 |
+
self.net_g = SynthesizerTrnMs768NSFsid(
|
97 |
+
*cpt["config"], is_half=Config.is_half
|
98 |
+
)
|
99 |
+
else:
|
100 |
+
self.net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
101 |
+
del self.net_g.enc_q
|
102 |
+
print(self.net_g.load_state_dict(cpt["weight"], strict=False))
|
103 |
+
self.net_g.eval().to(device)
|
104 |
+
if Config.is_half:
|
105 |
+
self.net_g = self.net_g.half()
|
106 |
+
else:
|
107 |
+
self.net_g = self.net_g.float()
|
108 |
+
except:
|
109 |
+
print(traceback.format_exc())
|
110 |
+
|
111 |
+
def get_f0(self, x, f0_up_key, inp_f0=None):
|
112 |
+
x_pad = 1
|
113 |
+
f0_min = 50
|
114 |
+
f0_max = 1100
|
115 |
+
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
|
116 |
+
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
|
117 |
+
f0, t = pyworld.harvest(
|
118 |
+
x.astype(np.double),
|
119 |
+
fs=self.sr,
|
120 |
+
f0_ceil=f0_max,
|
121 |
+
f0_floor=f0_min,
|
122 |
+
frame_period=10,
|
123 |
+
)
|
124 |
+
f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)
|
125 |
+
f0 = signal.medfilt(f0, 3)
|
126 |
+
f0 *= pow(2, f0_up_key / 12)
|
127 |
+
# with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
|
128 |
+
tf0 = self.sr // self.window # 每秒f0点数
|
129 |
+
if inp_f0 is not None:
|
130 |
+
delta_t = np.round(
|
131 |
+
(inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
|
132 |
+
).astype("int16")
|
133 |
+
replace_f0 = np.interp(
|
134 |
+
list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
|
135 |
+
)
|
136 |
+
shape = f0[x_pad * tf0 : x_pad * tf0 + len(replace_f0)].shape[0]
|
137 |
+
f0[x_pad * tf0 : x_pad * tf0 + len(replace_f0)] = replace_f0[:shape]
|
138 |
+
# with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
|
139 |
+
f0bak = f0.copy()
|
140 |
+
f0_mel = 1127 * np.log(1 + f0 / 700)
|
141 |
+
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
|
142 |
+
f0_mel_max - f0_mel_min
|
143 |
+
) + 1
|
144 |
+
f0_mel[f0_mel <= 1] = 1
|
145 |
+
f0_mel[f0_mel > 255] = 255
|
146 |
+
f0_coarse = np.rint(f0_mel).astype(np.int)
|
147 |
+
return f0_coarse, f0bak # 1-0
|
148 |
+
|
149 |
+
def infer(self, feats: torch.Tensor) -> np.ndarray:
|
150 |
+
"""
|
151 |
+
推理函数
|
152 |
+
"""
|
153 |
+
audio = feats.clone().cpu().numpy()
|
154 |
+
assert feats.dim() == 1, feats.dim()
|
155 |
+
feats = feats.view(1, -1)
|
156 |
+
padding_mask = torch.BoolTensor(feats.shape).fill_(False)
|
157 |
+
if Config.is_half:
|
158 |
+
feats = feats.half()
|
159 |
+
else:
|
160 |
+
feats = feats.float()
|
161 |
+
inputs = {
|
162 |
+
"source": feats.to(device),
|
163 |
+
"padding_mask": padding_mask.to(device),
|
164 |
+
"output_layer": 9 if self.version == "v1" else 12,
|
165 |
+
}
|
166 |
+
torch.cuda.synchronize()
|
167 |
+
with torch.no_grad():
|
168 |
+
logits = self.model.extract_features(**inputs)
|
169 |
+
feats = (
|
170 |
+
self.model.final_proj(logits[0]) if self.version == "v1" else logits[0]
|
171 |
+
)
|
172 |
+
|
173 |
+
####索引优化
|
174 |
+
try:
|
175 |
+
if (
|
176 |
+
hasattr(self, "index")
|
177 |
+
and hasattr(self, "big_npy")
|
178 |
+
and self.index_rate != 0
|
179 |
+
):
|
180 |
+
npy = feats[0].cpu().numpy().astype("float32")
|
181 |
+
score, ix = self.index.search(npy, k=8)
|
182 |
+
weight = np.square(1 / score)
|
183 |
+
weight /= weight.sum(axis=1, keepdims=True)
|
184 |
+
npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
|
185 |
+
if Config.is_half:
|
186 |
+
npy = npy.astype("float16")
|
187 |
+
feats = (
|
188 |
+
torch.from_numpy(npy).unsqueeze(0).to(device) * self.index_rate
|
189 |
+
+ (1 - self.index_rate) * feats
|
190 |
+
)
|
191 |
+
else:
|
192 |
+
print("index search FAIL or disabled")
|
193 |
+
except:
|
194 |
+
traceback.print_exc()
|
195 |
+
print("index search FAIL")
|
196 |
+
feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
|
197 |
+
torch.cuda.synchronize()
|
198 |
+
print(feats.shape)
|
199 |
+
if self.if_f0 == 1:
|
200 |
+
pitch, pitchf = self.get_f0(audio, self.f0_up_key)
|
201 |
+
p_len = min(feats.shape[1], 13000, pitch.shape[0]) # 太大了爆显存
|
202 |
+
else:
|
203 |
+
pitch, pitchf = None, None
|
204 |
+
p_len = min(feats.shape[1], 13000) # 太大了爆显存
|
205 |
+
torch.cuda.synchronize()
|
206 |
+
# print(feats.shape,pitch.shape)
|
207 |
+
feats = feats[:, :p_len, :]
|
208 |
+
if self.if_f0 == 1:
|
209 |
+
pitch = pitch[:p_len]
|
210 |
+
pitchf = pitchf[:p_len]
|
211 |
+
pitch = torch.LongTensor(pitch).unsqueeze(0).to(device)
|
212 |
+
pitchf = torch.FloatTensor(pitchf).unsqueeze(0).to(device)
|
213 |
+
p_len = torch.LongTensor([p_len]).to(device)
|
214 |
+
ii = 0 # sid
|
215 |
+
sid = torch.LongTensor([ii]).to(device)
|
216 |
+
with torch.no_grad():
|
217 |
+
if self.if_f0 == 1:
|
218 |
+
infered_audio = (
|
219 |
+
self.net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0]
|
220 |
+
.data.cpu()
|
221 |
+
.float()
|
222 |
+
)
|
223 |
+
else:
|
224 |
+
infered_audio = (
|
225 |
+
self.net_g.infer(feats, p_len, sid)[0][0, 0].data.cpu().float()
|
226 |
+
)
|
227 |
+
torch.cuda.synchronize()
|
228 |
+
return infered_audio
|
229 |
+
|
230 |
+
|
231 |
+
class GUIConfig:
|
232 |
+
def __init__(self) -> None:
|
233 |
+
self.hubert_path: str = ""
|
234 |
+
self.pth_path: str = ""
|
235 |
+
self.index_path: str = ""
|
236 |
+
self.npy_path: str = ""
|
237 |
+
self.pitch: int = 12
|
238 |
+
self.samplerate: int = 44100
|
239 |
+
self.block_time: float = 1.0 # s
|
240 |
+
self.buffer_num: int = 1
|
241 |
+
self.threhold: int = -30
|
242 |
+
self.crossfade_time: float = 0.08
|
243 |
+
self.extra_time: float = 0.04
|
244 |
+
self.I_noise_reduce = False
|
245 |
+
self.O_noise_reduce = False
|
246 |
+
self.index_rate = 0.3
|
247 |
+
|
248 |
+
|
249 |
+
class GUI:
|
250 |
+
def __init__(self) -> None:
|
251 |
+
self.config = GUIConfig()
|
252 |
+
self.flag_vc = False
|
253 |
+
|
254 |
+
self.launcher()
|
255 |
+
|
256 |
+
def load(self):
|
257 |
+
input_devices, output_devices, _, _ = self.get_devices()
|
258 |
+
try:
|
259 |
+
with open("values1.json", "r") as j:
|
260 |
+
data = json.load(j)
|
261 |
+
except:
|
262 |
+
with open("values1.json", "w") as j:
|
263 |
+
data = {
|
264 |
+
"pth_path": " ",
|
265 |
+
"index_path": " ",
|
266 |
+
"sg_input_device": input_devices[sd.default.device[0]],
|
267 |
+
"sg_output_device": output_devices[sd.default.device[1]],
|
268 |
+
"threhold": "-45",
|
269 |
+
"pitch": "0",
|
270 |
+
"index_rate": "0",
|
271 |
+
"block_time": "1",
|
272 |
+
"crossfade_length": "0.04",
|
273 |
+
"extra_time": "1",
|
274 |
+
}
|
275 |
+
return data
|
276 |
+
|
277 |
+
def launcher(self):
|
278 |
+
data = self.load()
|
279 |
+
sg.theme("LightBlue3")
|
280 |
+
input_devices, output_devices, _, _ = self.get_devices()
|
281 |
+
layout = [
|
282 |
+
[
|
283 |
+
sg.Frame(
|
284 |
+
title=i18n("加载模型"),
|
285 |
+
layout=[
|
286 |
+
[
|
287 |
+
sg.Input(
|
288 |
+
default_text="hubert_base.pt",
|
289 |
+
key="hubert_path",
|
290 |
+
disabled=True,
|
291 |
+
),
|
292 |
+
sg.FileBrowse(
|
293 |
+
i18n("Hubert模型"),
|
294 |
+
initial_folder=os.path.join(os.getcwd()),
|
295 |
+
file_types=((". pt"),),
|
296 |
+
),
|
297 |
+
],
|
298 |
+
[
|
299 |
+
sg.Input(
|
300 |
+
default_text=data.get("pth_path", ""),
|
301 |
+
key="pth_path",
|
302 |
+
),
|
303 |
+
sg.FileBrowse(
|
304 |
+
i18n("选择.pth文件"),
|
305 |
+
initial_folder=os.path.join(os.getcwd(), "weights"),
|
306 |
+
file_types=((". pth"),),
|
307 |
+
),
|
308 |
+
],
|
309 |
+
[
|
310 |
+
sg.Input(
|
311 |
+
default_text=data.get("index_path", ""),
|
312 |
+
key="index_path",
|
313 |
+
),
|
314 |
+
sg.FileBrowse(
|
315 |
+
i18n("选择.index文件"),
|
316 |
+
initial_folder=os.path.join(os.getcwd(), "logs"),
|
317 |
+
file_types=((". index"),),
|
318 |
+
),
|
319 |
+
],
|
320 |
+
[
|
321 |
+
sg.Input(
|
322 |
+
default_text="你不需要填写这个You don't need write this.",
|
323 |
+
key="npy_path",
|
324 |
+
disabled=True,
|
325 |
+
),
|
326 |
+
sg.FileBrowse(
|
327 |
+
i18n("选择.npy文件"),
|
328 |
+
initial_folder=os.path.join(os.getcwd(), "logs"),
|
329 |
+
file_types=((". npy"),),
|
330 |
+
),
|
331 |
+
],
|
332 |
+
],
|
333 |
+
)
|
334 |
+
],
|
335 |
+
[
|
336 |
+
sg.Frame(
|
337 |
+
layout=[
|
338 |
+
[
|
339 |
+
sg.Text(i18n("输入设备")),
|
340 |
+
sg.Combo(
|
341 |
+
input_devices,
|
342 |
+
key="sg_input_device",
|
343 |
+
default_value=data.get("sg_input_device", ""),
|
344 |
+
),
|
345 |
+
],
|
346 |
+
[
|
347 |
+
sg.Text(i18n("输出设备")),
|
348 |
+
sg.Combo(
|
349 |
+
output_devices,
|
350 |
+
key="sg_output_device",
|
351 |
+
default_value=data.get("sg_output_device", ""),
|
352 |
+
),
|
353 |
+
],
|
354 |
+
],
|
355 |
+
title=i18n("音频设备(请使用同种类驱动)"),
|
356 |
+
)
|
357 |
+
],
|
358 |
+
[
|
359 |
+
sg.Frame(
|
360 |
+
layout=[
|
361 |
+
[
|
362 |
+
sg.Text(i18n("响应阈值")),
|
363 |
+
sg.Slider(
|
364 |
+
range=(-60, 0),
|
365 |
+
key="threhold",
|
366 |
+
resolution=1,
|
367 |
+
orientation="h",
|
368 |
+
default_value=data.get("threhold", ""),
|
369 |
+
),
|
370 |
+
],
|
371 |
+
[
|
372 |
+
sg.Text(i18n("音调设置")),
|
373 |
+
sg.Slider(
|
374 |
+
range=(-24, 24),
|
375 |
+
key="pitch",
|
376 |
+
resolution=1,
|
377 |
+
orientation="h",
|
378 |
+
default_value=data.get("pitch", ""),
|
379 |
+
),
|
380 |
+
],
|
381 |
+
[
|
382 |
+
sg.Text(i18n("Index Rate")),
|
383 |
+
sg.Slider(
|
384 |
+
range=(0.0, 1.0),
|
385 |
+
key="index_rate",
|
386 |
+
resolution=0.01,
|
387 |
+
orientation="h",
|
388 |
+
default_value=data.get("index_rate", ""),
|
389 |
+
),
|
390 |
+
],
|
391 |
+
],
|
392 |
+
title=i18n("常规设置"),
|
393 |
+
),
|
394 |
+
sg.Frame(
|
395 |
+
layout=[
|
396 |
+
[
|
397 |
+
sg.Text(i18n("采样长度")),
|
398 |
+
sg.Slider(
|
399 |
+
range=(0.1, 3.0),
|
400 |
+
key="block_time",
|
401 |
+
resolution=0.1,
|
402 |
+
orientation="h",
|
403 |
+
default_value=data.get("block_time", ""),
|
404 |
+
),
|
405 |
+
],
|
406 |
+
[
|
407 |
+
sg.Text(i18n("淡入淡出长度")),
|
408 |
+
sg.Slider(
|
409 |
+
range=(0.01, 0.15),
|
410 |
+
key="crossfade_length",
|
411 |
+
resolution=0.01,
|
412 |
+
orientation="h",
|
413 |
+
default_value=data.get("crossfade_length", ""),
|
414 |
+
),
|
415 |
+
],
|
416 |
+
[
|
417 |
+
sg.Text(i18n("额外推理时长")),
|
418 |
+
sg.Slider(
|
419 |
+
range=(0.05, 3.00),
|
420 |
+
key="extra_time",
|
421 |
+
resolution=0.01,
|
422 |
+
orientation="h",
|
423 |
+
default_value=data.get("extra_time", ""),
|
424 |
+
),
|
425 |
+
],
|
426 |
+
[
|
427 |
+
sg.Checkbox(i18n("输入降噪"), key="I_noise_reduce"),
|
428 |
+
sg.Checkbox(i18n("输出降噪"), key="O_noise_reduce"),
|
429 |
+
],
|
430 |
+
],
|
431 |
+
title=i18n("性能设置"),
|
432 |
+
),
|
433 |
+
],
|
434 |
+
[
|
435 |
+
sg.Button(i18n("开始音频转换"), key="start_vc"),
|
436 |
+
sg.Button(i18n("停止音频转换"), key="stop_vc"),
|
437 |
+
sg.Text(i18n("推理时间(ms):")),
|
438 |
+
sg.Text("0", key="infer_time"),
|
439 |
+
],
|
440 |
+
]
|
441 |
+
self.window = sg.Window("RVC - GUI", layout=layout)
|
442 |
+
self.event_handler()
|
443 |
+
|
444 |
+
def event_handler(self):
|
445 |
+
while True:
|
446 |
+
event, values = self.window.read()
|
447 |
+
if event == sg.WINDOW_CLOSED:
|
448 |
+
self.flag_vc = False
|
449 |
+
exit()
|
450 |
+
if event == "start_vc" and self.flag_vc == False:
|
451 |
+
if self.set_values(values) == True:
|
452 |
+
print("using_cuda:" + str(torch.cuda.is_available()))
|
453 |
+
self.start_vc()
|
454 |
+
settings = {
|
455 |
+
"pth_path": values["pth_path"],
|
456 |
+
"index_path": values["index_path"],
|
457 |
+
"sg_input_device": values["sg_input_device"],
|
458 |
+
"sg_output_device": values["sg_output_device"],
|
459 |
+
"threhold": values["threhold"],
|
460 |
+
"pitch": values["pitch"],
|
461 |
+
"index_rate": values["index_rate"],
|
462 |
+
"block_time": values["block_time"],
|
463 |
+
"crossfade_length": values["crossfade_length"],
|
464 |
+
"extra_time": values["extra_time"],
|
465 |
+
}
|
466 |
+
with open("values1.json", "w") as j:
|
467 |
+
json.dump(settings, j)
|
468 |
+
if event == "stop_vc" and self.flag_vc == True:
|
469 |
+
self.flag_vc = False
|
470 |
+
|
471 |
+
def set_values(self, values):
|
472 |
+
if len(values["pth_path"].strip()) == 0:
|
473 |
+
sg.popup(i18n("请选择pth文件"))
|
474 |
+
return False
|
475 |
+
if len(values["index_path"].strip()) == 0:
|
476 |
+
sg.popup(i18n("请选择index文件"))
|
477 |
+
return False
|
478 |
+
pattern = re.compile("[^\x00-\x7F]+")
|
479 |
+
if pattern.findall(values["hubert_path"]):
|
480 |
+
sg.popup(i18n("hubert模型路径不可包含中文"))
|
481 |
+
return False
|
482 |
+
if pattern.findall(values["pth_path"]):
|
483 |
+
sg.popup(i18n("pth文件路径不可包含中文"))
|
484 |
+
return False
|
485 |
+
if pattern.findall(values["index_path"]):
|
486 |
+
sg.popup(i18n("index文件路径不可包含中文"))
|
487 |
+
return False
|
488 |
+
self.set_devices(values["sg_input_device"], values["sg_output_device"])
|
489 |
+
self.config.hubert_path = os.path.join(current_dir, "hubert_base.pt")
|
490 |
+
self.config.pth_path = values["pth_path"]
|
491 |
+
self.config.index_path = values["index_path"]
|
492 |
+
self.config.npy_path = values["npy_path"]
|
493 |
+
self.config.threhold = values["threhold"]
|
494 |
+
self.config.pitch = values["pitch"]
|
495 |
+
self.config.block_time = values["block_time"]
|
496 |
+
self.config.crossfade_time = values["crossfade_length"]
|
497 |
+
self.config.extra_time = values["extra_time"]
|
498 |
+
self.config.I_noise_reduce = values["I_noise_reduce"]
|
499 |
+
self.config.O_noise_reduce = values["O_noise_reduce"]
|
500 |
+
self.config.index_rate = values["index_rate"]
|
501 |
+
return True
|
502 |
+
|
503 |
+
def start_vc(self):
|
504 |
+
torch.cuda.empty_cache()
|
505 |
+
self.flag_vc = True
|
506 |
+
self.block_frame = int(self.config.block_time * self.config.samplerate)
|
507 |
+
self.crossfade_frame = int(self.config.crossfade_time * self.config.samplerate)
|
508 |
+
self.sola_search_frame = int(0.012 * self.config.samplerate)
|
509 |
+
self.delay_frame = int(0.01 * self.config.samplerate) # 往前预留0.02s
|
510 |
+
self.extra_frame = int(self.config.extra_time * self.config.samplerate)
|
511 |
+
self.rvc = None
|
512 |
+
self.rvc = RVC(
|
513 |
+
self.config.pitch,
|
514 |
+
self.config.hubert_path,
|
515 |
+
self.config.pth_path,
|
516 |
+
self.config.index_path,
|
517 |
+
self.config.npy_path,
|
518 |
+
self.config.index_rate,
|
519 |
+
)
|
520 |
+
self.input_wav: np.ndarray = np.zeros(
|
521 |
+
self.extra_frame
|
522 |
+
+ self.crossfade_frame
|
523 |
+
+ self.sola_search_frame
|
524 |
+
+ self.block_frame,
|
525 |
+
dtype="float32",
|
526 |
+
)
|
527 |
+
self.output_wav: torch.Tensor = torch.zeros(
|
528 |
+
self.block_frame, device=device, dtype=torch.float32
|
529 |
+
)
|
530 |
+
self.sola_buffer: torch.Tensor = torch.zeros(
|
531 |
+
self.crossfade_frame, device=device, dtype=torch.float32
|
532 |
+
)
|
533 |
+
self.fade_in_window: torch.Tensor = torch.linspace(
|
534 |
+
0.0, 1.0, steps=self.crossfade_frame, device=device, dtype=torch.float32
|
535 |
+
)
|
536 |
+
self.fade_out_window: torch.Tensor = 1 - self.fade_in_window
|
537 |
+
self.resampler1 = tat.Resample(
|
538 |
+
orig_freq=self.config.samplerate, new_freq=16000, dtype=torch.float32
|
539 |
+
)
|
540 |
+
self.resampler2 = tat.Resample(
|
541 |
+
orig_freq=self.rvc.tgt_sr,
|
542 |
+
new_freq=self.config.samplerate,
|
543 |
+
dtype=torch.float32,
|
544 |
+
)
|
545 |
+
thread_vc = threading.Thread(target=self.soundinput)
|
546 |
+
thread_vc.start()
|
547 |
+
|
548 |
+
def soundinput(self):
|
549 |
+
"""
|
550 |
+
接受音频输入
|
551 |
+
"""
|
552 |
+
with sd.Stream(
|
553 |
+
callback=self.audio_callback,
|
554 |
+
blocksize=self.block_frame,
|
555 |
+
samplerate=self.config.samplerate,
|
556 |
+
dtype="float32",
|
557 |
+
):
|
558 |
+
while self.flag_vc:
|
559 |
+
time.sleep(self.config.block_time)
|
560 |
+
print("Audio block passed.")
|
561 |
+
print("ENDing VC")
|
562 |
+
|
563 |
+
def audio_callback(
|
564 |
+
self, indata: np.ndarray, outdata: np.ndarray, frames, times, status
|
565 |
+
):
|
566 |
+
"""
|
567 |
+
音频处理
|
568 |
+
"""
|
569 |
+
start_time = time.perf_counter()
|
570 |
+
indata = librosa.to_mono(indata.T)
|
571 |
+
if self.config.I_noise_reduce:
|
572 |
+
indata[:] = nr.reduce_noise(y=indata, sr=self.config.samplerate)
|
573 |
+
|
574 |
+
"""noise gate"""
|
575 |
+
frame_length = 2048
|
576 |
+
hop_length = 1024
|
577 |
+
rms = librosa.feature.rms(
|
578 |
+
y=indata, frame_length=frame_length, hop_length=hop_length
|
579 |
+
)
|
580 |
+
db_threhold = librosa.amplitude_to_db(rms, ref=1.0)[0] < self.config.threhold
|
581 |
+
# print(rms.shape,db.shape,db)
|
582 |
+
for i in range(db_threhold.shape[0]):
|
583 |
+
if db_threhold[i]:
|
584 |
+
indata[i * hop_length : (i + 1) * hop_length] = 0
|
585 |
+
self.input_wav[:] = np.append(self.input_wav[self.block_frame :], indata)
|
586 |
+
|
587 |
+
# infer
|
588 |
+
print("input_wav:" + str(self.input_wav.shape))
|
589 |
+
# print('infered_wav:'+str(infer_wav.shape))
|
590 |
+
infer_wav: torch.Tensor = self.resampler2(
|
591 |
+
self.rvc.infer(self.resampler1(torch.from_numpy(self.input_wav)))
|
592 |
+
)[-self.crossfade_frame - self.sola_search_frame - self.block_frame :].to(
|
593 |
+
device
|
594 |
+
)
|
595 |
+
print("infer_wav:" + str(infer_wav.shape))
|
596 |
+
|
597 |
+
# SOLA algorithm from https://github.com/yxlllc/DDSP-SVC
|
598 |
+
cor_nom = F.conv1d(
|
599 |
+
infer_wav[None, None, : self.crossfade_frame + self.sola_search_frame],
|
600 |
+
self.sola_buffer[None, None, :],
|
601 |
+
)
|
602 |
+
cor_den = torch.sqrt(
|
603 |
+
F.conv1d(
|
604 |
+
infer_wav[None, None, : self.crossfade_frame + self.sola_search_frame]
|
605 |
+
** 2,
|
606 |
+
torch.ones(1, 1, self.crossfade_frame, device=device),
|
607 |
+
)
|
608 |
+
+ 1e-8
|
609 |
+
)
|
610 |
+
sola_offset = torch.argmax(cor_nom[0, 0] / cor_den[0, 0])
|
611 |
+
print("sola offset: " + str(int(sola_offset)))
|
612 |
+
|
613 |
+
# crossfade
|
614 |
+
self.output_wav[:] = infer_wav[sola_offset : sola_offset + self.block_frame]
|
615 |
+
self.output_wav[: self.crossfade_frame] *= self.fade_in_window
|
616 |
+
self.output_wav[: self.crossfade_frame] += self.sola_buffer[:]
|
617 |
+
if sola_offset < self.sola_search_frame:
|
618 |
+
self.sola_buffer[:] = (
|
619 |
+
infer_wav[
|
620 |
+
-self.sola_search_frame
|
621 |
+
- self.crossfade_frame
|
622 |
+
+ sola_offset : -self.sola_search_frame
|
623 |
+
+ sola_offset
|
624 |
+
]
|
625 |
+
* self.fade_out_window
|
626 |
+
)
|
627 |
+
else:
|
628 |
+
self.sola_buffer[:] = (
|
629 |
+
infer_wav[-self.crossfade_frame :] * self.fade_out_window
|
630 |
+
)
|
631 |
+
|
632 |
+
if self.config.O_noise_reduce:
|
633 |
+
outdata[:] = np.tile(
|
634 |
+
nr.reduce_noise(
|
635 |
+
y=self.output_wav[:].cpu().numpy(), sr=self.config.samplerate
|
636 |
+
),
|
637 |
+
(2, 1),
|
638 |
+
).T
|
639 |
+
else:
|
640 |
+
outdata[:] = self.output_wav[:].repeat(2, 1).t().cpu().numpy()
|
641 |
+
total_time = time.perf_counter() - start_time
|
642 |
+
self.window["infer_time"].update(int(total_time * 1000))
|
643 |
+
print("infer time:" + str(total_time))
|
644 |
+
|
645 |
+
def get_devices(self, update: bool = True):
|
646 |
+
"""获取设备列表"""
|
647 |
+
if update:
|
648 |
+
sd._terminate()
|
649 |
+
sd._initialize()
|
650 |
+
devices = sd.query_devices()
|
651 |
+
hostapis = sd.query_hostapis()
|
652 |
+
for hostapi in hostapis:
|
653 |
+
for device_idx in hostapi["devices"]:
|
654 |
+
devices[device_idx]["hostapi_name"] = hostapi["name"]
|
655 |
+
input_devices = [
|
656 |
+
f"{d['name']} ({d['hostapi_name']})"
|
657 |
+
for d in devices
|
658 |
+
if d["max_input_channels"] > 0
|
659 |
+
]
|
660 |
+
output_devices = [
|
661 |
+
f"{d['name']} ({d['hostapi_name']})"
|
662 |
+
for d in devices
|
663 |
+
if d["max_output_channels"] > 0
|
664 |
+
]
|
665 |
+
input_devices_indices = [
|
666 |
+
d["index"] if "index" in d else d["name"]
|
667 |
+
for d in devices
|
668 |
+
if d["max_input_channels"] > 0
|
669 |
+
]
|
670 |
+
output_devices_indices = [
|
671 |
+
d["index"] if "index" in d else d["name"]
|
672 |
+
for d in devices
|
673 |
+
if d["max_output_channels"] > 0
|
674 |
+
]
|
675 |
+
return (
|
676 |
+
input_devices,
|
677 |
+
output_devices,
|
678 |
+
input_devices_indices,
|
679 |
+
output_devices_indices,
|
680 |
+
)
|
681 |
+
|
682 |
+
def set_devices(self, input_device, output_device):
|
683 |
+
"""设置输出设备"""
|
684 |
+
(
|
685 |
+
input_devices,
|
686 |
+
output_devices,
|
687 |
+
input_device_indices,
|
688 |
+
output_device_indices,
|
689 |
+
) = self.get_devices()
|
690 |
+
sd.default.device[0] = input_device_indices[input_devices.index(input_device)]
|
691 |
+
sd.default.device[1] = output_device_indices[
|
692 |
+
output_devices.index(output_device)
|
693 |
+
]
|
694 |
+
print("input device:" + str(sd.default.device[0]) + ":" + str(input_device))
|
695 |
+
print("output device:" + str(sd.default.device[1]) + ":" + str(output_device))
|
696 |
+
|
697 |
+
|
698 |
+
gui = GUI()
|
host.json
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"version": "2.0",
|
3 |
+
"logging": {
|
4 |
+
"applicationInsights": {
|
5 |
+
"samplingSettings": {
|
6 |
+
"isEnabled": true,
|
7 |
+
"excludedTypes": "Request"
|
8 |
+
}
|
9 |
+
}
|
10 |
+
},
|
11 |
+
"extensionBundle": {
|
12 |
+
"id": "Microsoft.Azure.Functions.ExtensionBundle",
|
13 |
+
"version": "[4.*, 5.0.0)"
|
14 |
+
}
|
15 |
+
}
|
hubert_base.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f54b40fd2802423a5643779c4861af1e9ee9c1564dc9d32f54f20b5ffba7db96
|
3 |
+
size 189507909
|
i18n.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import locale
|
2 |
+
import json
|
3 |
+
import os
|
4 |
+
|
5 |
+
|
6 |
+
def load_language_list(language):
|
7 |
+
with open(f"./i18n/{language}.json", "r", encoding="utf-8") as f:
|
8 |
+
language_list = json.load(f)
|
9 |
+
return language_list
|
10 |
+
|
11 |
+
|
12 |
+
class I18nAuto:
|
13 |
+
def __init__(self, language=None):
|
14 |
+
if language in ["Auto", None]:
|
15 |
+
language = locale.getdefaultlocale()[
|
16 |
+
0
|
17 |
+
] # getlocale can't identify the system's language ((None, None))
|
18 |
+
if not os.path.exists(f"./i18n/{language}.json"):
|
19 |
+
language = "en_US"
|
20 |
+
self.language = language
|
21 |
+
# print("Use Language:", language)
|
22 |
+
self.language_map = load_language_list(language)
|
23 |
+
|
24 |
+
def __call__(self, key):
|
25 |
+
return self.language_map.get(key, key)
|
26 |
+
|
27 |
+
def print(self):
|
28 |
+
print("Use Language:", self.language)
|
infer-web.py
ADDED
@@ -0,0 +1,1999 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import shutil
|
3 |
+
import sys
|
4 |
+
# os.environ["CUDA_VISIBLE_DEVICES"] = ""
|
5 |
+
now_dir = os.getcwd()
|
6 |
+
sys.path.append(now_dir)
|
7 |
+
import traceback, pdb
|
8 |
+
import warnings
|
9 |
+
|
10 |
+
import numpy as np
|
11 |
+
import torch
|
12 |
+
|
13 |
+
os.environ['OPENBLAS_NUM_THREADS'] = '1'
|
14 |
+
os.environ["no_proxy"] = "localhost, 127.0.0.1, ::1"
|
15 |
+
import logging
|
16 |
+
import threading
|
17 |
+
from random import shuffle
|
18 |
+
from subprocess import Popen
|
19 |
+
from time import sleep
|
20 |
+
|
21 |
+
import faiss
|
22 |
+
import ffmpeg
|
23 |
+
import gradio as gr
|
24 |
+
import soundfile as sf
|
25 |
+
from config import Config
|
26 |
+
from fairseq import checkpoint_utils
|
27 |
+
from i18n import I18nAuto
|
28 |
+
from infer_pack.models import (
|
29 |
+
SynthesizerTrnMs256NSFsid,
|
30 |
+
SynthesizerTrnMs256NSFsid_nono,
|
31 |
+
SynthesizerTrnMs768NSFsid,
|
32 |
+
SynthesizerTrnMs768NSFsid_nono,
|
33 |
+
)
|
34 |
+
from infer_pack.models_onnx import SynthesizerTrnMsNSFsidM
|
35 |
+
from infer_uvr5 import _audio_pre_, _audio_pre_new
|
36 |
+
from MDXNet import MDXNetDereverb
|
37 |
+
from my_utils import load_audio
|
38 |
+
from train.process_ckpt import change_info, extract_small_model, merge, show_info
|
39 |
+
from vc_infer_pipeline import VC
|
40 |
+
from sklearn.cluster import MiniBatchKMeans
|
41 |
+
|
42 |
+
logging.getLogger("numba").setLevel(logging.WARNING)
|
43 |
+
|
44 |
+
|
45 |
+
tmp = os.path.join(now_dir, "TEMP")
|
46 |
+
shutil.rmtree(tmp, ignore_errors=True)
|
47 |
+
shutil.rmtree("%s/runtime/Lib/site-packages/infer_pack" % (now_dir), ignore_errors=True)
|
48 |
+
shutil.rmtree("%s/runtime/Lib/site-packages/uvr5_pack" % (now_dir), ignore_errors=True)
|
49 |
+
os.makedirs(tmp, exist_ok=True)
|
50 |
+
os.makedirs(os.path.join(now_dir, "logs"), exist_ok=True)
|
51 |
+
os.makedirs(os.path.join(now_dir, "weights"), exist_ok=True)
|
52 |
+
os.environ["TEMP"] = tmp
|
53 |
+
warnings.filterwarnings("ignore")
|
54 |
+
torch.manual_seed(114514)
|
55 |
+
|
56 |
+
|
57 |
+
config = Config()
|
58 |
+
i18n = I18nAuto()
|
59 |
+
i18n.print()
|
60 |
+
# 判断是否有能用来训练和加速推理的N卡
|
61 |
+
ngpu = torch.cuda.device_count()
|
62 |
+
gpu_infos = []
|
63 |
+
mem = []
|
64 |
+
if_gpu_ok = False
|
65 |
+
|
66 |
+
if torch.cuda.is_available() or ngpu != 0:
|
67 |
+
for i in range(ngpu):
|
68 |
+
gpu_name = torch.cuda.get_device_name(i)
|
69 |
+
if any(
|
70 |
+
value in gpu_name.upper()
|
71 |
+
for value in [
|
72 |
+
"10",
|
73 |
+
"16",
|
74 |
+
"20",
|
75 |
+
"30",
|
76 |
+
"40",
|
77 |
+
"A2",
|
78 |
+
"A3",
|
79 |
+
"A4",
|
80 |
+
"P4",
|
81 |
+
"A50",
|
82 |
+
"500",
|
83 |
+
"A60",
|
84 |
+
"70",
|
85 |
+
"80",
|
86 |
+
"90",
|
87 |
+
"M4",
|
88 |
+
"T4",
|
89 |
+
"TITAN",
|
90 |
+
]
|
91 |
+
):
|
92 |
+
# A10#A100#V100#A40#P40#M40#K80#A4500
|
93 |
+
if_gpu_ok = True # 至少有一张能用的N卡
|
94 |
+
gpu_infos.append("%s\t%s" % (i, gpu_name))
|
95 |
+
mem.append(
|
96 |
+
int(
|
97 |
+
torch.cuda.get_device_properties(i).total_memory
|
98 |
+
/ 1024
|
99 |
+
/ 1024
|
100 |
+
/ 1024
|
101 |
+
+ 0.4
|
102 |
+
)
|
103 |
+
)
|
104 |
+
if if_gpu_ok and len(gpu_infos) > 0:
|
105 |
+
gpu_info = "\n".join(gpu_infos)
|
106 |
+
default_batch_size = 1
|
107 |
+
else:
|
108 |
+
gpu_info = i18n("很遗憾您这没有能用的显卡来支持您训练")
|
109 |
+
default_batch_size = 1
|
110 |
+
gpus = "-".join([i[0] for i in gpu_infos])
|
111 |
+
|
112 |
+
|
113 |
+
class ToolButton(gr.Button, gr.components.FormComponent):
|
114 |
+
"""Small button with single emoji as text, fits inside gradio forms"""
|
115 |
+
|
116 |
+
def __init__(self, **kwargs):
|
117 |
+
super().__init__(variant="tool", **kwargs)
|
118 |
+
|
119 |
+
def get_block_name(self):
|
120 |
+
return "button"
|
121 |
+
|
122 |
+
|
123 |
+
hubert_model = None
|
124 |
+
|
125 |
+
|
126 |
+
def load_hubert():
|
127 |
+
global hubert_model
|
128 |
+
models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
|
129 |
+
["hubert_base.pt"],
|
130 |
+
suffix="",
|
131 |
+
)
|
132 |
+
hubert_model = models[0]
|
133 |
+
hubert_model = hubert_model.to(config.device)
|
134 |
+
if config.is_half:
|
135 |
+
hubert_model = hubert_model.half()
|
136 |
+
else:
|
137 |
+
hubert_model = hubert_model.float()
|
138 |
+
hubert_model.eval()
|
139 |
+
|
140 |
+
|
141 |
+
weight_root = "weights"
|
142 |
+
weight_uvr5_root = "uvr5_weights"
|
143 |
+
index_root = "logs"
|
144 |
+
names = []
|
145 |
+
for name in os.listdir(weight_root):
|
146 |
+
if name.endswith(".pth"):
|
147 |
+
names.append(name)
|
148 |
+
index_paths = []
|
149 |
+
for root, dirs, files in os.walk(index_root, topdown=False):
|
150 |
+
for name in files:
|
151 |
+
if name.endswith(".index") and "trained" not in name:
|
152 |
+
index_paths.append("%s/%s" % (root, name))
|
153 |
+
uvr5_names = []
|
154 |
+
for name in os.listdir(weight_uvr5_root):
|
155 |
+
if name.endswith(".pth") or "onnx" in name:
|
156 |
+
uvr5_names.append(name.replace(".pth", ""))
|
157 |
+
|
158 |
+
|
159 |
+
def vc_single(
|
160 |
+
sid,
|
161 |
+
input_audio_path,
|
162 |
+
f0_up_key,
|
163 |
+
f0_file,
|
164 |
+
f0_method,
|
165 |
+
file_index,
|
166 |
+
file_index2,
|
167 |
+
# file_big_npy,
|
168 |
+
index_rate,
|
169 |
+
filter_radius,
|
170 |
+
resample_sr,
|
171 |
+
rms_mix_rate,
|
172 |
+
protect,
|
173 |
+
): # spk_item, input_audio0, vc_transform0,f0_file,f0method0
|
174 |
+
global tgt_sr, net_g, vc, hubert_model, version
|
175 |
+
if input_audio_path is None:
|
176 |
+
return "You need to upload an audio", None
|
177 |
+
f0_up_key = int(f0_up_key)
|
178 |
+
try:
|
179 |
+
audio = load_audio(input_audio_path, 16000)
|
180 |
+
audio_max = np.abs(audio).max() / 0.95
|
181 |
+
if audio_max > 1:
|
182 |
+
audio /= audio_max
|
183 |
+
times = [0, 0, 0]
|
184 |
+
if not hubert_model:
|
185 |
+
load_hubert()
|
186 |
+
if_f0 = cpt.get("f0", 1)
|
187 |
+
file_index = (
|
188 |
+
(
|
189 |
+
file_index.strip(" ")
|
190 |
+
.strip('"')
|
191 |
+
.strip("\n")
|
192 |
+
.strip('"')
|
193 |
+
.strip(" ")
|
194 |
+
.replace("trained", "added")
|
195 |
+
)
|
196 |
+
if file_index != ""
|
197 |
+
else file_index2
|
198 |
+
) # 防止小白写错,自动帮他替换掉
|
199 |
+
# file_big_npy = (
|
200 |
+
# file_big_npy.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
201 |
+
# )
|
202 |
+
audio_opt = vc.pipeline(
|
203 |
+
hubert_model,
|
204 |
+
net_g,
|
205 |
+
sid,
|
206 |
+
audio,
|
207 |
+
input_audio_path,
|
208 |
+
times,
|
209 |
+
f0_up_key,
|
210 |
+
f0_method,
|
211 |
+
file_index,
|
212 |
+
# file_big_npy,
|
213 |
+
index_rate,
|
214 |
+
if_f0,
|
215 |
+
filter_radius,
|
216 |
+
tgt_sr,
|
217 |
+
resample_sr,
|
218 |
+
rms_mix_rate,
|
219 |
+
version,
|
220 |
+
protect,
|
221 |
+
f0_file=f0_file,
|
222 |
+
)
|
223 |
+
if tgt_sr != resample_sr >= 16000:
|
224 |
+
tgt_sr = resample_sr
|
225 |
+
index_info = (
|
226 |
+
"Using index:%s." % file_index
|
227 |
+
if os.path.exists(file_index)
|
228 |
+
else "Index not used."
|
229 |
+
)
|
230 |
+
return "Success.\n %s\nTime:\n npy:%ss, f0:%ss, infer:%ss" % (
|
231 |
+
index_info,
|
232 |
+
times[0],
|
233 |
+
times[1],
|
234 |
+
times[2],
|
235 |
+
), (tgt_sr, audio_opt)
|
236 |
+
except:
|
237 |
+
info = traceback.format_exc()
|
238 |
+
print(info)
|
239 |
+
return info, (None, None)
|
240 |
+
|
241 |
+
|
242 |
+
def vc_multi(
|
243 |
+
sid,
|
244 |
+
dir_path,
|
245 |
+
opt_root,
|
246 |
+
paths,
|
247 |
+
f0_up_key,
|
248 |
+
f0_method,
|
249 |
+
file_index,
|
250 |
+
file_index2,
|
251 |
+
# file_big_npy,
|
252 |
+
index_rate,
|
253 |
+
filter_radius,
|
254 |
+
resample_sr,
|
255 |
+
rms_mix_rate,
|
256 |
+
protect,
|
257 |
+
format1,
|
258 |
+
):
|
259 |
+
try:
|
260 |
+
dir_path = (
|
261 |
+
dir_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
262 |
+
) # 防止小白拷路径头尾带了空格和"和回车
|
263 |
+
opt_root = opt_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
264 |
+
os.makedirs(opt_root, exist_ok=True)
|
265 |
+
try:
|
266 |
+
if dir_path != "":
|
267 |
+
paths = [os.path.join(dir_path, name) for name in os.listdir(dir_path)]
|
268 |
+
else:
|
269 |
+
paths = [path.name for path in paths]
|
270 |
+
except:
|
271 |
+
traceback.print_exc()
|
272 |
+
paths = [path.name for path in paths]
|
273 |
+
infos = []
|
274 |
+
for path in paths:
|
275 |
+
info, opt = vc_single(
|
276 |
+
sid,
|
277 |
+
path,
|
278 |
+
f0_up_key,
|
279 |
+
None,
|
280 |
+
f0_method,
|
281 |
+
file_index,
|
282 |
+
file_index2,
|
283 |
+
# file_big_npy,
|
284 |
+
index_rate,
|
285 |
+
filter_radius,
|
286 |
+
resample_sr,
|
287 |
+
rms_mix_rate,
|
288 |
+
protect,
|
289 |
+
)
|
290 |
+
if "Success" in info:
|
291 |
+
try:
|
292 |
+
tgt_sr, audio_opt = opt
|
293 |
+
if format1 in ["wav", "flac"]:
|
294 |
+
sf.write(
|
295 |
+
"%s/%s.%s" % (opt_root, os.path.basename(path), format1),
|
296 |
+
audio_opt,
|
297 |
+
tgt_sr,
|
298 |
+
)
|
299 |
+
else:
|
300 |
+
path = "%s/%s.wav" % (opt_root, os.path.basename(path))
|
301 |
+
sf.write(
|
302 |
+
path,
|
303 |
+
audio_opt,
|
304 |
+
tgt_sr,
|
305 |
+
)
|
306 |
+
if os.path.exists(path):
|
307 |
+
os.system(
|
308 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
309 |
+
% (path, path[:-4] + ".%s" % format1)
|
310 |
+
)
|
311 |
+
except:
|
312 |
+
info += traceback.format_exc()
|
313 |
+
infos.append("%s->%s" % (os.path.basename(path), info))
|
314 |
+
yield "\n".join(infos)
|
315 |
+
yield "\n".join(infos)
|
316 |
+
except:
|
317 |
+
yield traceback.format_exc()
|
318 |
+
|
319 |
+
|
320 |
+
def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0):
|
321 |
+
infos = []
|
322 |
+
try:
|
323 |
+
inp_root = inp_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
324 |
+
save_root_vocal = (
|
325 |
+
save_root_vocal.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
326 |
+
)
|
327 |
+
save_root_ins = (
|
328 |
+
save_root_ins.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
329 |
+
)
|
330 |
+
if model_name == "onnx_dereverb_By_FoxJoy":
|
331 |
+
pre_fun = MDXNetDereverb(15)
|
332 |
+
else:
|
333 |
+
func = _audio_pre_ if "DeEcho" not in model_name else _audio_pre_new
|
334 |
+
pre_fun = func(
|
335 |
+
agg=int(agg),
|
336 |
+
model_path=os.path.join(weight_uvr5_root, model_name + ".pth"),
|
337 |
+
device=config.device,
|
338 |
+
is_half=config.is_half,
|
339 |
+
)
|
340 |
+
if inp_root != "":
|
341 |
+
paths = [os.path.join(inp_root, name) for name in os.listdir(inp_root)]
|
342 |
+
else:
|
343 |
+
paths = [path.name for path in paths]
|
344 |
+
for path in paths:
|
345 |
+
inp_path = os.path.join(inp_root, path)
|
346 |
+
need_reformat = 1
|
347 |
+
done = 0
|
348 |
+
try:
|
349 |
+
info = ffmpeg.probe(inp_path, cmd="ffprobe")
|
350 |
+
if (
|
351 |
+
info["streams"][0]["channels"] == 2
|
352 |
+
and info["streams"][0]["sample_rate"] == "44100"
|
353 |
+
):
|
354 |
+
need_reformat = 0
|
355 |
+
pre_fun._path_audio_(
|
356 |
+
inp_path, save_root_ins, save_root_vocal, format0
|
357 |
+
)
|
358 |
+
done = 1
|
359 |
+
except:
|
360 |
+
need_reformat = 1
|
361 |
+
traceback.print_exc()
|
362 |
+
if need_reformat == 1:
|
363 |
+
tmp_path = "%s/%s.reformatted.wav" % (tmp, os.path.basename(inp_path))
|
364 |
+
os.system(
|
365 |
+
"ffmpeg -i %s -vn -acodec pcm_s16le -ac 2 -ar 44100 %s -y"
|
366 |
+
% (inp_path, tmp_path)
|
367 |
+
)
|
368 |
+
inp_path = tmp_path
|
369 |
+
try:
|
370 |
+
if done == 0:
|
371 |
+
pre_fun._path_audio_(
|
372 |
+
inp_path, save_root_ins, save_root_vocal, format0
|
373 |
+
)
|
374 |
+
infos.append("%s->Success" % (os.path.basename(inp_path)))
|
375 |
+
yield "\n".join(infos)
|
376 |
+
except:
|
377 |
+
infos.append(
|
378 |
+
"%s->%s" % (os.path.basename(inp_path), traceback.format_exc())
|
379 |
+
)
|
380 |
+
yield "\n".join(infos)
|
381 |
+
except:
|
382 |
+
infos.append(traceback.format_exc())
|
383 |
+
yield "\n".join(infos)
|
384 |
+
finally:
|
385 |
+
try:
|
386 |
+
if model_name == "onnx_dereverb_By_FoxJoy":
|
387 |
+
del pre_fun.pred.model
|
388 |
+
del pre_fun.pred.model_
|
389 |
+
else:
|
390 |
+
del pre_fun.model
|
391 |
+
del pre_fun
|
392 |
+
except:
|
393 |
+
traceback.print_exc()
|
394 |
+
print("clean_empty_cache")
|
395 |
+
if torch.cuda.is_available():
|
396 |
+
torch.cuda.empty_cache()
|
397 |
+
yield "\n".join(infos)
|
398 |
+
|
399 |
+
|
400 |
+
# 一个选项卡全局只能有一个音色
|
401 |
+
def get_vc(sid, to_return_protect0, to_return_protect1):
|
402 |
+
global n_spk, tgt_sr, net_g, vc, cpt, version
|
403 |
+
if sid == "" or sid == []:
|
404 |
+
global hubert_model
|
405 |
+
if hubert_model is not None: # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
|
406 |
+
print("clean_empty_cache")
|
407 |
+
del net_g, n_spk, vc, hubert_model, tgt_sr # ,cpt
|
408 |
+
hubert_model = net_g = n_spk = vc = hubert_model = tgt_sr = None
|
409 |
+
if torch.cuda.is_available():
|
410 |
+
torch.cuda.empty_cache()
|
411 |
+
###楼下不这么折腾清理不干净
|
412 |
+
if_f0 = cpt.get("f0", 1)
|
413 |
+
version = cpt.get("version", "v1")
|
414 |
+
if version == "v1":
|
415 |
+
if if_f0 == 1:
|
416 |
+
net_g = SynthesizerTrnMs256NSFsid(
|
417 |
+
*cpt["config"], is_half=config.is_half
|
418 |
+
)
|
419 |
+
else:
|
420 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
421 |
+
elif version == "v2":
|
422 |
+
if if_f0 == 1:
|
423 |
+
net_g = SynthesizerTrnMs768NSFsid(
|
424 |
+
*cpt["config"], is_half=config.is_half
|
425 |
+
)
|
426 |
+
else:
|
427 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
428 |
+
del net_g, cpt
|
429 |
+
if torch.cuda.is_available():
|
430 |
+
torch.cuda.empty_cache()
|
431 |
+
cpt = None
|
432 |
+
return {"visible": False, "__type__": "update"}
|
433 |
+
person = "%s/%s" % (weight_root, sid)
|
434 |
+
print("loading %s" % person)
|
435 |
+
cpt = torch.load(person, map_location="cpu")
|
436 |
+
tgt_sr = cpt["config"][-1]
|
437 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
|
438 |
+
if_f0 = cpt.get("f0", 1)
|
439 |
+
if if_f0 == 0:
|
440 |
+
to_return_protect0 = to_return_protect1 = {
|
441 |
+
"visible": False,
|
442 |
+
"value": 0.5,
|
443 |
+
"__type__": "update",
|
444 |
+
}
|
445 |
+
else:
|
446 |
+
to_return_protect0 = {
|
447 |
+
"visible": True,
|
448 |
+
"value": to_return_protect0,
|
449 |
+
"__type__": "update",
|
450 |
+
}
|
451 |
+
to_return_protect1 = {
|
452 |
+
"visible": True,
|
453 |
+
"value": to_return_protect1,
|
454 |
+
"__type__": "update",
|
455 |
+
}
|
456 |
+
version = cpt.get("version", "v1")
|
457 |
+
if version == "v1":
|
458 |
+
if if_f0 == 1:
|
459 |
+
net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
|
460 |
+
else:
|
461 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
462 |
+
elif version == "v2":
|
463 |
+
if if_f0 == 1:
|
464 |
+
net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
|
465 |
+
else:
|
466 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
467 |
+
del net_g.enc_q
|
468 |
+
print(net_g.load_state_dict(cpt["weight"], strict=False))
|
469 |
+
net_g.eval().to(config.device)
|
470 |
+
if config.is_half:
|
471 |
+
net_g = net_g.half()
|
472 |
+
else:
|
473 |
+
net_g = net_g.float()
|
474 |
+
vc = VC(tgt_sr, config)
|
475 |
+
n_spk = cpt["config"][-3]
|
476 |
+
return (
|
477 |
+
{"visible": True, "maximum": n_spk, "__type__": "update"},
|
478 |
+
to_return_protect0,
|
479 |
+
to_return_protect1,
|
480 |
+
)
|
481 |
+
|
482 |
+
|
483 |
+
def change_choices():
|
484 |
+
names = []
|
485 |
+
for name in os.listdir(weight_root):
|
486 |
+
if name.endswith(".pth"):
|
487 |
+
names.append(name)
|
488 |
+
index_paths = []
|
489 |
+
for root, dirs, files in os.walk(index_root, topdown=False):
|
490 |
+
for name in files:
|
491 |
+
if name.endswith(".index") and "trained" not in name:
|
492 |
+
index_paths.append("%s/%s" % (root, name))
|
493 |
+
return {"choices": sorted(names), "__type__": "update"}, {
|
494 |
+
"choices": sorted(index_paths),
|
495 |
+
"__type__": "update",
|
496 |
+
}
|
497 |
+
|
498 |
+
|
499 |
+
def clean():
|
500 |
+
return {"value": "", "__type__": "update"}
|
501 |
+
|
502 |
+
|
503 |
+
sr_dict = {
|
504 |
+
"32k": 32000,
|
505 |
+
"40k": 40000,
|
506 |
+
"48k": 48000,
|
507 |
+
}
|
508 |
+
|
509 |
+
|
510 |
+
def if_done(done, p):
|
511 |
+
while 1:
|
512 |
+
if p.poll() is None:
|
513 |
+
sleep(0.5)
|
514 |
+
else:
|
515 |
+
break
|
516 |
+
done[0] = True
|
517 |
+
|
518 |
+
|
519 |
+
def if_done_multi(done, ps):
|
520 |
+
while 1:
|
521 |
+
# poll==None代表进程未结束
|
522 |
+
# 只要有一个进程未结束都不停
|
523 |
+
flag = 1
|
524 |
+
for p in ps:
|
525 |
+
if p.poll() is None:
|
526 |
+
flag = 0
|
527 |
+
sleep(0.5)
|
528 |
+
break
|
529 |
+
if flag == 1:
|
530 |
+
break
|
531 |
+
done[0] = True
|
532 |
+
|
533 |
+
|
534 |
+
def preprocess_dataset(trainset_dir, exp_dir, sr, n_p):
|
535 |
+
sr = sr_dict[sr]
|
536 |
+
os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
|
537 |
+
f = open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "w")
|
538 |
+
f.close()
|
539 |
+
cmd = (
|
540 |
+
config.python_cmd
|
541 |
+
+ " trainset_preprocess_pipeline_print.py %s %s %s %s/logs/%s "
|
542 |
+
% (trainset_dir, sr, n_p, now_dir, exp_dir)
|
543 |
+
+ str(config.noparallel)
|
544 |
+
)
|
545 |
+
print(cmd)
|
546 |
+
p = Popen(cmd, shell=True) # , stdin=PIPE, stdout=PIPE,stderr=PIPE,cwd=now_dir
|
547 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
548 |
+
done = [False]
|
549 |
+
threading.Thread(
|
550 |
+
target=if_done,
|
551 |
+
args=(
|
552 |
+
done,
|
553 |
+
p,
|
554 |
+
),
|
555 |
+
).start()
|
556 |
+
while 1:
|
557 |
+
with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
|
558 |
+
yield (f.read())
|
559 |
+
sleep(1)
|
560 |
+
if done[0]:
|
561 |
+
break
|
562 |
+
with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
|
563 |
+
log = f.read()
|
564 |
+
print(log)
|
565 |
+
yield log
|
566 |
+
|
567 |
+
|
568 |
+
# but2.click(extract_f0,[gpus6,np7,f0method8,if_f0_3,trainset_dir4],[info2])
|
569 |
+
def extract_f0_feature(gpus, n_p, f0method, if_f0, exp_dir, version19):
|
570 |
+
gpus = gpus.split("-")
|
571 |
+
os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
|
572 |
+
f = open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "w")
|
573 |
+
f.close()
|
574 |
+
if if_f0:
|
575 |
+
cmd = config.python_cmd + " extract_f0_print.py %s/logs/%s %s %s" % (
|
576 |
+
now_dir,
|
577 |
+
exp_dir,
|
578 |
+
n_p,
|
579 |
+
f0method,
|
580 |
+
)
|
581 |
+
print(cmd)
|
582 |
+
p = Popen(cmd, shell=True, cwd=now_dir) # , stdin=PIPE, stdout=PIPE,stderr=PIPE
|
583 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
584 |
+
done = [False]
|
585 |
+
threading.Thread(
|
586 |
+
target=if_done,
|
587 |
+
args=(
|
588 |
+
done,
|
589 |
+
p,
|
590 |
+
),
|
591 |
+
).start()
|
592 |
+
while 1:
|
593 |
+
with open(
|
594 |
+
"%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r"
|
595 |
+
) as f:
|
596 |
+
yield (f.read())
|
597 |
+
sleep(1)
|
598 |
+
if done[0]:
|
599 |
+
break
|
600 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
601 |
+
log = f.read()
|
602 |
+
print(log)
|
603 |
+
yield log
|
604 |
+
####对不同part分别开多进程
|
605 |
+
"""
|
606 |
+
n_part=int(sys.argv[1])
|
607 |
+
i_part=int(sys.argv[2])
|
608 |
+
i_gpu=sys.argv[3]
|
609 |
+
exp_dir=sys.argv[4]
|
610 |
+
os.environ["CUDA_VISIBLE_DEVICES"]=str(i_gpu)
|
611 |
+
"""
|
612 |
+
leng = len(gpus)
|
613 |
+
ps = []
|
614 |
+
for idx, n_g in enumerate(gpus):
|
615 |
+
cmd = (
|
616 |
+
config.python_cmd
|
617 |
+
+ " extract_feature_print.py %s %s %s %s %s/logs/%s %s"
|
618 |
+
% (
|
619 |
+
config.device,
|
620 |
+
leng,
|
621 |
+
idx,
|
622 |
+
n_g,
|
623 |
+
now_dir,
|
624 |
+
exp_dir,
|
625 |
+
version19,
|
626 |
+
)
|
627 |
+
)
|
628 |
+
print(cmd)
|
629 |
+
p = Popen(
|
630 |
+
cmd, shell=True, cwd=now_dir
|
631 |
+
) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
|
632 |
+
ps.append(p)
|
633 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
634 |
+
done = [False]
|
635 |
+
threading.Thread(
|
636 |
+
target=if_done_multi,
|
637 |
+
args=(
|
638 |
+
done,
|
639 |
+
ps,
|
640 |
+
),
|
641 |
+
).start()
|
642 |
+
while 1:
|
643 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
644 |
+
yield (f.read())
|
645 |
+
sleep(1)
|
646 |
+
if done[0]:
|
647 |
+
break
|
648 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
649 |
+
log = f.read()
|
650 |
+
print(log)
|
651 |
+
yield log
|
652 |
+
|
653 |
+
|
654 |
+
def change_sr2(sr2, if_f0_3, version19):
|
655 |
+
path_str = "" if version19 == "v1" else "_v2"
|
656 |
+
f0_str = "f0" if if_f0_3 else ""
|
657 |
+
if_pretrained_generator_exist = os.access(
|
658 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
659 |
+
)
|
660 |
+
if_pretrained_discriminator_exist = os.access(
|
661 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
662 |
+
)
|
663 |
+
if not if_pretrained_generator_exist:
|
664 |
+
print(
|
665 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2),
|
666 |
+
"not exist, will not use pretrained model",
|
667 |
+
)
|
668 |
+
if not if_pretrained_discriminator_exist:
|
669 |
+
print(
|
670 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2),
|
671 |
+
"not exist, will not use pretrained model",
|
672 |
+
)
|
673 |
+
return (
|
674 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2)
|
675 |
+
if if_pretrained_generator_exist
|
676 |
+
else "",
|
677 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2)
|
678 |
+
if if_pretrained_discriminator_exist
|
679 |
+
else "",
|
680 |
+
)
|
681 |
+
|
682 |
+
|
683 |
+
def change_version19(sr2, if_f0_3, version19):
|
684 |
+
path_str = "" if version19 == "v1" else "_v2"
|
685 |
+
if sr2 == "32k" and version19 == "v1":
|
686 |
+
sr2 = "40k"
|
687 |
+
to_return_sr2 = (
|
688 |
+
{"choices": ["40k", "48k"], "__type__": "update", "value": sr2}
|
689 |
+
if version19 == "v1"
|
690 |
+
else {"choices": ["40k", "48k", "32k"], "__type__": "update", "value": sr2}
|
691 |
+
)
|
692 |
+
f0_str = "f0" if if_f0_3 else ""
|
693 |
+
if_pretrained_generator_exist = os.access(
|
694 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
695 |
+
)
|
696 |
+
if_pretrained_discriminator_exist = os.access(
|
697 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
698 |
+
)
|
699 |
+
if not if_pretrained_generator_exist:
|
700 |
+
print(
|
701 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2),
|
702 |
+
"not exist, will not use pretrained model",
|
703 |
+
)
|
704 |
+
if not if_pretrained_discriminator_exist:
|
705 |
+
print(
|
706 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2),
|
707 |
+
"not exist, will not use pretrained model",
|
708 |
+
)
|
709 |
+
return (
|
710 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2)
|
711 |
+
if if_pretrained_generator_exist
|
712 |
+
else "",
|
713 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2)
|
714 |
+
if if_pretrained_discriminator_exist
|
715 |
+
else "",
|
716 |
+
to_return_sr2,
|
717 |
+
)
|
718 |
+
|
719 |
+
|
720 |
+
def change_f0(if_f0_3, sr2, version19): # f0method8,pretrained_G14,pretrained_D15
|
721 |
+
path_str = "" if version19 == "v1" else "_v2"
|
722 |
+
if_pretrained_generator_exist = os.access(
|
723 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2), os.F_OK
|
724 |
+
)
|
725 |
+
if_pretrained_discriminator_exist = os.access(
|
726 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2), os.F_OK
|
727 |
+
)
|
728 |
+
if not if_pretrained_generator_exist:
|
729 |
+
print(
|
730 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2),
|
731 |
+
"not exist, will not use pretrained model",
|
732 |
+
)
|
733 |
+
if not if_pretrained_discriminator_exist:
|
734 |
+
print(
|
735 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2),
|
736 |
+
"not exist, will not use pretrained model",
|
737 |
+
)
|
738 |
+
if if_f0_3:
|
739 |
+
return (
|
740 |
+
{"visible": True, "__type__": "update"},
|
741 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2)
|
742 |
+
if if_pretrained_generator_exist
|
743 |
+
else "",
|
744 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2)
|
745 |
+
if if_pretrained_discriminator_exist
|
746 |
+
else "",
|
747 |
+
)
|
748 |
+
return (
|
749 |
+
{"visible": False, "__type__": "update"},
|
750 |
+
("pretrained%s/G%s.pth" % (path_str, sr2))
|
751 |
+
if if_pretrained_generator_exist
|
752 |
+
else "",
|
753 |
+
("pretrained%s/D%s.pth" % (path_str, sr2))
|
754 |
+
if if_pretrained_discriminator_exist
|
755 |
+
else "",
|
756 |
+
)
|
757 |
+
|
758 |
+
|
759 |
+
# but3.click(click_train,[exp_dir1,sr2,if_f0_3,save_epoch10,total_epoch11,batch_size12,if_save_latest13,pretrained_G14,pretrained_D15,gpus16])
|
760 |
+
def click_train(
|
761 |
+
exp_dir1,
|
762 |
+
sr2,
|
763 |
+
if_f0_3,
|
764 |
+
spk_id5,
|
765 |
+
save_epoch10,
|
766 |
+
total_epoch11,
|
767 |
+
batch_size12,
|
768 |
+
if_save_latest13,
|
769 |
+
pretrained_G14,
|
770 |
+
pretrained_D15,
|
771 |
+
gpus16,
|
772 |
+
if_cache_gpu17,
|
773 |
+
if_save_every_weights18,
|
774 |
+
version19,
|
775 |
+
):
|
776 |
+
# 生成filelist
|
777 |
+
exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
778 |
+
os.makedirs(exp_dir, exist_ok=True)
|
779 |
+
gt_wavs_dir = "%s/0_gt_wavs" % (exp_dir)
|
780 |
+
feature_dir = (
|
781 |
+
"%s/3_feature256" % (exp_dir)
|
782 |
+
if version19 == "v1"
|
783 |
+
else "%s/3_feature768" % (exp_dir)
|
784 |
+
)
|
785 |
+
if if_f0_3:
|
786 |
+
f0_dir = "%s/2a_f0" % (exp_dir)
|
787 |
+
f0nsf_dir = "%s/2b-f0nsf" % (exp_dir)
|
788 |
+
names = (
|
789 |
+
set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)])
|
790 |
+
& set([name.split(".")[0] for name in os.listdir(feature_dir)])
|
791 |
+
& set([name.split(".")[0] for name in os.listdir(f0_dir)])
|
792 |
+
& set([name.split(".")[0] for name in os.listdir(f0nsf_dir)])
|
793 |
+
)
|
794 |
+
else:
|
795 |
+
names = set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)]) & set(
|
796 |
+
[name.split(".")[0] for name in os.listdir(feature_dir)]
|
797 |
+
)
|
798 |
+
opt = []
|
799 |
+
for name in names:
|
800 |
+
if if_f0_3:
|
801 |
+
opt.append(
|
802 |
+
"%s/%s.wav|%s/%s.npy|%s/%s.wav.npy|%s/%s.wav.npy|%s"
|
803 |
+
% (
|
804 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
805 |
+
name,
|
806 |
+
feature_dir.replace("\\", "\\\\"),
|
807 |
+
name,
|
808 |
+
f0_dir.replace("\\", "\\\\"),
|
809 |
+
name,
|
810 |
+
f0nsf_dir.replace("\\", "\\\\"),
|
811 |
+
name,
|
812 |
+
spk_id5,
|
813 |
+
)
|
814 |
+
)
|
815 |
+
else:
|
816 |
+
opt.append(
|
817 |
+
"%s/%s.wav|%s/%s.npy|%s"
|
818 |
+
% (
|
819 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
820 |
+
name,
|
821 |
+
feature_dir.replace("\\", "\\\\"),
|
822 |
+
name,
|
823 |
+
spk_id5,
|
824 |
+
)
|
825 |
+
)
|
826 |
+
fea_dim = 256 if version19 == "v1" else 768
|
827 |
+
if if_f0_3:
|
828 |
+
for _ in range(2):
|
829 |
+
opt.append(
|
830 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s/logs/mute/2a_f0/mute.wav.npy|%s/logs/mute/2b-f0nsf/mute.wav.npy|%s"
|
831 |
+
% (now_dir, sr2, now_dir, fea_dim, now_dir, now_dir, spk_id5)
|
832 |
+
)
|
833 |
+
else:
|
834 |
+
for _ in range(2):
|
835 |
+
opt.append(
|
836 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s"
|
837 |
+
% (now_dir, sr2, now_dir, fea_dim, spk_id5)
|
838 |
+
)
|
839 |
+
shuffle(opt)
|
840 |
+
with open("%s/filelist.txt" % exp_dir, "w") as f:
|
841 |
+
f.write("\n".join(opt))
|
842 |
+
print("write filelist done")
|
843 |
+
# 生成config#无需生成config
|
844 |
+
# cmd = python_cmd + " train_nsf_sim_cache_sid_load_pretrain.py -e mi-test -sr 40k -f0 1 -bs 4 -g 0 -te 10 -se 5 -pg pretrained/f0G40k.pth -pd pretrained/f0D40k.pth -l 1 -c 0"
|
845 |
+
print("use gpus:", gpus16)
|
846 |
+
if pretrained_G14 == "":
|
847 |
+
print("no pretrained Generator")
|
848 |
+
if pretrained_D15 == "":
|
849 |
+
print("no pretrained Discriminator")
|
850 |
+
if gpus16:
|
851 |
+
cmd = (
|
852 |
+
config.python_cmd
|
853 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -g %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
854 |
+
% (
|
855 |
+
exp_dir1,
|
856 |
+
sr2,
|
857 |
+
1 if if_f0_3 else 0,
|
858 |
+
batch_size12,
|
859 |
+
gpus16,
|
860 |
+
total_epoch11,
|
861 |
+
save_epoch10,
|
862 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
863 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
864 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
865 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
866 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
867 |
+
version19,
|
868 |
+
)
|
869 |
+
)
|
870 |
+
else:
|
871 |
+
cmd = (
|
872 |
+
config.python_cmd
|
873 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
874 |
+
% (
|
875 |
+
exp_dir1,
|
876 |
+
sr2,
|
877 |
+
1 if if_f0_3 else 0,
|
878 |
+
batch_size12,
|
879 |
+
total_epoch11,
|
880 |
+
save_epoch10,
|
881 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "\b",
|
882 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "\b",
|
883 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
884 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
885 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
886 |
+
version19,
|
887 |
+
)
|
888 |
+
)
|
889 |
+
print(cmd)
|
890 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
891 |
+
p.wait()
|
892 |
+
return "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log"
|
893 |
+
|
894 |
+
|
895 |
+
# but4.click(train_index, [exp_dir1], info3)
|
896 |
+
def train_index(exp_dir1, version19):
|
897 |
+
exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
898 |
+
os.makedirs(exp_dir, exist_ok=True)
|
899 |
+
feature_dir = (
|
900 |
+
"%s/3_feature256" % (exp_dir)
|
901 |
+
if version19 == "v1"
|
902 |
+
else "%s/3_feature768" % (exp_dir)
|
903 |
+
)
|
904 |
+
if not os.path.exists(feature_dir):
|
905 |
+
return "请先进行特征提取!"
|
906 |
+
listdir_res = list(os.listdir(feature_dir))
|
907 |
+
if len(listdir_res) == 0:
|
908 |
+
return "请先进行特征提取!"
|
909 |
+
infos = []
|
910 |
+
npys = []
|
911 |
+
for name in sorted(listdir_res):
|
912 |
+
phone = np.load("%s/%s" % (feature_dir, name))
|
913 |
+
npys.append(phone)
|
914 |
+
big_npy = np.concatenate(npys, 0)
|
915 |
+
big_npy_idx = np.arange(big_npy.shape[0])
|
916 |
+
np.random.shuffle(big_npy_idx)
|
917 |
+
big_npy = big_npy[big_npy_idx]
|
918 |
+
if big_npy.shape[0] > 2e5:
|
919 |
+
# if(1):
|
920 |
+
infos.append("Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0])
|
921 |
+
yield "\n".join(infos)
|
922 |
+
try:
|
923 |
+
big_npy = (
|
924 |
+
MiniBatchKMeans(
|
925 |
+
n_clusters=10000,
|
926 |
+
verbose=True,
|
927 |
+
batch_size=256 * config.n_cpu,
|
928 |
+
compute_labels=False,
|
929 |
+
init="random",
|
930 |
+
)
|
931 |
+
.fit(big_npy)
|
932 |
+
.cluster_centers_
|
933 |
+
)
|
934 |
+
except:
|
935 |
+
info = traceback.format_exc()
|
936 |
+
print(info)
|
937 |
+
infos.append(info)
|
938 |
+
yield "\n".join(infos)
|
939 |
+
|
940 |
+
np.save("%s/total_fea.npy" % exp_dir, big_npy)
|
941 |
+
n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
|
942 |
+
infos.append("%s,%s" % (big_npy.shape, n_ivf))
|
943 |
+
yield "\n".join(infos)
|
944 |
+
index = faiss.index_factory(256 if version19 == "v1" else 768, "IVF%s,Flat" % n_ivf)
|
945 |
+
# index = faiss.index_factory(256if version19=="v1"else 768, "IVF%s,PQ128x4fs,RFlat"%n_ivf)
|
946 |
+
infos.append("training")
|
947 |
+
yield "\n".join(infos)
|
948 |
+
index_ivf = faiss.extract_index_ivf(index) #
|
949 |
+
index_ivf.nprobe = 1
|
950 |
+
index.train(big_npy)
|
951 |
+
faiss.write_index(
|
952 |
+
index,
|
953 |
+
"%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
954 |
+
% (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
955 |
+
)
|
956 |
+
# faiss.write_index(index, '%s/trained_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19))
|
957 |
+
infos.append("adding")
|
958 |
+
yield "\n".join(infos)
|
959 |
+
batch_size_add = 8192
|
960 |
+
for i in range(0, big_npy.shape[0], batch_size_add):
|
961 |
+
index.add(big_npy[i : i + batch_size_add])
|
962 |
+
faiss.write_index(
|
963 |
+
index,
|
964 |
+
"%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
965 |
+
% (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
966 |
+
)
|
967 |
+
infos.append(
|
968 |
+
"成功构建索引,added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
969 |
+
% (n_ivf, index_ivf.nprobe, exp_dir1, version19)
|
970 |
+
)
|
971 |
+
# faiss.write_index(index, '%s/added_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19))
|
972 |
+
# infos.append("成功构建索引,added_IVF%s_Flat_FastScan_%s.index"%(n_ivf,version19))
|
973 |
+
yield "\n".join(infos)
|
974 |
+
|
975 |
+
|
976 |
+
# but5.click(train1key, [exp_dir1, sr2, if_f0_3, trainset_dir4, spk_id5, gpus6, np7, f0method8, save_epoch10, total_epoch11, batch_size12, if_save_latest13, pretrained_G14, pretrained_D15, gpus16, if_cache_gpu17], info3)
|
977 |
+
def train1key(
|
978 |
+
exp_dir1,
|
979 |
+
sr2,
|
980 |
+
if_f0_3,
|
981 |
+
trainset_dir4,
|
982 |
+
spk_id5,
|
983 |
+
np7,
|
984 |
+
f0method8,
|
985 |
+
save_epoch10,
|
986 |
+
total_epoch11,
|
987 |
+
batch_size12,
|
988 |
+
if_save_latest13,
|
989 |
+
pretrained_G14,
|
990 |
+
pretrained_D15,
|
991 |
+
gpus16,
|
992 |
+
if_cache_gpu17,
|
993 |
+
if_save_every_weights18,
|
994 |
+
version19,
|
995 |
+
):
|
996 |
+
infos = []
|
997 |
+
|
998 |
+
def get_info_str(strr):
|
999 |
+
infos.append(strr)
|
1000 |
+
return "\n".join(infos)
|
1001 |
+
|
1002 |
+
model_log_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
1003 |
+
preprocess_log_path = "%s/preprocess.log" % model_log_dir
|
1004 |
+
extract_f0_feature_log_path = "%s/extract_f0_feature.log" % model_log_dir
|
1005 |
+
gt_wavs_dir = "%s/0_gt_wavs" % model_log_dir
|
1006 |
+
feature_dir = (
|
1007 |
+
"%s/3_feature256" % model_log_dir
|
1008 |
+
if version19 == "v1"
|
1009 |
+
else "%s/3_feature768" % model_log_dir
|
1010 |
+
)
|
1011 |
+
|
1012 |
+
os.makedirs(model_log_dir, exist_ok=True)
|
1013 |
+
#########step1:处理数据
|
1014 |
+
open(preprocess_log_path, "w").close()
|
1015 |
+
cmd = (
|
1016 |
+
config.python_cmd
|
1017 |
+
+ " trainset_preprocess_pipeline_print.py %s %s %s %s "
|
1018 |
+
% (trainset_dir4, sr_dict[sr2], np7, model_log_dir)
|
1019 |
+
+ str(config.noparallel)
|
1020 |
+
)
|
1021 |
+
yield get_info_str(i18n("step1:正在处理数据"))
|
1022 |
+
yield get_info_str(cmd)
|
1023 |
+
p = Popen(cmd, shell=True)
|
1024 |
+
p.wait()
|
1025 |
+
with open(preprocess_log_path, "r") as f:
|
1026 |
+
print(f.read())
|
1027 |
+
#########step2a:提取音高
|
1028 |
+
open(extract_f0_feature_log_path, "w")
|
1029 |
+
if if_f0_3:
|
1030 |
+
yield get_info_str("step2a:正在提取音高")
|
1031 |
+
cmd = config.python_cmd + " extract_f0_print.py %s %s %s" % (
|
1032 |
+
model_log_dir,
|
1033 |
+
np7,
|
1034 |
+
f0method8,
|
1035 |
+
)
|
1036 |
+
yield get_info_str(cmd)
|
1037 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
1038 |
+
p.wait()
|
1039 |
+
with open(extract_f0_feature_log_path, "r") as f:
|
1040 |
+
print(f.read())
|
1041 |
+
else:
|
1042 |
+
yield get_info_str(i18n("step2a:无需提取音高"))
|
1043 |
+
#######step2b:提取特征
|
1044 |
+
yield get_info_str(i18n("step2b:正在提取特征"))
|
1045 |
+
gpus = gpus16.split("-")
|
1046 |
+
leng = len(gpus)
|
1047 |
+
ps = []
|
1048 |
+
for idx, n_g in enumerate(gpus):
|
1049 |
+
cmd = config.python_cmd + " extract_feature_print.py %s %s %s %s %s %s" % (
|
1050 |
+
config.device,
|
1051 |
+
leng,
|
1052 |
+
idx,
|
1053 |
+
n_g,
|
1054 |
+
model_log_dir,
|
1055 |
+
version19,
|
1056 |
+
)
|
1057 |
+
yield get_info_str(cmd)
|
1058 |
+
p = Popen(
|
1059 |
+
cmd, shell=True, cwd=now_dir
|
1060 |
+
) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
|
1061 |
+
ps.append(p)
|
1062 |
+
for p in ps:
|
1063 |
+
p.wait()
|
1064 |
+
with open(extract_f0_feature_log_path, "r") as f:
|
1065 |
+
print(f.read())
|
1066 |
+
#######step3a:训练模型
|
1067 |
+
yield get_info_str(i18n("step3a:正在训练模型"))
|
1068 |
+
# 生成filelist
|
1069 |
+
if if_f0_3:
|
1070 |
+
f0_dir = "%s/2a_f0" % model_log_dir
|
1071 |
+
f0nsf_dir = "%s/2b-f0nsf" % model_log_dir
|
1072 |
+
names = (
|
1073 |
+
set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)])
|
1074 |
+
& set([name.split(".")[0] for name in os.listdir(feature_dir)])
|
1075 |
+
& set([name.split(".")[0] for name in os.listdir(f0_dir)])
|
1076 |
+
& set([name.split(".")[0] for name in os.listdir(f0nsf_dir)])
|
1077 |
+
)
|
1078 |
+
else:
|
1079 |
+
names = set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)]) & set(
|
1080 |
+
[name.split(".")[0] for name in os.listdir(feature_dir)]
|
1081 |
+
)
|
1082 |
+
opt = []
|
1083 |
+
for name in names:
|
1084 |
+
if if_f0_3:
|
1085 |
+
opt.append(
|
1086 |
+
"%s/%s.wav|%s/%s.npy|%s/%s.wav.npy|%s/%s.wav.npy|%s"
|
1087 |
+
% (
|
1088 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
1089 |
+
name,
|
1090 |
+
feature_dir.replace("\\", "\\\\"),
|
1091 |
+
name,
|
1092 |
+
f0_dir.replace("\\", "\\\\"),
|
1093 |
+
name,
|
1094 |
+
f0nsf_dir.replace("\\", "\\\\"),
|
1095 |
+
name,
|
1096 |
+
spk_id5,
|
1097 |
+
)
|
1098 |
+
)
|
1099 |
+
else:
|
1100 |
+
opt.append(
|
1101 |
+
"%s/%s.wav|%s/%s.npy|%s"
|
1102 |
+
% (
|
1103 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
1104 |
+
name,
|
1105 |
+
feature_dir.replace("\\", "\\\\"),
|
1106 |
+
name,
|
1107 |
+
spk_id5,
|
1108 |
+
)
|
1109 |
+
)
|
1110 |
+
fea_dim = 256 if version19 == "v1" else 768
|
1111 |
+
if if_f0_3:
|
1112 |
+
for _ in range(2):
|
1113 |
+
opt.append(
|
1114 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s/logs/mute/2a_f0/mute.wav.npy|%s/logs/mute/2b-f0nsf/mute.wav.npy|%s"
|
1115 |
+
% (now_dir, sr2, now_dir, fea_dim, now_dir, now_dir, spk_id5)
|
1116 |
+
)
|
1117 |
+
else:
|
1118 |
+
for _ in range(2):
|
1119 |
+
opt.append(
|
1120 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s"
|
1121 |
+
% (now_dir, sr2, now_dir, fea_dim, spk_id5)
|
1122 |
+
)
|
1123 |
+
shuffle(opt)
|
1124 |
+
with open("%s/filelist.txt" % model_log_dir, "w") as f:
|
1125 |
+
f.write("\n".join(opt))
|
1126 |
+
yield get_info_str("write filelist done")
|
1127 |
+
if gpus16:
|
1128 |
+
cmd = (
|
1129 |
+
config.python_cmd
|
1130 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -g %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
1131 |
+
% (
|
1132 |
+
exp_dir1,
|
1133 |
+
sr2,
|
1134 |
+
1 if if_f0_3 else 0,
|
1135 |
+
batch_size12,
|
1136 |
+
gpus16,
|
1137 |
+
total_epoch11,
|
1138 |
+
save_epoch10,
|
1139 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
1140 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
1141 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
1142 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
1143 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
1144 |
+
version19,
|
1145 |
+
)
|
1146 |
+
)
|
1147 |
+
else:
|
1148 |
+
cmd = (
|
1149 |
+
config.python_cmd
|
1150 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
1151 |
+
% (
|
1152 |
+
exp_dir1,
|
1153 |
+
sr2,
|
1154 |
+
1 if if_f0_3 else 0,
|
1155 |
+
batch_size12,
|
1156 |
+
total_epoch11,
|
1157 |
+
save_epoch10,
|
1158 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
1159 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
1160 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
1161 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
1162 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
1163 |
+
version19,
|
1164 |
+
)
|
1165 |
+
)
|
1166 |
+
yield get_info_str(cmd)
|
1167 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
1168 |
+
p.wait()
|
1169 |
+
yield get_info_str(i18n("训练结束, 您可查看控制台训练日志或实验文件夹下的train.log"))
|
1170 |
+
#######step3b:训练索引
|
1171 |
+
npys = []
|
1172 |
+
listdir_res = list(os.listdir(feature_dir))
|
1173 |
+
for name in sorted(listdir_res):
|
1174 |
+
phone = np.load("%s/%s" % (feature_dir, name))
|
1175 |
+
npys.append(phone)
|
1176 |
+
big_npy = np.concatenate(npys, 0)
|
1177 |
+
|
1178 |
+
big_npy_idx = np.arange(big_npy.shape[0])
|
1179 |
+
np.random.shuffle(big_npy_idx)
|
1180 |
+
big_npy = big_npy[big_npy_idx]
|
1181 |
+
|
1182 |
+
if big_npy.shape[0] > 2e5:
|
1183 |
+
# if(1):
|
1184 |
+
info = "Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0]
|
1185 |
+
print(info)
|
1186 |
+
yield get_info_str(info)
|
1187 |
+
try:
|
1188 |
+
big_npy = (
|
1189 |
+
MiniBatchKMeans(
|
1190 |
+
n_clusters=10000,
|
1191 |
+
verbose=True,
|
1192 |
+
batch_size=256 * config.n_cpu,
|
1193 |
+
compute_labels=False,
|
1194 |
+
init="random",
|
1195 |
+
)
|
1196 |
+
.fit(big_npy)
|
1197 |
+
.cluster_centers_
|
1198 |
+
)
|
1199 |
+
except:
|
1200 |
+
info = traceback.format_exc()
|
1201 |
+
print(info)
|
1202 |
+
yield get_info_str(info)
|
1203 |
+
|
1204 |
+
np.save("%s/total_fea.npy" % model_log_dir, big_npy)
|
1205 |
+
|
1206 |
+
# n_ivf = big_npy.shape[0] // 39
|
1207 |
+
n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
|
1208 |
+
yield get_info_str("%s,%s" % (big_npy.shape, n_ivf))
|
1209 |
+
index = faiss.index_factory(256 if version19 == "v1" else 768, "IVF%s,Flat" % n_ivf)
|
1210 |
+
yield get_info_str("training index")
|
1211 |
+
index_ivf = faiss.extract_index_ivf(index) #
|
1212 |
+
index_ivf.nprobe = 1
|
1213 |
+
index.train(big_npy)
|
1214 |
+
faiss.write_index(
|
1215 |
+
index,
|
1216 |
+
"%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1217 |
+
% (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
1218 |
+
)
|
1219 |
+
yield get_info_str("adding index")
|
1220 |
+
batch_size_add = 8192
|
1221 |
+
for i in range(0, big_npy.shape[0], batch_size_add):
|
1222 |
+
index.add(big_npy[i : i + batch_size_add])
|
1223 |
+
faiss.write_index(
|
1224 |
+
index,
|
1225 |
+
"%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1226 |
+
% (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
1227 |
+
)
|
1228 |
+
yield get_info_str(
|
1229 |
+
"成功构建索引, added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1230 |
+
% (n_ivf, index_ivf.nprobe, exp_dir1, version19)
|
1231 |
+
)
|
1232 |
+
yield get_info_str(i18n("全流程结束!"))
|
1233 |
+
|
1234 |
+
|
1235 |
+
# ckpt_path2.change(change_info_,[ckpt_path2],[sr__,if_f0__])
|
1236 |
+
def change_info_(ckpt_path):
|
1237 |
+
if not os.path.exists(ckpt_path.replace(os.path.basename(ckpt_path), "train.log")):
|
1238 |
+
return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
|
1239 |
+
try:
|
1240 |
+
with open(
|
1241 |
+
ckpt_path.replace(os.path.basename(ckpt_path), "train.log"), "r"
|
1242 |
+
) as f:
|
1243 |
+
info = eval(f.read().strip("\n").split("\n")[0].split("\t")[-1])
|
1244 |
+
sr, f0 = info["sample_rate"], info["if_f0"]
|
1245 |
+
version = "v2" if ("version" in info and info["version"] == "v2") else "v1"
|
1246 |
+
return sr, str(f0), version
|
1247 |
+
except:
|
1248 |
+
traceback.print_exc()
|
1249 |
+
return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
|
1250 |
+
|
1251 |
+
|
1252 |
+
def export_onnx(ModelPath, ExportedPath):
|
1253 |
+
cpt = torch.load(ModelPath, map_location="cpu")
|
1254 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
|
1255 |
+
vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768
|
1256 |
+
|
1257 |
+
test_phone = torch.rand(1, 200, vec_channels) # hidden unit
|
1258 |
+
test_phone_lengths = torch.tensor([200]).long() # hidden unit 长度(貌似没啥用)
|
1259 |
+
test_pitch = torch.randint(size=(1, 200), low=5, high=255) # 基频(单位赫兹)
|
1260 |
+
test_pitchf = torch.rand(1, 200) # nsf基频
|
1261 |
+
test_ds = torch.LongTensor([0]) # 说话人ID
|
1262 |
+
test_rnd = torch.rand(1, 192, 200) # 噪声(加入随机因子)
|
1263 |
+
|
1264 |
+
device = "cpu" # 导出时设备(不影响使用模型)
|
1265 |
+
|
1266 |
+
net_g = SynthesizerTrnMsNSFsidM(
|
1267 |
+
*cpt["config"], is_half=False, version=cpt.get("version", "v1")
|
1268 |
+
) # fp32导出(C++要支持fp16必须手动将内存重新排列所以暂时不用fp16)
|
1269 |
+
net_g.load_state_dict(cpt["weight"], strict=False)
|
1270 |
+
input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
|
1271 |
+
output_names = [
|
1272 |
+
"audio",
|
1273 |
+
]
|
1274 |
+
# net_g.construct_spkmixmap(n_speaker) 多角色混合轨道导出
|
1275 |
+
torch.onnx.export(
|
1276 |
+
net_g,
|
1277 |
+
(
|
1278 |
+
test_phone.to(device),
|
1279 |
+
test_phone_lengths.to(device),
|
1280 |
+
test_pitch.to(device),
|
1281 |
+
test_pitchf.to(device),
|
1282 |
+
test_ds.to(device),
|
1283 |
+
test_rnd.to(device),
|
1284 |
+
),
|
1285 |
+
ExportedPath,
|
1286 |
+
dynamic_axes={
|
1287 |
+
"phone": [1],
|
1288 |
+
"pitch": [1],
|
1289 |
+
"pitchf": [1],
|
1290 |
+
"rnd": [2],
|
1291 |
+
},
|
1292 |
+
do_constant_folding=False,
|
1293 |
+
opset_version=13,
|
1294 |
+
verbose=False,
|
1295 |
+
input_names=input_names,
|
1296 |
+
output_names=output_names,
|
1297 |
+
)
|
1298 |
+
return "Finished"
|
1299 |
+
|
1300 |
+
|
1301 |
+
with gr.Blocks() as app:
|
1302 |
+
gr.Markdown(
|
1303 |
+
value=i18n(
|
1304 |
+
"本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责. <br>如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录<b>使用需遵守的协议-LICENSE.txt</b>."
|
1305 |
+
)
|
1306 |
+
)
|
1307 |
+
with gr.Tabs():
|
1308 |
+
with gr.TabItem(i18n("模型推理")):
|
1309 |
+
with gr.Row():
|
1310 |
+
sid0 = gr.Dropdown(label=i18n("推理音色"), choices=sorted(names))
|
1311 |
+
refresh_button = gr.Button(i18n("刷新音色列表和索引路径"), variant="primary")
|
1312 |
+
clean_button = gr.Button(i18n("卸载音色省显存"), variant="primary")
|
1313 |
+
spk_item = gr.Slider(
|
1314 |
+
minimum=0,
|
1315 |
+
maximum=2333,
|
1316 |
+
step=1,
|
1317 |
+
label=i18n("请选择说话人id"),
|
1318 |
+
value=0,
|
1319 |
+
visible=False,
|
1320 |
+
interactive=True,
|
1321 |
+
)
|
1322 |
+
clean_button.click(fn=clean, inputs=[], outputs=[sid0])
|
1323 |
+
with gr.Group():
|
1324 |
+
gr.Markdown(
|
1325 |
+
value=i18n("男转女推荐+12key, 女转男推荐-12key, 如果音域爆炸导致音色失真也可以自己调整到合适音域. ")
|
1326 |
+
)
|
1327 |
+
with gr.Row():
|
1328 |
+
with gr.Column():
|
1329 |
+
vc_transform0 = gr.Number(
|
1330 |
+
label=i18n("变调(整数, 半音数量, 升八度12降八度-12)"), value=0
|
1331 |
+
)
|
1332 |
+
input_audio0 = gr.Textbox(
|
1333 |
+
label=i18n("输入待处理音频文件路径(默认是正确格式示例)"),
|
1334 |
+
value="E:\\codes\\py39\\test-20230416b\\todo-songs\\冬之花clip1.wav",
|
1335 |
+
)
|
1336 |
+
f0method0 = gr.Radio(
|
1337 |
+
label=i18n(
|
1338 |
+
"选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比,crepe效果好但吃GPU"
|
1339 |
+
),
|
1340 |
+
choices=["pm", "harvest", "crepe"],
|
1341 |
+
value="pm",
|
1342 |
+
interactive=True,
|
1343 |
+
)
|
1344 |
+
filter_radius0 = gr.Slider(
|
1345 |
+
minimum=0,
|
1346 |
+
maximum=7,
|
1347 |
+
label=i18n(">=3则使用对harvest音高识别的结果使用中值滤波,数值为滤波半径,使用可以削弱哑音"),
|
1348 |
+
value=3,
|
1349 |
+
step=1,
|
1350 |
+
interactive=True,
|
1351 |
+
)
|
1352 |
+
with gr.Column():
|
1353 |
+
file_index1 = gr.Textbox(
|
1354 |
+
label=i18n("特征检索库文件路径,为空则使用下拉的选择结果"),
|
1355 |
+
value="",
|
1356 |
+
interactive=True,
|
1357 |
+
)
|
1358 |
+
file_index2 = gr.Dropdown(
|
1359 |
+
label=i18n("自动检测index路径,下拉式选择(dropdown)"),
|
1360 |
+
choices=sorted(index_paths),
|
1361 |
+
interactive=True,
|
1362 |
+
)
|
1363 |
+
refresh_button.click(
|
1364 |
+
fn=change_choices, inputs=[], outputs=[sid0, file_index2]
|
1365 |
+
)
|
1366 |
+
# file_big_npy1 = gr.Textbox(
|
1367 |
+
# label=i18n("特征文件路径"),
|
1368 |
+
# value="E:\\codes\py39\\vits_vc_gpu_train\\logs\\mi-test-1key\\total_fea.npy",
|
1369 |
+
# interactive=True,
|
1370 |
+
# )
|
1371 |
+
index_rate1 = gr.Slider(
|
1372 |
+
minimum=0,
|
1373 |
+
maximum=1,
|
1374 |
+
label=i18n("检索特征占比"),
|
1375 |
+
value=0.75,
|
1376 |
+
interactive=True,
|
1377 |
+
)
|
1378 |
+
with gr.Column():
|
1379 |
+
resample_sr0 = gr.Slider(
|
1380 |
+
minimum=0,
|
1381 |
+
maximum=48000,
|
1382 |
+
label=i18n("后处理重采样至最终采样率,0为不进行重采样"),
|
1383 |
+
value=0,
|
1384 |
+
step=1,
|
1385 |
+
interactive=True,
|
1386 |
+
)
|
1387 |
+
rms_mix_rate0 = gr.Slider(
|
1388 |
+
minimum=0,
|
1389 |
+
maximum=1,
|
1390 |
+
label=i18n("输入源音量包络替换输出音量包络融合比例,越靠近1越使用输出包络"),
|
1391 |
+
value=0.25,
|
1392 |
+
interactive=True,
|
1393 |
+
)
|
1394 |
+
protect0 = gr.Slider(
|
1395 |
+
minimum=0,
|
1396 |
+
maximum=0.5,
|
1397 |
+
label=i18n(
|
1398 |
+
"保护清辅音和呼吸声,防止电音撕裂等artifact,拉满0.5不开启,调低加大保护力度但可能降低索引效果"
|
1399 |
+
),
|
1400 |
+
value=0.33,
|
1401 |
+
step=0.01,
|
1402 |
+
interactive=True,
|
1403 |
+
)
|
1404 |
+
f0_file = gr.File(label=i18n("F0曲线文件, 可选, 一行一个音高, 代替默认F0及升降调"))
|
1405 |
+
but0 = gr.Button(i18n("转换"), variant="primary")
|
1406 |
+
with gr.Row():
|
1407 |
+
vc_output1 = gr.Textbox(label=i18n("输出信息"))
|
1408 |
+
vc_output2 = gr.Audio(label=i18n("输出音频(右下角三个点,点了可以下载)"))
|
1409 |
+
but0.click(
|
1410 |
+
vc_single,
|
1411 |
+
[
|
1412 |
+
spk_item,
|
1413 |
+
input_audio0,
|
1414 |
+
vc_transform0,
|
1415 |
+
f0_file,
|
1416 |
+
f0method0,
|
1417 |
+
file_index1,
|
1418 |
+
file_index2,
|
1419 |
+
# file_big_npy1,
|
1420 |
+
index_rate1,
|
1421 |
+
filter_radius0,
|
1422 |
+
resample_sr0,
|
1423 |
+
rms_mix_rate0,
|
1424 |
+
protect0,
|
1425 |
+
],
|
1426 |
+
[vc_output1, vc_output2],
|
1427 |
+
)
|
1428 |
+
with gr.Group():
|
1429 |
+
gr.Markdown(
|
1430 |
+
value=i18n("批量转换, 输入待转换音频文件夹, 或上传多个音频文件, 在指定文件夹(默认opt)下输出转换的音频. ")
|
1431 |
+
)
|
1432 |
+
with gr.Row():
|
1433 |
+
with gr.Column():
|
1434 |
+
vc_transform1 = gr.Number(
|
1435 |
+
label=i18n("变调(整数, 半音数量, 升八度12降八度-12)"), value=0
|
1436 |
+
)
|
1437 |
+
opt_input = gr.Textbox(label=i18n("指定输出文件夹"), value="opt")
|
1438 |
+
f0method1 = gr.Radio(
|
1439 |
+
label=i18n(
|
1440 |
+
"选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比,crepe效果好但吃GPU"
|
1441 |
+
),
|
1442 |
+
choices=["pm", "harvest", "crepe"],
|
1443 |
+
value="pm",
|
1444 |
+
interactive=True,
|
1445 |
+
)
|
1446 |
+
filter_radius1 = gr.Slider(
|
1447 |
+
minimum=0,
|
1448 |
+
maximum=7,
|
1449 |
+
label=i18n(">=3则使用对harvest音高识别的结果使用中值滤波,数值为滤波半径,使用可以削弱哑音"),
|
1450 |
+
value=3,
|
1451 |
+
step=1,
|
1452 |
+
interactive=True,
|
1453 |
+
)
|
1454 |
+
with gr.Column():
|
1455 |
+
file_index3 = gr.Textbox(
|
1456 |
+
label=i18n("特征检索库文件路径,为空则使用下拉的选择结果"),
|
1457 |
+
value="",
|
1458 |
+
interactive=True,
|
1459 |
+
)
|
1460 |
+
file_index4 = gr.Dropdown(
|
1461 |
+
label=i18n("自动检测index路径,下拉式选择(dropdown)"),
|
1462 |
+
choices=sorted(index_paths),
|
1463 |
+
interactive=True,
|
1464 |
+
)
|
1465 |
+
refresh_button.click(
|
1466 |
+
fn=lambda: change_choices()[1],
|
1467 |
+
inputs=[],
|
1468 |
+
outputs=file_index4,
|
1469 |
+
)
|
1470 |
+
# file_big_npy2 = gr.Textbox(
|
1471 |
+
# label=i18n("特征文件路径"),
|
1472 |
+
# value="E:\\codes\\py39\\vits_vc_gpu_train\\logs\\mi-test-1key\\total_fea.npy",
|
1473 |
+
# interactive=True,
|
1474 |
+
# )
|
1475 |
+
index_rate2 = gr.Slider(
|
1476 |
+
minimum=0,
|
1477 |
+
maximum=1,
|
1478 |
+
label=i18n("检索特征占比"),
|
1479 |
+
value=1,
|
1480 |
+
interactive=True,
|
1481 |
+
)
|
1482 |
+
with gr.Column():
|
1483 |
+
resample_sr1 = gr.Slider(
|
1484 |
+
minimum=0,
|
1485 |
+
maximum=48000,
|
1486 |
+
label=i18n("后处理重采样至最终采样率,0为不进行重采样"),
|
1487 |
+
value=0,
|
1488 |
+
step=1,
|
1489 |
+
interactive=True,
|
1490 |
+
)
|
1491 |
+
rms_mix_rate1 = gr.Slider(
|
1492 |
+
minimum=0,
|
1493 |
+
maximum=1,
|
1494 |
+
label=i18n("输入源音量包络替换输出音量包络融合比例,越靠近1越使用输出包络"),
|
1495 |
+
value=1,
|
1496 |
+
interactive=True,
|
1497 |
+
)
|
1498 |
+
protect1 = gr.Slider(
|
1499 |
+
minimum=0,
|
1500 |
+
maximum=0.5,
|
1501 |
+
label=i18n(
|
1502 |
+
"保护清辅音和呼吸声,防止电音撕裂等artifact,拉满0.5不开启,调低加大保护力度但可能降低索引效果"
|
1503 |
+
),
|
1504 |
+
value=0.33,
|
1505 |
+
step=0.01,
|
1506 |
+
interactive=True,
|
1507 |
+
)
|
1508 |
+
with gr.Column():
|
1509 |
+
dir_input = gr.Textbox(
|
1510 |
+
label=i18n("输入待处理音频文件夹路径(去文件管理器地址栏拷就行了)"),
|
1511 |
+
value="E:\codes\py39\\test-20230416b\\todo-songs",
|
1512 |
+
)
|
1513 |
+
inputs = gr.File(
|
1514 |
+
file_count="multiple", label=i18n("也可批量输入音频文件, 二选一, 优先读文件夹")
|
1515 |
+
)
|
1516 |
+
with gr.Row():
|
1517 |
+
format1 = gr.Radio(
|
1518 |
+
label=i18n("导出文件格式"),
|
1519 |
+
choices=["wav", "flac", "mp3", "m4a"],
|
1520 |
+
value="flac",
|
1521 |
+
interactive=True,
|
1522 |
+
)
|
1523 |
+
but1 = gr.Button(i18n("转换"), variant="primary")
|
1524 |
+
vc_output3 = gr.Textbox(label=i18n("输出信息"))
|
1525 |
+
but1.click(
|
1526 |
+
vc_multi,
|
1527 |
+
[
|
1528 |
+
spk_item,
|
1529 |
+
dir_input,
|
1530 |
+
opt_input,
|
1531 |
+
inputs,
|
1532 |
+
vc_transform1,
|
1533 |
+
f0method1,
|
1534 |
+
file_index3,
|
1535 |
+
file_index4,
|
1536 |
+
# file_big_npy2,
|
1537 |
+
index_rate2,
|
1538 |
+
filter_radius1,
|
1539 |
+
resample_sr1,
|
1540 |
+
rms_mix_rate1,
|
1541 |
+
protect1,
|
1542 |
+
format1,
|
1543 |
+
],
|
1544 |
+
[vc_output3],
|
1545 |
+
)
|
1546 |
+
sid0.change(
|
1547 |
+
fn=get_vc,
|
1548 |
+
inputs=[sid0, protect0, protect1],
|
1549 |
+
outputs=[spk_item, protect0, protect1],
|
1550 |
+
)
|
1551 |
+
with gr.TabItem(i18n("伴奏人声分离&去混响&去回声")):
|
1552 |
+
with gr.Group():
|
1553 |
+
gr.Markdown(
|
1554 |
+
value=i18n(
|
1555 |
+
"人声伴奏分离批量处理, 使用UVR5模型。 <br>"
|
1556 |
+
"合格的文件夹路径格式举例: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例(去文件管理器地址栏拷就行了)。 <br>"
|
1557 |
+
"模型分为三类: <br>"
|
1558 |
+
"1、保留人声:不带和声的音频选这个,对主人声保留比HP5更好。内置HP2和HP3两个模型,HP3可能轻微漏伴奏但对主人声保留比HP2稍微好一丁点; <br>"
|
1559 |
+
"2、仅保留主人声:带和声的音频选这个,对主人声可能有削弱。内置HP5一个模型; <br> "
|
1560 |
+
"3、去混响、去延迟模型(by FoxJoy):<br>"
|
1561 |
+
" (1)MDX-Net(onnx_dereverb):对于双通道混响是最好的选择,不能去除单通道混响;<br>"
|
1562 |
+
" (234)DeEcho:去除延迟效果。Aggressive比Normal去除得更彻底,DeReverb额外去除混响,可去除单声道混响,但是对高频重的板式混响去不干净。<br>"
|
1563 |
+
"去混响/去延迟,附:<br>"
|
1564 |
+
"1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍;<br>"
|
1565 |
+
"2、MDX-Net-Dereverb模型挺慢的;<br>"
|
1566 |
+
"3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。"
|
1567 |
+
)
|
1568 |
+
)
|
1569 |
+
with gr.Row():
|
1570 |
+
with gr.Column():
|
1571 |
+
dir_wav_input = gr.Textbox(
|
1572 |
+
label=i18n("输入待处理音频文件夹路径"),
|
1573 |
+
value="E:\\codes\\py39\\test-20230416b\\todo-songs\\todo-songs",
|
1574 |
+
)
|
1575 |
+
wav_inputs = gr.File(
|
1576 |
+
file_count="multiple", label=i18n("也可批量输入音频文件, 二选一, 优先读文件夹")
|
1577 |
+
)
|
1578 |
+
with gr.Column():
|
1579 |
+
model_choose = gr.Dropdown(label=i18n("模型"), choices=uvr5_names)
|
1580 |
+
agg = gr.Slider(
|
1581 |
+
minimum=0,
|
1582 |
+
maximum=20,
|
1583 |
+
step=1,
|
1584 |
+
label="人声提取激进程度",
|
1585 |
+
value=10,
|
1586 |
+
interactive=True,
|
1587 |
+
visible=False, # 先不开放调整
|
1588 |
+
)
|
1589 |
+
opt_vocal_root = gr.Textbox(
|
1590 |
+
label=i18n("指定输出主人声文件夹"), value="opt"
|
1591 |
+
)
|
1592 |
+
opt_ins_root = gr.Textbox(
|
1593 |
+
label=i18n("指定输出非主人声文件夹"), value="opt"
|
1594 |
+
)
|
1595 |
+
format0 = gr.Radio(
|
1596 |
+
label=i18n("导出文件格式"),
|
1597 |
+
choices=["wav", "flac", "mp3", "m4a"],
|
1598 |
+
value="flac",
|
1599 |
+
interactive=True,
|
1600 |
+
)
|
1601 |
+
but2 = gr.Button(i18n("转换"), variant="primary")
|
1602 |
+
vc_output4 = gr.Textbox(label=i18n("输出信息"))
|
1603 |
+
but2.click(
|
1604 |
+
uvr,
|
1605 |
+
[
|
1606 |
+
model_choose,
|
1607 |
+
dir_wav_input,
|
1608 |
+
opt_vocal_root,
|
1609 |
+
wav_inputs,
|
1610 |
+
opt_ins_root,
|
1611 |
+
agg,
|
1612 |
+
format0,
|
1613 |
+
],
|
1614 |
+
[vc_output4],
|
1615 |
+
)
|
1616 |
+
with gr.TabItem(i18n("训练")):
|
1617 |
+
gr.Markdown(
|
1618 |
+
value=i18n(
|
1619 |
+
"step1: 填写实验配置. 实验数据放在logs下, 每个实验一个文件夹, 需手工输入实验名路径, 内含实验配置, 日志, 训练得到的模型文件. "
|
1620 |
+
)
|
1621 |
+
)
|
1622 |
+
with gr.Row():
|
1623 |
+
exp_dir1 = gr.Textbox(label=i18n("输入实验名"), value="mi-test")
|
1624 |
+
sr2 = gr.Radio(
|
1625 |
+
label=i18n("目标采样率"),
|
1626 |
+
choices=["40k", "48k"],
|
1627 |
+
value="40k",
|
1628 |
+
interactive=True,
|
1629 |
+
)
|
1630 |
+
if_f0_3 = gr.Radio(
|
1631 |
+
label=i18n("模型是否带音高指导(唱歌一定要, 语音可以不要)"),
|
1632 |
+
choices=[True, False],
|
1633 |
+
value=True,
|
1634 |
+
interactive=True,
|
1635 |
+
)
|
1636 |
+
version19 = gr.Radio(
|
1637 |
+
label=i18n("版本"),
|
1638 |
+
choices=["v1", "v2"],
|
1639 |
+
value="v1",
|
1640 |
+
interactive=True,
|
1641 |
+
visible=True,
|
1642 |
+
)
|
1643 |
+
np7 = gr.Slider(
|
1644 |
+
minimum=0,
|
1645 |
+
maximum=config.n_cpu,
|
1646 |
+
step=1,
|
1647 |
+
label=i18n("提取音高和处理数据使用的CPU进程数"),
|
1648 |
+
value=int(np.ceil(config.n_cpu / 1.5)),
|
1649 |
+
interactive=True,
|
1650 |
+
)
|
1651 |
+
with gr.Group(): # 暂时单人的, 后面支持最多4人的#数据处理
|
1652 |
+
gr.Markdown(
|
1653 |
+
value=i18n(
|
1654 |
+
"step2a: 自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练. "
|
1655 |
+
)
|
1656 |
+
)
|
1657 |
+
with gr.Row():
|
1658 |
+
trainset_dir4 = gr.Textbox(
|
1659 |
+
label=i18n("输入训练文件夹路径"), value="E:\\语音音频+标注\\米津玄师\\src"
|
1660 |
+
)
|
1661 |
+
spk_id5 = gr.Slider(
|
1662 |
+
minimum=0,
|
1663 |
+
maximum=4,
|
1664 |
+
step=1,
|
1665 |
+
label=i18n("请指定说话人id"),
|
1666 |
+
value=0,
|
1667 |
+
interactive=True,
|
1668 |
+
)
|
1669 |
+
but1 = gr.Button(i18n("处理数据"), variant="primary")
|
1670 |
+
info1 = gr.Textbox(label=i18n("输出信息"), value="")
|
1671 |
+
but1.click(
|
1672 |
+
preprocess_dataset, [trainset_dir4, exp_dir1, sr2, np7], [info1]
|
1673 |
+
)
|
1674 |
+
with gr.Group():
|
1675 |
+
gr.Markdown(value=i18n("step2b: 使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号)"))
|
1676 |
+
with gr.Row():
|
1677 |
+
with gr.Column():
|
1678 |
+
gpus6 = gr.Textbox(
|
1679 |
+
label=i18n("以-分隔输入使用的卡号, 例如 0-1-2 使用卡0和卡1和卡2"),
|
1680 |
+
value=gpus,
|
1681 |
+
interactive=True,
|
1682 |
+
)
|
1683 |
+
gpu_info9 = gr.Textbox(label=i18n("显卡信息"), value=gpu_info)
|
1684 |
+
with gr.Column():
|
1685 |
+
f0method8 = gr.Radio(
|
1686 |
+
label=i18n(
|
1687 |
+
"选择音高提取算法:输入歌声可用pm提速,高质量语音但CPU差可用dio提速,harvest质量更好但慢"
|
1688 |
+
),
|
1689 |
+
choices=["pm", "harvest", "dio"],
|
1690 |
+
value="harvest",
|
1691 |
+
interactive=True,
|
1692 |
+
)
|
1693 |
+
but2 = gr.Button(i18n("特征提取"), variant="primary")
|
1694 |
+
info2 = gr.Textbox(label=i18n("输出信息"), value="", max_lines=8)
|
1695 |
+
but2.click(
|
1696 |
+
extract_f0_feature,
|
1697 |
+
[gpus6, np7, f0method8, if_f0_3, exp_dir1, version19],
|
1698 |
+
[info2],
|
1699 |
+
)
|
1700 |
+
with gr.Group():
|
1701 |
+
gr.Markdown(value=i18n("step3: 填写训练设置, 开始训练模型和索引"))
|
1702 |
+
with gr.Row():
|
1703 |
+
save_epoch10 = gr.Slider(
|
1704 |
+
minimum=0,
|
1705 |
+
maximum=50,
|
1706 |
+
step=1,
|
1707 |
+
label=i18n("保存频率save_every_epoch"),
|
1708 |
+
value=5,
|
1709 |
+
interactive=True,
|
1710 |
+
)
|
1711 |
+
total_epoch11 = gr.Slider(
|
1712 |
+
minimum=0,
|
1713 |
+
maximum=1000,
|
1714 |
+
step=1,
|
1715 |
+
label=i18n("总训练轮数total_epoch"),
|
1716 |
+
value=20,
|
1717 |
+
interactive=True,
|
1718 |
+
)
|
1719 |
+
batch_size12 = gr.Slider(
|
1720 |
+
minimum=1,
|
1721 |
+
maximum=40,
|
1722 |
+
step=1,
|
1723 |
+
label=i18n("每张显卡的batch_size"),
|
1724 |
+
value=default_batch_size,
|
1725 |
+
interactive=True,
|
1726 |
+
)
|
1727 |
+
if_save_latest13 = gr.Radio(
|
1728 |
+
label=i18n("是否仅保存最新的ckpt文件以节省硬盘空间"),
|
1729 |
+
choices=[i18n("是"), i18n("否")],
|
1730 |
+
value=i18n("否"),
|
1731 |
+
interactive=True,
|
1732 |
+
)
|
1733 |
+
if_cache_gpu17 = gr.Radio(
|
1734 |
+
label=i18n(
|
1735 |
+
"是否缓存所有训练集至显存. 10min以下小数据可缓存以加速训练, 大数据缓存会炸显存也加不了多少速"
|
1736 |
+
),
|
1737 |
+
choices=[i18n("是"), i18n("否")],
|
1738 |
+
value=i18n("否"),
|
1739 |
+
interactive=True,
|
1740 |
+
)
|
1741 |
+
if_save_every_weights18 = gr.Radio(
|
1742 |
+
label=i18n("是否在每次保存时间点将最终小模型保存至weights文件夹"),
|
1743 |
+
choices=[i18n("是"), i18n("否")],
|
1744 |
+
value=i18n("否"),
|
1745 |
+
interactive=True,
|
1746 |
+
)
|
1747 |
+
with gr.Row():
|
1748 |
+
pretrained_G14 = gr.Textbox(
|
1749 |
+
label=i18n("加载预训练底模G路径"),
|
1750 |
+
value="pretrained/f0G40k.pth",
|
1751 |
+
interactive=True,
|
1752 |
+
)
|
1753 |
+
pretrained_D15 = gr.Textbox(
|
1754 |
+
label=i18n("加载预训练底模D路径"),
|
1755 |
+
value="pretrained/f0D40k.pth",
|
1756 |
+
interactive=True,
|
1757 |
+
)
|
1758 |
+
sr2.change(
|
1759 |
+
change_sr2,
|
1760 |
+
[sr2, if_f0_3, version19],
|
1761 |
+
[pretrained_G14, pretrained_D15],
|
1762 |
+
)
|
1763 |
+
version19.change(
|
1764 |
+
change_version19,
|
1765 |
+
[sr2, if_f0_3, version19],
|
1766 |
+
[pretrained_G14, pretrained_D15, sr2],
|
1767 |
+
)
|
1768 |
+
if_f0_3.change(
|
1769 |
+
change_f0,
|
1770 |
+
[if_f0_3, sr2, version19],
|
1771 |
+
[f0method8, pretrained_G14, pretrained_D15],
|
1772 |
+
)
|
1773 |
+
gpus16 = gr.Textbox(
|
1774 |
+
label=i18n("以-分隔输入使用的卡号, 例如 0-1-2 使用卡0和卡1和卡2"),
|
1775 |
+
value=gpus,
|
1776 |
+
interactive=True,
|
1777 |
+
)
|
1778 |
+
but3 = gr.Button(i18n("训练模型"), variant="primary")
|
1779 |
+
but4 = gr.Button(i18n("训练特征索引"), variant="primary")
|
1780 |
+
but5 = gr.Button(i18n("一键训练"), variant="primary")
|
1781 |
+
info3 = gr.Textbox(label=i18n("输出信息"), value="", max_lines=10)
|
1782 |
+
but3.click(
|
1783 |
+
click_train,
|
1784 |
+
[
|
1785 |
+
exp_dir1,
|
1786 |
+
sr2,
|
1787 |
+
if_f0_3,
|
1788 |
+
spk_id5,
|
1789 |
+
save_epoch10,
|
1790 |
+
total_epoch11,
|
1791 |
+
batch_size12,
|
1792 |
+
if_save_latest13,
|
1793 |
+
pretrained_G14,
|
1794 |
+
pretrained_D15,
|
1795 |
+
gpus16,
|
1796 |
+
if_cache_gpu17,
|
1797 |
+
if_save_every_weights18,
|
1798 |
+
version19,
|
1799 |
+
],
|
1800 |
+
info3,
|
1801 |
+
)
|
1802 |
+
but4.click(train_index, [exp_dir1, version19], info3)
|
1803 |
+
but5.click(
|
1804 |
+
train1key,
|
1805 |
+
[
|
1806 |
+
exp_dir1,
|
1807 |
+
sr2,
|
1808 |
+
if_f0_3,
|
1809 |
+
trainset_dir4,
|
1810 |
+
spk_id5,
|
1811 |
+
np7,
|
1812 |
+
f0method8,
|
1813 |
+
save_epoch10,
|
1814 |
+
total_epoch11,
|
1815 |
+
batch_size12,
|
1816 |
+
if_save_latest13,
|
1817 |
+
pretrained_G14,
|
1818 |
+
pretrained_D15,
|
1819 |
+
gpus16,
|
1820 |
+
if_cache_gpu17,
|
1821 |
+
if_save_every_weights18,
|
1822 |
+
version19,
|
1823 |
+
],
|
1824 |
+
info3,
|
1825 |
+
)
|
1826 |
+
|
1827 |
+
with gr.TabItem(i18n("ckpt处理")):
|
1828 |
+
with gr.Group():
|
1829 |
+
gr.Markdown(value=i18n("模型融合, 可用于测试音色融合"))
|
1830 |
+
with gr.Row():
|
1831 |
+
ckpt_a = gr.Textbox(label=i18n("A模型路径"), value="", interactive=True)
|
1832 |
+
ckpt_b = gr.Textbox(label=i18n("B模型路径"), value="", interactive=True)
|
1833 |
+
alpha_a = gr.Slider(
|
1834 |
+
minimum=0,
|
1835 |
+
maximum=1,
|
1836 |
+
label=i18n("A模型权重"),
|
1837 |
+
value=0.5,
|
1838 |
+
interactive=True,
|
1839 |
+
)
|
1840 |
+
with gr.Row():
|
1841 |
+
sr_ = gr.Radio(
|
1842 |
+
label=i18n("目标采样率"),
|
1843 |
+
choices=["40k", "48k"],
|
1844 |
+
value="40k",
|
1845 |
+
interactive=True,
|
1846 |
+
)
|
1847 |
+
if_f0_ = gr.Radio(
|
1848 |
+
label=i18n("模型是否带音高指导"),
|
1849 |
+
choices=[i18n("是"), i18n("否")],
|
1850 |
+
value=i18n("是"),
|
1851 |
+
interactive=True,
|
1852 |
+
)
|
1853 |
+
info__ = gr.Textbox(
|
1854 |
+
label=i18n("要置入的模型信息"), value="", max_lines=8, interactive=True
|
1855 |
+
)
|
1856 |
+
name_to_save0 = gr.Textbox(
|
1857 |
+
label=i18n("保存的模型名不带后缀"),
|
1858 |
+
value="",
|
1859 |
+
max_lines=1,
|
1860 |
+
interactive=True,
|
1861 |
+
)
|
1862 |
+
version_2 = gr.Radio(
|
1863 |
+
label=i18n("模型版本型号"),
|
1864 |
+
choices=["v1", "v2"],
|
1865 |
+
value="v1",
|
1866 |
+
interactive=True,
|
1867 |
+
)
|
1868 |
+
with gr.Row():
|
1869 |
+
but6 = gr.Button(i18n("融合"), variant="primary")
|
1870 |
+
info4 = gr.Textbox(label=i18n("输出信息"), value="", max_lines=8)
|
1871 |
+
but6.click(
|
1872 |
+
merge,
|
1873 |
+
[
|
1874 |
+
ckpt_a,
|
1875 |
+
ckpt_b,
|
1876 |
+
alpha_a,
|
1877 |
+
sr_,
|
1878 |
+
if_f0_,
|
1879 |
+
info__,
|
1880 |
+
name_to_save0,
|
1881 |
+
version_2,
|
1882 |
+
],
|
1883 |
+
info4,
|
1884 |
+
) # def merge(path1,path2,alpha1,sr,f0,info):
|
1885 |
+
with gr.Group():
|
1886 |
+
gr.Markdown(value=i18n("修改模型信息(仅支持weights文件夹下提取的小模型文件)"))
|
1887 |
+
with gr.Row():
|
1888 |
+
ckpt_path0 = gr.Textbox(
|
1889 |
+
label=i18n("模型路径"), value="", interactive=True
|
1890 |
+
)
|
1891 |
+
info_ = gr.Textbox(
|
1892 |
+
label=i18n("要改的模型信息"), value="", max_lines=8, interactive=True
|
1893 |
+
)
|
1894 |
+
name_to_save1 = gr.Textbox(
|
1895 |
+
label=i18n("保存的文件名, 默认空为和源文件同名"),
|
1896 |
+
value="",
|
1897 |
+
max_lines=8,
|
1898 |
+
interactive=True,
|
1899 |
+
)
|
1900 |
+
with gr.Row():
|
1901 |
+
but7 = gr.Button(i18n("修改"), variant="primary")
|
1902 |
+
info5 = gr.Textbox(label=i18n("输出信息"), value="", max_lines=8)
|
1903 |
+
but7.click(change_info, [ckpt_path0, info_, name_to_save1], info5)
|
1904 |
+
with gr.Group():
|
1905 |
+
gr.Markdown(value=i18n("查看模型信息(仅支持weights文件夹下提取的小模型文件)"))
|
1906 |
+
with gr.Row():
|
1907 |
+
ckpt_path1 = gr.Textbox(
|
1908 |
+
label=i18n("模型路径"), value="", interactive=True
|
1909 |
+
)
|
1910 |
+
but8 = gr.Button(i18n("查看"), variant="primary")
|
1911 |
+
info6 = gr.Textbox(label=i18n("输出信息"), value="", max_lines=8)
|
1912 |
+
but8.click(show_info, [ckpt_path1], info6)
|
1913 |
+
with gr.Group():
|
1914 |
+
gr.Markdown(
|
1915 |
+
value=i18n(
|
1916 |
+
"模型提取(输入logs文件夹下大文件模型路径),适用于训一半不想训了模型没有自动提取保存小文件模型,或者想测试中间模型的情况"
|
1917 |
+
)
|
1918 |
+
)
|
1919 |
+
with gr.Row():
|
1920 |
+
ckpt_path2 = gr.Textbox(
|
1921 |
+
label=i18n("模型路径"),
|
1922 |
+
value="E:\\codes\\py39\\logs\\mi-test_f0_48k\\G_23333.pth",
|
1923 |
+
interactive=True,
|
1924 |
+
)
|
1925 |
+
save_name = gr.Textbox(
|
1926 |
+
label=i18n("保存名"), value="", interactive=True
|
1927 |
+
)
|
1928 |
+
sr__ = gr.Radio(
|
1929 |
+
label=i18n("目标采样率"),
|
1930 |
+
choices=["32k", "40k", "48k"],
|
1931 |
+
value="40k",
|
1932 |
+
interactive=True,
|
1933 |
+
)
|
1934 |
+
if_f0__ = gr.Radio(
|
1935 |
+
label=i18n("模型是否带音高指导,1是0否"),
|
1936 |
+
choices=["1", "0"],
|
1937 |
+
value="1",
|
1938 |
+
interactive=True,
|
1939 |
+
)
|
1940 |
+
version_1 = gr.Radio(
|
1941 |
+
label=i18n("模型版本型号"),
|
1942 |
+
choices=["v1", "v2"],
|
1943 |
+
value="v2",
|
1944 |
+
interactive=True,
|
1945 |
+
)
|
1946 |
+
info___ = gr.Textbox(
|
1947 |
+
label=i18n("要置入的模型信息"), value="", max_lines=8, interactive=True
|
1948 |
+
)
|
1949 |
+
but9 = gr.Button(i18n("提取"), variant="primary")
|
1950 |
+
info7 = gr.Textbox(label=i18n("输出信息"), value="", max_lines=8)
|
1951 |
+
ckpt_path2.change(
|
1952 |
+
change_info_, [ckpt_path2], [sr__, if_f0__, version_1]
|
1953 |
+
)
|
1954 |
+
but9.click(
|
1955 |
+
extract_small_model,
|
1956 |
+
[ckpt_path2, save_name, sr__, if_f0__, info___, version_1],
|
1957 |
+
info7,
|
1958 |
+
)
|
1959 |
+
|
1960 |
+
with gr.TabItem(i18n("Onnx导出")):
|
1961 |
+
with gr.Row():
|
1962 |
+
ckpt_dir = gr.Textbox(label=i18n("RVC模型路径"), value="", interactive=True)
|
1963 |
+
with gr.Row():
|
1964 |
+
onnx_dir = gr.Textbox(
|
1965 |
+
label=i18n("Onnx输出路径"), value="", interactive=True
|
1966 |
+
)
|
1967 |
+
with gr.Row():
|
1968 |
+
infoOnnx = gr.Label(label="info")
|
1969 |
+
with gr.Row():
|
1970 |
+
butOnnx = gr.Button(i18n("导出Onnx模型"), variant="primary")
|
1971 |
+
butOnnx.click(export_onnx, [ckpt_dir, onnx_dir], infoOnnx)
|
1972 |
+
|
1973 |
+
tab_faq = i18n("常见问题解答")
|
1974 |
+
with gr.TabItem(tab_faq):
|
1975 |
+
try:
|
1976 |
+
if tab_faq == "常见问题解答":
|
1977 |
+
with open("docs/faq.md", "r", encoding="utf8") as f:
|
1978 |
+
info = f.read()
|
1979 |
+
else:
|
1980 |
+
with open("docs/faq_en.md", "r", encoding="utf8") as f:
|
1981 |
+
info = f.read()
|
1982 |
+
gr.Markdown(value=info)
|
1983 |
+
except:
|
1984 |
+
gr.Markdown(traceback.format_exc())
|
1985 |
+
|
1986 |
+
# with gr.TabItem(i18n("招募音高曲线前端编辑器")):
|
1987 |
+
# gr.Markdown(value=i18n("加开发群联系我xxxxx"))
|
1988 |
+
# with gr.TabItem(i18n("点击查看交流、问题反馈群号")):
|
1989 |
+
# gr.Markdown(value=i18n("xxxxx"))
|
1990 |
+
|
1991 |
+
if config.iscolab:
|
1992 |
+
app.queue(concurrency_count=511, max_size=1022).launch(share=True)
|
1993 |
+
else:
|
1994 |
+
app.queue(concurrency_count=511, max_size=1022).launch(
|
1995 |
+
server_name="0.0.0.0",
|
1996 |
+
inbrowser=not config.noautoopen,
|
1997 |
+
server_port=config.listen_port,
|
1998 |
+
quiet=True,
|
1999 |
+
)
|
infer_batch_rvc.py
ADDED
@@ -0,0 +1,220 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
v1
|
3 |
+
runtime\python.exe myinfer-v2-0528.py 0 "E:\codes\py39\RVC-beta\todo-songs" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" harvest "E:\codes\py39\RVC-beta\output" "E:\codes\py39\test-20230416b\weights\mi-test.pth" 0.66 cuda:0 True 3 0 1 0.33
|
4 |
+
v2
|
5 |
+
runtime\python.exe myinfer-v2-0528.py 0 "E:\codes\py39\RVC-beta\todo-songs" "E:\codes\py39\test-20230416b\logs\mi-test-v2\aadded_IVF677_Flat_nprobe_1_v2.index" harvest "E:\codes\py39\RVC-beta\output_v2" "E:\codes\py39\test-20230416b\weights\mi-test-v2.pth" 0.66 cuda:0 True 3 0 1 0.33
|
6 |
+
"""
|
7 |
+
import os, sys, pdb, torch
|
8 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = ""
|
9 |
+
now_dir = os.getcwd()
|
10 |
+
sys.path.append(now_dir)
|
11 |
+
import argparse
|
12 |
+
import glob
|
13 |
+
import sys
|
14 |
+
import torch
|
15 |
+
import tqdm as tq
|
16 |
+
from multiprocessing import cpu_count
|
17 |
+
|
18 |
+
|
19 |
+
class Config:
|
20 |
+
def __init__(self, device, is_half):
|
21 |
+
self.device = device
|
22 |
+
self.is_half = is_half
|
23 |
+
self.n_cpu = 0
|
24 |
+
self.gpu_name = None
|
25 |
+
self.gpu_mem = None
|
26 |
+
self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
|
27 |
+
|
28 |
+
def device_config(self) -> tuple:
|
29 |
+
if torch.cuda.is_available():
|
30 |
+
i_device = int(self.device.split(":")[-1])
|
31 |
+
self.gpu_name = torch.cuda.get_device_name(i_device)
|
32 |
+
if (
|
33 |
+
("16" in self.gpu_name and "V100" not in self.gpu_name.upper())
|
34 |
+
or "P40" in self.gpu_name.upper()
|
35 |
+
or "1060" in self.gpu_name
|
36 |
+
or "1070" in self.gpu_name
|
37 |
+
or "1080" in self.gpu_name
|
38 |
+
):
|
39 |
+
print("16系/10系显卡和P40强制单精度")
|
40 |
+
self.is_half = False
|
41 |
+
for config_file in ["32k.json", "40k.json", "48k.json"]:
|
42 |
+
with open(f"configs/{config_file}", "r") as f:
|
43 |
+
strr = f.read().replace("true", "false")
|
44 |
+
with open(f"configs/{config_file}", "w") as f:
|
45 |
+
f.write(strr)
|
46 |
+
with open("trainset_preprocess_pipeline_print.py", "r") as f:
|
47 |
+
strr = f.read().replace("3.7", "3.0")
|
48 |
+
with open("trainset_preprocess_pipeline_print.py", "w") as f:
|
49 |
+
f.write(strr)
|
50 |
+
else:
|
51 |
+
self.gpu_name = None
|
52 |
+
self.gpu_mem = int(
|
53 |
+
torch.cuda.get_device_properties(i_device).total_memory
|
54 |
+
/ 1024
|
55 |
+
/ 1024
|
56 |
+
/ 1024
|
57 |
+
+ 0.4
|
58 |
+
)
|
59 |
+
if self.gpu_mem <= 4:
|
60 |
+
with open("trainset_preprocess_pipeline_print.py", "r") as f:
|
61 |
+
strr = f.read().replace("3.7", "3.0")
|
62 |
+
with open("trainset_preprocess_pipeline_print.py", "w") as f:
|
63 |
+
f.write(strr)
|
64 |
+
elif torch.backends.mps.is_available():
|
65 |
+
print("没有发现支持的N卡, 使用MPS进行推理")
|
66 |
+
self.device = "mps"
|
67 |
+
else:
|
68 |
+
print("没有发现支持的N卡, 使用CPU进行推理")
|
69 |
+
self.device = "cpu"
|
70 |
+
self.is_half = True
|
71 |
+
|
72 |
+
if self.n_cpu == 0:
|
73 |
+
self.n_cpu = cpu_count()
|
74 |
+
|
75 |
+
if self.is_half:
|
76 |
+
# 6G显存配置
|
77 |
+
x_pad = 3
|
78 |
+
x_query = 10
|
79 |
+
x_center = 60
|
80 |
+
x_max = 65
|
81 |
+
else:
|
82 |
+
# 5G显存配置
|
83 |
+
x_pad = 1
|
84 |
+
x_query = 6
|
85 |
+
x_center = 38
|
86 |
+
x_max = 41
|
87 |
+
|
88 |
+
if self.gpu_mem != None and self.gpu_mem <= 4:
|
89 |
+
x_pad = 1
|
90 |
+
x_query = 5
|
91 |
+
x_center = 30
|
92 |
+
x_max = 32
|
93 |
+
|
94 |
+
return x_pad, x_query, x_center, x_max
|
95 |
+
|
96 |
+
|
97 |
+
f0up_key = sys.argv[1] #-12.0
|
98 |
+
input_path = sys.argv[2]
|
99 |
+
index_path = ''
|
100 |
+
abc = sys.argv[3]
|
101 |
+
f0method = sys.argv[4] # harvest or pm
|
102 |
+
opt_path = sys.argv[5]
|
103 |
+
model_path = sys.argv[6]
|
104 |
+
index_rate = float(sys.argv[7])
|
105 |
+
device = sys.argv[8]
|
106 |
+
device = None
|
107 |
+
is_half = bool(sys.argv[9])
|
108 |
+
is_half = None
|
109 |
+
filter_radius = int(sys.argv[10])
|
110 |
+
resample_sr = int(sys.argv[11])
|
111 |
+
rms_mix_rate = float(sys.argv[12])
|
112 |
+
protect = float(sys.argv[13])
|
113 |
+
print(sys.argv)
|
114 |
+
config = Config(device, is_half)
|
115 |
+
now_dir = os.getcwd()
|
116 |
+
sys.path.append(now_dir)
|
117 |
+
from vc_infer_pipeline import VC
|
118 |
+
from infer_pack.models import (
|
119 |
+
SynthesizerTrnMs256NSFsid,
|
120 |
+
SynthesizerTrnMs256NSFsid_nono,
|
121 |
+
SynthesizerTrnMs768NSFsid,
|
122 |
+
SynthesizerTrnMs768NSFsid_nono,
|
123 |
+
)
|
124 |
+
from my_utils import load_audio
|
125 |
+
from fairseq import checkpoint_utils
|
126 |
+
from scipy.io import wavfile
|
127 |
+
|
128 |
+
hubert_model = None
|
129 |
+
|
130 |
+
|
131 |
+
def load_hubert():
|
132 |
+
global hubert_model
|
133 |
+
models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
|
134 |
+
["hubert_base.pt"],
|
135 |
+
suffix="",
|
136 |
+
)
|
137 |
+
hubert_model = models[0]
|
138 |
+
hubert_model = hubert_model.to(device)
|
139 |
+
if is_half:
|
140 |
+
hubert_model = hubert_model.half()
|
141 |
+
else:
|
142 |
+
hubert_model = hubert_model.float()
|
143 |
+
hubert_model.eval()
|
144 |
+
|
145 |
+
|
146 |
+
def vc_single(sid, input_audio, f0_up_key, f0_file, f0_method, file_index, index_rate):
|
147 |
+
global tgt_sr, net_g, vc, hubert_model, version
|
148 |
+
if input_audio is None:
|
149 |
+
return "You need to upload an audio", None
|
150 |
+
f0_up_key = int(f0_up_key)
|
151 |
+
audio = load_audio(input_audio, 16000)
|
152 |
+
times = [0, 0, 0]
|
153 |
+
if hubert_model == None:
|
154 |
+
load_hubert()
|
155 |
+
if_f0 = cpt.get("f0", 1)
|
156 |
+
# audio_opt=vc.pipeline(hubert_model,net_g,sid,audio,times,f0_up_key,f0_method,file_index,file_big_npy,index_rate,if_f0,f0_file=f0_file)
|
157 |
+
audio_opt = vc.pipeline(
|
158 |
+
hubert_model,
|
159 |
+
net_g,
|
160 |
+
sid,
|
161 |
+
audio,
|
162 |
+
input_audio,
|
163 |
+
times,
|
164 |
+
f0_up_key,
|
165 |
+
f0_method,
|
166 |
+
file_index,
|
167 |
+
index_rate,
|
168 |
+
if_f0,
|
169 |
+
filter_radius,
|
170 |
+
tgt_sr,
|
171 |
+
resample_sr,
|
172 |
+
rms_mix_rate,
|
173 |
+
version,
|
174 |
+
protect,
|
175 |
+
f0_file=f0_file,
|
176 |
+
)
|
177 |
+
print(times)
|
178 |
+
return audio_opt
|
179 |
+
|
180 |
+
|
181 |
+
def get_vc(model_path):
|
182 |
+
global n_spk, tgt_sr, net_g, vc, cpt, device, is_half, version
|
183 |
+
print("loading pth %s" % model_path)
|
184 |
+
cpt = torch.load(model_path, map_location="cpu")
|
185 |
+
tgt_sr = cpt["config"][-1]
|
186 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
|
187 |
+
if_f0 = cpt.get("f0", 1)
|
188 |
+
version = cpt.get("version", "v1")
|
189 |
+
if version == "v1":
|
190 |
+
if if_f0 == 1:
|
191 |
+
net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=is_half)
|
192 |
+
else:
|
193 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
194 |
+
elif version == "v2":
|
195 |
+
if if_f0 == 1: #
|
196 |
+
net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=is_half)
|
197 |
+
else:
|
198 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
199 |
+
del net_g.enc_q
|
200 |
+
print(net_g.load_state_dict(cpt["weight"], strict=False)) # 不加这一行清不干净,真奇葩
|
201 |
+
net_g.eval().to(device)
|
202 |
+
if is_half:
|
203 |
+
net_g = net_g.half()
|
204 |
+
else:
|
205 |
+
net_g = net_g.float()
|
206 |
+
vc = VC(tgt_sr, config)
|
207 |
+
n_spk = cpt["config"][-3]
|
208 |
+
# return {"visible": True,"maximum": n_spk, "__type__": "update"}
|
209 |
+
|
210 |
+
|
211 |
+
get_vc(model_path)
|
212 |
+
audios = os.listdir(input_path)
|
213 |
+
for file in tq.tqdm(audios):
|
214 |
+
if file.endswith(".wav"):
|
215 |
+
file_path = input_path + "/" + file
|
216 |
+
wav_opt = vc_single(
|
217 |
+
0, file_path, f0up_key, None, f0method, index_path, index_rate
|
218 |
+
)
|
219 |
+
out_path = opt_path + "/" + file
|
220 |
+
wavfile.write(out_path, tgt_sr, wav_opt)
|
infer_uvr5.py
ADDED
@@ -0,0 +1,364 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os, sys, torch, warnings, pdb
|
2 |
+
|
3 |
+
now_dir = os.getcwd()
|
4 |
+
sys.path.append(now_dir)
|
5 |
+
from json import load as ll
|
6 |
+
|
7 |
+
warnings.filterwarnings("ignore")
|
8 |
+
import librosa
|
9 |
+
import importlib
|
10 |
+
import numpy as np
|
11 |
+
import hashlib, math
|
12 |
+
from tqdm import tqdm
|
13 |
+
from uvr5_pack.lib_v5 import spec_utils
|
14 |
+
from uvr5_pack.utils import _get_name_params, inference
|
15 |
+
from uvr5_pack.lib_v5.model_param_init import ModelParameters
|
16 |
+
import soundfile as sf
|
17 |
+
from uvr5_pack.lib_v5.nets_new import CascadedNet
|
18 |
+
from uvr5_pack.lib_v5 import nets_61968KB as nets
|
19 |
+
|
20 |
+
|
21 |
+
class _audio_pre_:
|
22 |
+
def __init__(self, agg, model_path, device, is_half):
|
23 |
+
self.model_path = model_path
|
24 |
+
self.device = device
|
25 |
+
self.data = {
|
26 |
+
# Processing Options
|
27 |
+
"postprocess": False,
|
28 |
+
"tta": False,
|
29 |
+
# Constants
|
30 |
+
"window_size": 512,
|
31 |
+
"agg": agg,
|
32 |
+
"high_end_process": "mirroring",
|
33 |
+
}
|
34 |
+
mp = ModelParameters("uvr5_pack/lib_v5/modelparams/4band_v2.json")
|
35 |
+
model = nets.CascadedASPPNet(mp.param["bins"] * 2)
|
36 |
+
cpk = torch.load(model_path, map_location="cpu")
|
37 |
+
model.load_state_dict(cpk)
|
38 |
+
model.eval()
|
39 |
+
if is_half:
|
40 |
+
model = model.half().to(device)
|
41 |
+
else:
|
42 |
+
model = model.to(device)
|
43 |
+
|
44 |
+
self.mp = mp
|
45 |
+
self.model = model
|
46 |
+
|
47 |
+
def _path_audio_(self, music_file, ins_root=None, vocal_root=None, format="flac"):
|
48 |
+
if ins_root is None and vocal_root is None:
|
49 |
+
return "No save root."
|
50 |
+
|
51 |
+
if ins_root is not None:
|
52 |
+
os.makedirs(ins_root, exist_ok=True)
|
53 |
+
if vocal_root is not None:
|
54 |
+
os.makedirs(vocal_root, exist_ok=True)
|
55 |
+
X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
|
56 |
+
bands_n = len(self.mp.param["band"])
|
57 |
+
# print(bands_n)
|
58 |
+
for d in range(bands_n, 0, -1):
|
59 |
+
bp = self.mp.param["band"][d]
|
60 |
+
if d == bands_n: # high-end band
|
61 |
+
(
|
62 |
+
X_wave[d],
|
63 |
+
_,
|
64 |
+
) = librosa.core.load(
|
65 |
+
music_file,
|
66 |
+
bp["sr"],
|
67 |
+
False,
|
68 |
+
dtype=np.float32,
|
69 |
+
res_type=bp["res_type"],
|
70 |
+
)
|
71 |
+
if X_wave[d].ndim == 1:
|
72 |
+
X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]])
|
73 |
+
else: # lower bands
|
74 |
+
X_wave[d] = librosa.core.resample(
|
75 |
+
X_wave[d + 1],
|
76 |
+
self.mp.param["band"][d + 1]["sr"],
|
77 |
+
bp["sr"],
|
78 |
+
res_type=bp["res_type"],
|
79 |
+
)
|
80 |
+
# Stft of wave source
|
81 |
+
X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(
|
82 |
+
X_wave[d],
|
83 |
+
bp["hl"],
|
84 |
+
bp["n_fft"],
|
85 |
+
self.mp.param["mid_side"],
|
86 |
+
self.mp.param["mid_side_b2"],
|
87 |
+
self.mp.param["reverse"],
|
88 |
+
)
|
89 |
+
# pdb.set_trace()
|
90 |
+
if d == bands_n and self.data["high_end_process"] != "none":
|
91 |
+
input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + (
|
92 |
+
self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"]
|
93 |
+
)
|
94 |
+
input_high_end = X_spec_s[d][
|
95 |
+
:, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, :
|
96 |
+
]
|
97 |
+
|
98 |
+
X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp)
|
99 |
+
aggresive_set = float(self.data["agg"] / 100)
|
100 |
+
aggressiveness = {
|
101 |
+
"value": aggresive_set,
|
102 |
+
"split_bin": self.mp.param["band"][1]["crop_stop"],
|
103 |
+
}
|
104 |
+
with torch.no_grad():
|
105 |
+
pred, X_mag, X_phase = inference(
|
106 |
+
X_spec_m, self.device, self.model, aggressiveness, self.data
|
107 |
+
)
|
108 |
+
# Postprocess
|
109 |
+
if self.data["postprocess"]:
|
110 |
+
pred_inv = np.clip(X_mag - pred, 0, np.inf)
|
111 |
+
pred = spec_utils.mask_silence(pred, pred_inv)
|
112 |
+
y_spec_m = pred * X_phase
|
113 |
+
v_spec_m = X_spec_m - y_spec_m
|
114 |
+
|
115 |
+
if ins_root is not None:
|
116 |
+
if self.data["high_end_process"].startswith("mirroring"):
|
117 |
+
input_high_end_ = spec_utils.mirroring(
|
118 |
+
self.data["high_end_process"], y_spec_m, input_high_end, self.mp
|
119 |
+
)
|
120 |
+
wav_instrument = spec_utils.cmb_spectrogram_to_wave(
|
121 |
+
y_spec_m, self.mp, input_high_end_h, input_high_end_
|
122 |
+
)
|
123 |
+
else:
|
124 |
+
wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp)
|
125 |
+
print("%s instruments done")
|
126 |
+
if format in ["wav", "flac"]:
|
127 |
+
sf.write(
|
128 |
+
os.path.join(
|
129 |
+
ins_root,
|
130 |
+
"instrument.{}".format(format),
|
131 |
+
),
|
132 |
+
(np.array(wav_instrument) * 32768).astype("int16"),
|
133 |
+
self.mp.param["sr"],
|
134 |
+
) #
|
135 |
+
else:
|
136 |
+
path = os.path.join(
|
137 |
+
ins_root, "instrument.wav"
|
138 |
+
)
|
139 |
+
sf.write(
|
140 |
+
path,
|
141 |
+
(np.array(wav_instrument) * 32768).astype("int16"),
|
142 |
+
self.mp.param["sr"],
|
143 |
+
)
|
144 |
+
if os.path.exists(path):
|
145 |
+
os.system(
|
146 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
147 |
+
% (path, path[:-4] + ".%s" % format)
|
148 |
+
)
|
149 |
+
if vocal_root is not None:
|
150 |
+
if self.data["high_end_process"].startswith("mirroring"):
|
151 |
+
input_high_end_ = spec_utils.mirroring(
|
152 |
+
self.data["high_end_process"], v_spec_m, input_high_end, self.mp
|
153 |
+
)
|
154 |
+
wav_vocals = spec_utils.cmb_spectrogram_to_wave(
|
155 |
+
v_spec_m, self.mp, input_high_end_h, input_high_end_
|
156 |
+
)
|
157 |
+
else:
|
158 |
+
wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp)
|
159 |
+
print("%s vocals done")
|
160 |
+
if format in ["wav", "flac"]:
|
161 |
+
sf.write(
|
162 |
+
os.path.join(
|
163 |
+
vocal_root,
|
164 |
+
"vocal.{}".format(format),
|
165 |
+
),
|
166 |
+
(np.array(wav_vocals) * 32768).astype("int16"),
|
167 |
+
self.mp.param["sr"],
|
168 |
+
)
|
169 |
+
else:
|
170 |
+
path = os.path.join(
|
171 |
+
vocal_root, "vocal.wav"
|
172 |
+
)
|
173 |
+
sf.write(
|
174 |
+
path,
|
175 |
+
(np.array(wav_vocals) * 32768).astype("int16"),
|
176 |
+
self.mp.param["sr"],
|
177 |
+
)
|
178 |
+
if os.path.exists(path):
|
179 |
+
os.system(
|
180 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
181 |
+
% (path, path[:-4] + ".%s" % format)
|
182 |
+
)
|
183 |
+
|
184 |
+
|
185 |
+
class _audio_pre_new:
|
186 |
+
def __init__(self, agg, model_path, device, is_half):
|
187 |
+
print('_audio_pre_new')
|
188 |
+
self.model_path = model_path
|
189 |
+
self.device = device
|
190 |
+
self.data = {
|
191 |
+
# Processing Options
|
192 |
+
"postprocess": False,
|
193 |
+
"tta": False,
|
194 |
+
# Constants
|
195 |
+
"window_size": 512,
|
196 |
+
"agg": agg,
|
197 |
+
"high_end_process": "mirroring",
|
198 |
+
}
|
199 |
+
mp = ModelParameters("uvr5_pack/lib_v5/modelparams/4band_v3.json")
|
200 |
+
nout = 64 if "DeReverb" in model_path else 48
|
201 |
+
model = CascadedNet(mp.param["bins"] * 2, nout)
|
202 |
+
cpk = torch.load(model_path, map_location="cpu")
|
203 |
+
model.load_state_dict(cpk)
|
204 |
+
model.eval()
|
205 |
+
if is_half:
|
206 |
+
model = model.half().to(device)
|
207 |
+
else:
|
208 |
+
model = model.to(device)
|
209 |
+
|
210 |
+
self.mp = mp
|
211 |
+
self.model = model
|
212 |
+
|
213 |
+
def _path_audio_(
|
214 |
+
self, music_file, vocal_root=None, ins_root=None, format="flac"
|
215 |
+
): # 3个VR模型vocal和ins是反的
|
216 |
+
if ins_root is None and vocal_root is None:
|
217 |
+
return "No save root."
|
218 |
+
name = os.path.basename(music_file)
|
219 |
+
if ins_root is not None:
|
220 |
+
os.makedirs(ins_root, exist_ok=True)
|
221 |
+
if vocal_root is not None:
|
222 |
+
os.makedirs(vocal_root, exist_ok=True)
|
223 |
+
X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
|
224 |
+
bands_n = len(self.mp.param["band"])
|
225 |
+
# print(bands_n)
|
226 |
+
for d in range(bands_n, 0, -1):
|
227 |
+
bp = self.mp.param["band"][d]
|
228 |
+
if d == bands_n: # high-end band
|
229 |
+
(
|
230 |
+
X_wave[d],
|
231 |
+
_,
|
232 |
+
) = librosa.core.load( # 理论上librosa读取可能对某些音频有bug,应该上ffmpeg读取,但是太麻烦了弃坑
|
233 |
+
music_file,
|
234 |
+
bp["sr"],
|
235 |
+
False,
|
236 |
+
dtype=np.float32,
|
237 |
+
res_type=bp["res_type"],
|
238 |
+
)
|
239 |
+
if X_wave[d].ndim == 1:
|
240 |
+
X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]])
|
241 |
+
else: # lower bands
|
242 |
+
X_wave[d] = librosa.core.resample(
|
243 |
+
X_wave[d + 1],
|
244 |
+
self.mp.param["band"][d + 1]["sr"],
|
245 |
+
bp["sr"],
|
246 |
+
res_type=bp["res_type"],
|
247 |
+
)
|
248 |
+
# Stft of wave source
|
249 |
+
X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(
|
250 |
+
X_wave[d],
|
251 |
+
bp["hl"],
|
252 |
+
bp["n_fft"],
|
253 |
+
self.mp.param["mid_side"],
|
254 |
+
self.mp.param["mid_side_b2"],
|
255 |
+
self.mp.param["reverse"],
|
256 |
+
)
|
257 |
+
# pdb.set_trace()
|
258 |
+
if d == bands_n and self.data["high_end_process"] != "none":
|
259 |
+
input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + (
|
260 |
+
self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"]
|
261 |
+
)
|
262 |
+
input_high_end = X_spec_s[d][
|
263 |
+
:, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, :
|
264 |
+
]
|
265 |
+
|
266 |
+
X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp)
|
267 |
+
aggresive_set = float(self.data["agg"] / 100)
|
268 |
+
aggressiveness = {
|
269 |
+
"value": aggresive_set,
|
270 |
+
"split_bin": self.mp.param["band"][1]["crop_stop"],
|
271 |
+
}
|
272 |
+
with torch.no_grad():
|
273 |
+
pred, X_mag, X_phase = inference(
|
274 |
+
X_spec_m, self.device, self.model, aggressiveness, self.data
|
275 |
+
)
|
276 |
+
# Postprocess
|
277 |
+
if self.data["postprocess"]:
|
278 |
+
pred_inv = np.clip(X_mag - pred, 0, np.inf)
|
279 |
+
pred = spec_utils.mask_silence(pred, pred_inv)
|
280 |
+
y_spec_m = pred * X_phase
|
281 |
+
v_spec_m = X_spec_m - y_spec_m
|
282 |
+
|
283 |
+
if ins_root is not None:
|
284 |
+
if self.data["high_end_process"].startswith("mirroring"):
|
285 |
+
input_high_end_ = spec_utils.mirroring(
|
286 |
+
self.data["high_end_process"], y_spec_m, input_high_end, self.mp
|
287 |
+
)
|
288 |
+
wav_instrument = spec_utils.cmb_spectrogram_to_wave(
|
289 |
+
y_spec_m, self.mp, input_high_end_h, input_high_end_
|
290 |
+
)
|
291 |
+
else:
|
292 |
+
wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp)
|
293 |
+
print("%s instruments done" % name)
|
294 |
+
if format in ["wav", "flac"]:
|
295 |
+
sf.write(
|
296 |
+
os.path.join(
|
297 |
+
ins_root,
|
298 |
+
"instrument_{}_{}.{}".format(name, self.data["agg"], format),
|
299 |
+
),
|
300 |
+
(np.array(wav_instrument) * 32768).astype("int16"),
|
301 |
+
self.mp.param["sr"],
|
302 |
+
) #
|
303 |
+
else:
|
304 |
+
path = os.path.join(
|
305 |
+
ins_root, "instrument_{}_{}.wav".format(name, self.data["agg"])
|
306 |
+
)
|
307 |
+
sf.write(
|
308 |
+
path,
|
309 |
+
(np.array(wav_instrument) * 32768).astype("int16"),
|
310 |
+
self.mp.param["sr"],
|
311 |
+
)
|
312 |
+
if os.path.exists(path):
|
313 |
+
os.system(
|
314 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
315 |
+
% (path, path[:-4] + ".%s" % format)
|
316 |
+
)
|
317 |
+
if vocal_root is not None:
|
318 |
+
if self.data["high_end_process"].startswith("mirroring"):
|
319 |
+
input_high_end_ = spec_utils.mirroring(
|
320 |
+
self.data["high_end_process"], v_spec_m, input_high_end, self.mp
|
321 |
+
)
|
322 |
+
wav_vocals = spec_utils.cmb_spectrogram_to_wave(
|
323 |
+
v_spec_m, self.mp, input_high_end_h, input_high_end_
|
324 |
+
)
|
325 |
+
else:
|
326 |
+
wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp)
|
327 |
+
print("%s vocals done" % name)
|
328 |
+
if format in ["wav", "flac"]:
|
329 |
+
sf.write(
|
330 |
+
os.path.join(
|
331 |
+
vocal_root,
|
332 |
+
"vocal_{}_{}.{}".format(name, self.data["agg"], format),
|
333 |
+
),
|
334 |
+
(np.array(wav_vocals) * 32768).astype("int16"),
|
335 |
+
self.mp.param["sr"],
|
336 |
+
)
|
337 |
+
else:
|
338 |
+
path = os.path.join(
|
339 |
+
vocal_root, "vocal_{}_{}.wav".format(name, self.data["agg"])
|
340 |
+
)
|
341 |
+
sf.write(
|
342 |
+
path,
|
343 |
+
(np.array(wav_vocals) * 32768).astype("int16"),
|
344 |
+
self.mp.param["sr"],
|
345 |
+
)
|
346 |
+
if os.path.exists(path):
|
347 |
+
os.system(
|
348 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
349 |
+
% (path, path[:-4] + ".%s" % format)
|
350 |
+
)
|
351 |
+
|
352 |
+
|
353 |
+
if __name__ == "__main__":
|
354 |
+
device = "cuda"
|
355 |
+
is_half = True
|
356 |
+
# model_path = "uvr5_weights/2_HP-UVR.pth"
|
357 |
+
# model_path = "uvr5_weights/VR-DeEchoDeReverb.pth"
|
358 |
+
# model_path = "uvr5_weights/VR-DeEchoNormal.pth"
|
359 |
+
model_path = "models/mymodelimran.pth"
|
360 |
+
# pre_fun = _audio_pre_(model_path=model_path, device=device, is_half=True,agg=10)
|
361 |
+
pre_fun = _audio_pre_new(model_path=model_path, device=device, is_half=True, agg=10)
|
362 |
+
audio_path = "audios/abc.mp3"
|
363 |
+
save_path = "results"
|
364 |
+
pre_fun._path_audio_(audio_path, save_path, save_path)
|
local.settings.json
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"IsEncrypted": false,
|
3 |
+
"Values": {
|
4 |
+
"AzureWebJobsStorage": "",
|
5 |
+
"FUNCTIONS_WORKER_RUNTIME": "python",
|
6 |
+
"AzureWebJobs.AltVoiceClone.Disabled": "true"
|
7 |
+
}
|
8 |
+
}
|
main.py
ADDED
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import functions_framework
|
2 |
+
import os
|
3 |
+
import logging
|
4 |
+
|
5 |
+
|
6 |
+
@functions_framework.http
|
7 |
+
def hello_http(request):
|
8 |
+
|
9 |
+
request_args = request.args
|
10 |
+
|
11 |
+
# Extract parameters from request
|
12 |
+
audio_file = request_args.get('audio_file')
|
13 |
+
model_name = request_args.get('model_name')
|
14 |
+
transform = request_args.get('transform')
|
15 |
+
song = request_args.get('song')
|
16 |
+
|
17 |
+
# Check if any parameter is None
|
18 |
+
if any(param is None for param in [audio_file, model_name, transform, song]):
|
19 |
+
return "Please provide all the required arguments: audio_file, model_name, transform, song."
|
20 |
+
else:
|
21 |
+
is_song = song.lower() == "true"
|
22 |
+
transform = int(transform)
|
23 |
+
import RVC_class
|
24 |
+
|
25 |
+
|
26 |
+
is_song = song.lower() == "true"
|
27 |
+
|
28 |
+
# Create an instance of VoiceConverter
|
29 |
+
converter = RVC_class.VoiceConverter()
|
30 |
+
|
31 |
+
# Call single_run method and get the result
|
32 |
+
result = converter.single_run(audio_file, model_name, transform, is_song)
|
33 |
+
|
34 |
+
return result
|
35 |
+
|
36 |
+
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
|
my_utils.py
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import ffmpeg
|
2 |
+
import numpy as np
|
3 |
+
import requests
|
4 |
+
import logging
|
5 |
+
import wave
|
6 |
+
import librosa
|
7 |
+
|
8 |
+
def load_audio(file_path, sr):
|
9 |
+
try:
|
10 |
+
with wave.open(file_path, 'rb') as audio_file:
|
11 |
+
channels = audio_file.getnchannels()
|
12 |
+
sample_width = audio_file.getsampwidth()
|
13 |
+
frame_rate = audio_file.getframerate()
|
14 |
+
frames = audio_file.readframes(audio_file.getnframes())
|
15 |
+
|
16 |
+
audio_data = np.frombuffer(frames, dtype=np.int16)
|
17 |
+
audio_data = audio_data.astype(np.float32) / np.iinfo(np.int16).max # Convert to float and normalize to the range [-1, 1]
|
18 |
+
|
19 |
+
# Resample the audio if the sample rate is different
|
20 |
+
if frame_rate != sr:
|
21 |
+
audio_data = librosa.resample(audio_data, orig_sr=frame_rate, target_sr=sr)
|
22 |
+
|
23 |
+
# Perform any required audio processing or conversion
|
24 |
+
# ...
|
25 |
+
|
26 |
+
except Exception as e:
|
27 |
+
raise RuntimeError(f"Failed to load audio: {e}")
|
28 |
+
|
29 |
+
return audio_data
|
30 |
+
|
31 |
+
|
32 |
+
|
33 |
+
|
34 |
+
|
35 |
+
|
onnx_inference_demo.py
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import soundfile
|
2 |
+
from infer_pack.onnx_inference import OnnxRVC
|
3 |
+
|
4 |
+
hop_size = 512
|
5 |
+
sampling_rate = 40000 # 采样率
|
6 |
+
f0_up_key = 0 # 升降调
|
7 |
+
sid = 0 # 角色ID
|
8 |
+
f0_method = "dio" # F0提取算法
|
9 |
+
model_path = "ShirohaRVC.onnx" # 模型的完整路径
|
10 |
+
vec_name = "vec-256-layer-9" # 内部自动补齐为 f"pretrained/{vec_name}.onnx" 需要onnx的vec模型
|
11 |
+
wav_path = "123.wav" # 输入路径或ByteIO实例
|
12 |
+
out_path = "out.wav" # 输出路径或ByteIO实例
|
13 |
+
|
14 |
+
model = OnnxRVC(
|
15 |
+
model_path, vec_path=vec_name, sr=sampling_rate, hop_size=hop_size, device="cuda"
|
16 |
+
)
|
17 |
+
|
18 |
+
audio = model.inference(wav_path, sid, f0_method=f0_method, f0_up_key=f0_up_key)
|
19 |
+
|
20 |
+
soundfile.write(out_path, audio, sampling_rate)
|
poetry.lock
ADDED
The diff for this file is too large to render.
See raw diff
|
|
pyproject.toml
ADDED
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[tool.poetry]
|
2 |
+
name = "rvc-beta"
|
3 |
+
version = "0.1.0"
|
4 |
+
description = ""
|
5 |
+
authors = ["lj1995"]
|
6 |
+
license = "MIT"
|
7 |
+
|
8 |
+
[tool.poetry.dependencies]
|
9 |
+
python = "^3.8"
|
10 |
+
torch = "^2.0.0"
|
11 |
+
torchaudio = "^2.0.1"
|
12 |
+
Cython = "^0.29.34"
|
13 |
+
gradio = "^3.24.1"
|
14 |
+
future = "^0.18.3"
|
15 |
+
pydub = "^0.25.1"
|
16 |
+
soundfile = "^0.12.1"
|
17 |
+
ffmpeg-python = "^0.2.0"
|
18 |
+
tensorboardX = "^2.6"
|
19 |
+
functorch = "^2.0.0"
|
20 |
+
fairseq = "^0.12.2"
|
21 |
+
faiss-cpu = "^1.7.2"
|
22 |
+
Jinja2 = "^3.1.2"
|
23 |
+
json5 = "^0.9.11"
|
24 |
+
librosa = "0.9.2"
|
25 |
+
llvmlite = "0.39.0"
|
26 |
+
Markdown = "^3.4.3"
|
27 |
+
matplotlib = "^3.7.1"
|
28 |
+
matplotlib-inline = "^0.1.6"
|
29 |
+
numba = "0.56.4"
|
30 |
+
numpy = "1.23.5"
|
31 |
+
scipy = "1.9.3"
|
32 |
+
praat-parselmouth = "^0.4.3"
|
33 |
+
Pillow = "9.3.0"
|
34 |
+
pyworld = "^0.3.2"
|
35 |
+
resampy = "^0.4.2"
|
36 |
+
scikit-learn = "^1.2.2"
|
37 |
+
starlette = "^0.27.0"
|
38 |
+
tensorboard = "^2.12.1"
|
39 |
+
tensorboard-data-server = "^0.7.0"
|
40 |
+
tensorboard-plugin-wit = "^1.8.1"
|
41 |
+
torchgen = "^0.0.1"
|
42 |
+
tqdm = "^4.65.0"
|
43 |
+
tornado = "^6.3"
|
44 |
+
Werkzeug = "^2.2.3"
|
45 |
+
uc-micro-py = "^1.0.1"
|
46 |
+
sympy = "^1.11.1"
|
47 |
+
tabulate = "^0.9.0"
|
48 |
+
PyYAML = "^6.0"
|
49 |
+
pyasn1 = "^0.4.8"
|
50 |
+
pyasn1-modules = "^0.2.8"
|
51 |
+
fsspec = "^2023.3.0"
|
52 |
+
absl-py = "^1.4.0"
|
53 |
+
audioread = "^3.0.0"
|
54 |
+
uvicorn = "^0.21.1"
|
55 |
+
colorama = "^0.4.6"
|
56 |
+
|
57 |
+
[tool.poetry.dev-dependencies]
|
58 |
+
|
59 |
+
[build-system]
|
60 |
+
requires = ["poetry-core>=1.0.0"]
|
61 |
+
build-backend = "poetry.core.masonry.api"
|
requirements-win-for-realtime_vc_gui.txt
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#1.Install torch from pytorch.org:
|
2 |
+
#torch 2.0 with cuda 11.8
|
3 |
+
#pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
|
4 |
+
#torch 1.11.0 with cuda 11.3
|
5 |
+
#pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
|
6 |
+
einops
|
7 |
+
fairseq
|
8 |
+
flask
|
9 |
+
flask_cors
|
10 |
+
gin
|
11 |
+
gin_config
|
12 |
+
librosa
|
13 |
+
local_attention
|
14 |
+
matplotlib
|
15 |
+
praat-parselmouth
|
16 |
+
pyworld
|
17 |
+
PyYAML
|
18 |
+
resampy
|
19 |
+
scikit_learn
|
20 |
+
scipy
|
21 |
+
SoundFile
|
22 |
+
tensorboard
|
23 |
+
tqdm
|
24 |
+
wave
|
25 |
+
PySimpleGUI
|
26 |
+
sounddevice
|
27 |
+
gradio
|
28 |
+
noisereduce
|
requirements.txt
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# DO NOT include azure-functions-worker in this file
|
2 |
+
# The Python Worker is managed by Azure Functions platform
|
3 |
+
# Manually managing azure-functions-worker may cause unexpected issues
|
4 |
+
|
5 |
+
functions-framework
|
6 |
+
joblib>=1.1.0
|
7 |
+
numba==0.56.4
|
8 |
+
numpy==1.23.5
|
9 |
+
scipy==1.9.3
|
10 |
+
librosa==0.9.1
|
11 |
+
llvmlite==0.39.0
|
12 |
+
fairseq==0.12.2
|
13 |
+
faiss-cpu==1.7.3
|
14 |
+
gradio==3.14.0
|
15 |
+
Cython
|
16 |
+
pydub>=0.25.1
|
17 |
+
soundfile>=0.12.1
|
18 |
+
ffmpeg-python>=0.2.0
|
19 |
+
tensorboardX
|
20 |
+
Jinja2>=3.1.2
|
21 |
+
json5
|
22 |
+
Markdown
|
23 |
+
matplotlib>=3.7.0
|
24 |
+
matplotlib-inline>=0.1.3
|
25 |
+
praat-parselmouth>=0.4.2
|
26 |
+
Pillow>=9.1.1
|
27 |
+
resampy>=0.4.2
|
28 |
+
scikit-learn
|
29 |
+
starlette>=0.25.0
|
30 |
+
tensorboard
|
31 |
+
tensorboard-data-server
|
32 |
+
tensorboard-plugin-wit
|
33 |
+
torchgen>=0.0.1
|
34 |
+
torch==2.0.0
|
35 |
+
tqdm>=4.63.1
|
36 |
+
tornado>=6.1
|
37 |
+
Werkzeug>=2.2.3
|
38 |
+
uc-micro-py>=1.0.1
|
39 |
+
sympy>=1.11.1
|
40 |
+
tabulate>=0.8.10
|
41 |
+
PyYAML>=6.0
|
42 |
+
pyasn1>=0.4.8
|
43 |
+
pyasn1-modules>=0.2.8
|
44 |
+
fsspec>=2022.11.0
|
45 |
+
absl-py>=1.2.0
|
46 |
+
audioread
|
47 |
+
uvicorn>=0.21.1
|
48 |
+
colorama>=0.4.5
|
49 |
+
pyworld>=0.3.2
|
50 |
+
httpx==0.23.0
|
51 |
+
onnxruntime-gpu
|
52 |
+
torchcrepe==0.0.20
|
53 |
+
wave
|
54 |
+
pydub
|
55 |
+
requests
|
slicer2.py
ADDED
@@ -0,0 +1,260 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
|
3 |
+
|
4 |
+
# This function is obtained from librosa.
|
5 |
+
def get_rms(
|
6 |
+
y,
|
7 |
+
frame_length=2048,
|
8 |
+
hop_length=512,
|
9 |
+
pad_mode="constant",
|
10 |
+
):
|
11 |
+
padding = (int(frame_length // 2), int(frame_length // 2))
|
12 |
+
y = np.pad(y, padding, mode=pad_mode)
|
13 |
+
|
14 |
+
axis = -1
|
15 |
+
# put our new within-frame axis at the end for now
|
16 |
+
out_strides = y.strides + tuple([y.strides[axis]])
|
17 |
+
# Reduce the shape on the framing axis
|
18 |
+
x_shape_trimmed = list(y.shape)
|
19 |
+
x_shape_trimmed[axis] -= frame_length - 1
|
20 |
+
out_shape = tuple(x_shape_trimmed) + tuple([frame_length])
|
21 |
+
xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)
|
22 |
+
if axis < 0:
|
23 |
+
target_axis = axis - 1
|
24 |
+
else:
|
25 |
+
target_axis = axis + 1
|
26 |
+
xw = np.moveaxis(xw, -1, target_axis)
|
27 |
+
# Downsample along the target axis
|
28 |
+
slices = [slice(None)] * xw.ndim
|
29 |
+
slices[axis] = slice(0, None, hop_length)
|
30 |
+
x = xw[tuple(slices)]
|
31 |
+
|
32 |
+
# Calculate power
|
33 |
+
power = np.mean(np.abs(x) ** 2, axis=-2, keepdims=True)
|
34 |
+
|
35 |
+
return np.sqrt(power)
|
36 |
+
|
37 |
+
|
38 |
+
class Slicer:
|
39 |
+
def __init__(
|
40 |
+
self,
|
41 |
+
sr: int,
|
42 |
+
threshold: float = -40.0,
|
43 |
+
min_length: int = 5000,
|
44 |
+
min_interval: int = 300,
|
45 |
+
hop_size: int = 20,
|
46 |
+
max_sil_kept: int = 5000,
|
47 |
+
):
|
48 |
+
if not min_length >= min_interval >= hop_size:
|
49 |
+
raise ValueError(
|
50 |
+
"The following condition must be satisfied: min_length >= min_interval >= hop_size"
|
51 |
+
)
|
52 |
+
if not max_sil_kept >= hop_size:
|
53 |
+
raise ValueError(
|
54 |
+
"The following condition must be satisfied: max_sil_kept >= hop_size"
|
55 |
+
)
|
56 |
+
min_interval = sr * min_interval / 1000
|
57 |
+
self.threshold = 10 ** (threshold / 20.0)
|
58 |
+
self.hop_size = round(sr * hop_size / 1000)
|
59 |
+
self.win_size = min(round(min_interval), 4 * self.hop_size)
|
60 |
+
self.min_length = round(sr * min_length / 1000 / self.hop_size)
|
61 |
+
self.min_interval = round(min_interval / self.hop_size)
|
62 |
+
self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
|
63 |
+
|
64 |
+
def _apply_slice(self, waveform, begin, end):
|
65 |
+
if len(waveform.shape) > 1:
|
66 |
+
return waveform[
|
67 |
+
:, begin * self.hop_size : min(waveform.shape[1], end * self.hop_size)
|
68 |
+
]
|
69 |
+
else:
|
70 |
+
return waveform[
|
71 |
+
begin * self.hop_size : min(waveform.shape[0], end * self.hop_size)
|
72 |
+
]
|
73 |
+
|
74 |
+
# @timeit
|
75 |
+
def slice(self, waveform):
|
76 |
+
if len(waveform.shape) > 1:
|
77 |
+
samples = waveform.mean(axis=0)
|
78 |
+
else:
|
79 |
+
samples = waveform
|
80 |
+
if samples.shape[0] <= self.min_length:
|
81 |
+
return [waveform]
|
82 |
+
rms_list = get_rms(
|
83 |
+
y=samples, frame_length=self.win_size, hop_length=self.hop_size
|
84 |
+
).squeeze(0)
|
85 |
+
sil_tags = []
|
86 |
+
silence_start = None
|
87 |
+
clip_start = 0
|
88 |
+
for i, rms in enumerate(rms_list):
|
89 |
+
# Keep looping while frame is silent.
|
90 |
+
if rms < self.threshold:
|
91 |
+
# Record start of silent frames.
|
92 |
+
if silence_start is None:
|
93 |
+
silence_start = i
|
94 |
+
continue
|
95 |
+
# Keep looping while frame is not silent and silence start has not been recorded.
|
96 |
+
if silence_start is None:
|
97 |
+
continue
|
98 |
+
# Clear recorded silence start if interval is not enough or clip is too short
|
99 |
+
is_leading_silence = silence_start == 0 and i > self.max_sil_kept
|
100 |
+
need_slice_middle = (
|
101 |
+
i - silence_start >= self.min_interval
|
102 |
+
and i - clip_start >= self.min_length
|
103 |
+
)
|
104 |
+
if not is_leading_silence and not need_slice_middle:
|
105 |
+
silence_start = None
|
106 |
+
continue
|
107 |
+
# Need slicing. Record the range of silent frames to be removed.
|
108 |
+
if i - silence_start <= self.max_sil_kept:
|
109 |
+
pos = rms_list[silence_start : i + 1].argmin() + silence_start
|
110 |
+
if silence_start == 0:
|
111 |
+
sil_tags.append((0, pos))
|
112 |
+
else:
|
113 |
+
sil_tags.append((pos, pos))
|
114 |
+
clip_start = pos
|
115 |
+
elif i - silence_start <= self.max_sil_kept * 2:
|
116 |
+
pos = rms_list[
|
117 |
+
i - self.max_sil_kept : silence_start + self.max_sil_kept + 1
|
118 |
+
].argmin()
|
119 |
+
pos += i - self.max_sil_kept
|
120 |
+
pos_l = (
|
121 |
+
rms_list[
|
122 |
+
silence_start : silence_start + self.max_sil_kept + 1
|
123 |
+
].argmin()
|
124 |
+
+ silence_start
|
125 |
+
)
|
126 |
+
pos_r = (
|
127 |
+
rms_list[i - self.max_sil_kept : i + 1].argmin()
|
128 |
+
+ i
|
129 |
+
- self.max_sil_kept
|
130 |
+
)
|
131 |
+
if silence_start == 0:
|
132 |
+
sil_tags.append((0, pos_r))
|
133 |
+
clip_start = pos_r
|
134 |
+
else:
|
135 |
+
sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
|
136 |
+
clip_start = max(pos_r, pos)
|
137 |
+
else:
|
138 |
+
pos_l = (
|
139 |
+
rms_list[
|
140 |
+
silence_start : silence_start + self.max_sil_kept + 1
|
141 |
+
].argmin()
|
142 |
+
+ silence_start
|
143 |
+
)
|
144 |
+
pos_r = (
|
145 |
+
rms_list[i - self.max_sil_kept : i + 1].argmin()
|
146 |
+
+ i
|
147 |
+
- self.max_sil_kept
|
148 |
+
)
|
149 |
+
if silence_start == 0:
|
150 |
+
sil_tags.append((0, pos_r))
|
151 |
+
else:
|
152 |
+
sil_tags.append((pos_l, pos_r))
|
153 |
+
clip_start = pos_r
|
154 |
+
silence_start = None
|
155 |
+
# Deal with trailing silence.
|
156 |
+
total_frames = rms_list.shape[0]
|
157 |
+
if (
|
158 |
+
silence_start is not None
|
159 |
+
and total_frames - silence_start >= self.min_interval
|
160 |
+
):
|
161 |
+
silence_end = min(total_frames, silence_start + self.max_sil_kept)
|
162 |
+
pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start
|
163 |
+
sil_tags.append((pos, total_frames + 1))
|
164 |
+
# Apply and return slices.
|
165 |
+
if len(sil_tags) == 0:
|
166 |
+
return [waveform]
|
167 |
+
else:
|
168 |
+
chunks = []
|
169 |
+
if sil_tags[0][0] > 0:
|
170 |
+
chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))
|
171 |
+
for i in range(len(sil_tags) - 1):
|
172 |
+
chunks.append(
|
173 |
+
self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0])
|
174 |
+
)
|
175 |
+
if sil_tags[-1][1] < total_frames:
|
176 |
+
chunks.append(
|
177 |
+
self._apply_slice(waveform, sil_tags[-1][1], total_frames)
|
178 |
+
)
|
179 |
+
return chunks
|
180 |
+
|
181 |
+
|
182 |
+
def main():
|
183 |
+
import os.path
|
184 |
+
from argparse import ArgumentParser
|
185 |
+
|
186 |
+
import librosa
|
187 |
+
import soundfile
|
188 |
+
|
189 |
+
parser = ArgumentParser()
|
190 |
+
parser.add_argument("audio", type=str, help="The audio to be sliced")
|
191 |
+
parser.add_argument(
|
192 |
+
"--out", type=str, help="Output directory of the sliced audio clips"
|
193 |
+
)
|
194 |
+
parser.add_argument(
|
195 |
+
"--db_thresh",
|
196 |
+
type=float,
|
197 |
+
required=False,
|
198 |
+
default=-40,
|
199 |
+
help="The dB threshold for silence detection",
|
200 |
+
)
|
201 |
+
parser.add_argument(
|
202 |
+
"--min_length",
|
203 |
+
type=int,
|
204 |
+
required=False,
|
205 |
+
default=5000,
|
206 |
+
help="The minimum milliseconds required for each sliced audio clip",
|
207 |
+
)
|
208 |
+
parser.add_argument(
|
209 |
+
"--min_interval",
|
210 |
+
type=int,
|
211 |
+
required=False,
|
212 |
+
default=300,
|
213 |
+
help="The minimum milliseconds for a silence part to be sliced",
|
214 |
+
)
|
215 |
+
parser.add_argument(
|
216 |
+
"--hop_size",
|
217 |
+
type=int,
|
218 |
+
required=False,
|
219 |
+
default=10,
|
220 |
+
help="Frame length in milliseconds",
|
221 |
+
)
|
222 |
+
parser.add_argument(
|
223 |
+
"--max_sil_kept",
|
224 |
+
type=int,
|
225 |
+
required=False,
|
226 |
+
default=500,
|
227 |
+
help="The maximum silence length kept around the sliced clip, presented in milliseconds",
|
228 |
+
)
|
229 |
+
args = parser.parse_args()
|
230 |
+
out = args.out
|
231 |
+
if out is None:
|
232 |
+
out = os.path.dirname(os.path.abspath(args.audio))
|
233 |
+
audio, sr = librosa.load(args.audio, sr=None, mono=False)
|
234 |
+
slicer = Slicer(
|
235 |
+
sr=sr,
|
236 |
+
threshold=args.db_thresh,
|
237 |
+
min_length=args.min_length,
|
238 |
+
min_interval=args.min_interval,
|
239 |
+
hop_size=args.hop_size,
|
240 |
+
max_sil_kept=args.max_sil_kept,
|
241 |
+
)
|
242 |
+
chunks = slicer.slice(audio)
|
243 |
+
if not os.path.exists(out):
|
244 |
+
os.makedirs(out)
|
245 |
+
for i, chunk in enumerate(chunks):
|
246 |
+
if len(chunk.shape) > 1:
|
247 |
+
chunk = chunk.T
|
248 |
+
soundfile.write(
|
249 |
+
os.path.join(
|
250 |
+
out,
|
251 |
+
f"%s_%d.wav"
|
252 |
+
% (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
|
253 |
+
),
|
254 |
+
chunk,
|
255 |
+
sr,
|
256 |
+
)
|
257 |
+
|
258 |
+
|
259 |
+
if __name__ == "__main__":
|
260 |
+
main()
|
temp.py
ADDED
@@ -0,0 +1,1392 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import shutil
|
3 |
+
import sys
|
4 |
+
import tempfile
|
5 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = ""
|
6 |
+
now_dir = os.getcwd()
|
7 |
+
sys.path.append(now_dir)
|
8 |
+
import traceback, pdb
|
9 |
+
import warnings
|
10 |
+
|
11 |
+
import numpy as np
|
12 |
+
import torch
|
13 |
+
|
14 |
+
os.environ['OPENBLAS_NUM_THREADS'] = '1'
|
15 |
+
os.environ["no_proxy"] = "localhost, 127.0.0.1, ::1"
|
16 |
+
import logging
|
17 |
+
import threading
|
18 |
+
from random import shuffle
|
19 |
+
from subprocess import Popen
|
20 |
+
from time import sleep
|
21 |
+
|
22 |
+
import faiss
|
23 |
+
import ffmpeg
|
24 |
+
import gradio as gr
|
25 |
+
import soundfile as sf
|
26 |
+
from config import Config
|
27 |
+
from fairseq import checkpoint_utils
|
28 |
+
from i18n import I18nAuto
|
29 |
+
from infer_pack.models import (
|
30 |
+
SynthesizerTrnMs256NSFsid,
|
31 |
+
SynthesizerTrnMs256NSFsid_nono,
|
32 |
+
SynthesizerTrnMs768NSFsid,
|
33 |
+
SynthesizerTrnMs768NSFsid_nono,
|
34 |
+
)
|
35 |
+
from infer_pack.models_onnx import SynthesizerTrnMsNSFsidM
|
36 |
+
from infer_uvr5 import _audio_pre_, _audio_pre_new
|
37 |
+
from MDXNet import MDXNetDereverb
|
38 |
+
from my_utils import load_audio
|
39 |
+
from train.process_ckpt import change_info, extract_small_model, merge, show_info
|
40 |
+
from vc_infer_pipeline import VC
|
41 |
+
from sklearn.cluster import MiniBatchKMeans
|
42 |
+
|
43 |
+
logging.getLogger("numba").setLevel(logging.WARNING)
|
44 |
+
|
45 |
+
|
46 |
+
tmp = os.path.join(now_dir, "TEMP")
|
47 |
+
shutil.rmtree(tmp, ignore_errors=True)
|
48 |
+
shutil.rmtree("%s/runtime/Lib/site-packages/infer_pack" % (now_dir), ignore_errors=True)
|
49 |
+
shutil.rmtree("%s/runtime/Lib/site-packages/uvr5_pack" % (now_dir), ignore_errors=True)
|
50 |
+
os.makedirs(tmp, exist_ok=True)
|
51 |
+
os.makedirs(os.path.join(now_dir, "logs"), exist_ok=True)
|
52 |
+
os.makedirs(os.path.join(now_dir, "weights"), exist_ok=True)
|
53 |
+
os.environ["TEMP"] = tmp
|
54 |
+
warnings.filterwarnings("ignore")
|
55 |
+
torch.manual_seed(114514)
|
56 |
+
from scipy.io import wavfile
|
57 |
+
|
58 |
+
config = Config()
|
59 |
+
i18n = I18nAuto()
|
60 |
+
i18n.print()
|
61 |
+
# 判断是否有能用来训练和加速推理的N卡
|
62 |
+
ngpu = torch.cuda.device_count()
|
63 |
+
gpu_infos = []
|
64 |
+
mem = []
|
65 |
+
if_gpu_ok = False
|
66 |
+
|
67 |
+
if torch.cuda.is_available() or ngpu != 0:
|
68 |
+
for i in range(ngpu):
|
69 |
+
gpu_name = torch.cuda.get_device_name(i)
|
70 |
+
if any(
|
71 |
+
value in gpu_name.upper()
|
72 |
+
for value in [
|
73 |
+
"10",
|
74 |
+
"16",
|
75 |
+
"20",
|
76 |
+
"30",
|
77 |
+
"40",
|
78 |
+
"A2",
|
79 |
+
"A3",
|
80 |
+
"A4",
|
81 |
+
"P4",
|
82 |
+
"A50",
|
83 |
+
"500",
|
84 |
+
"A60",
|
85 |
+
"70",
|
86 |
+
"80",
|
87 |
+
"90",
|
88 |
+
"M4",
|
89 |
+
"T4",
|
90 |
+
"TITAN",
|
91 |
+
]
|
92 |
+
):
|
93 |
+
# A10#A100#V100#A40#P40#M40#K80#A4500
|
94 |
+
if_gpu_ok = True # 至少有一张能用的N卡
|
95 |
+
gpu_infos.append("%s\t%s" % (i, gpu_name))
|
96 |
+
mem.append(
|
97 |
+
int(
|
98 |
+
torch.cuda.get_device_properties(i).total_memory
|
99 |
+
/ 1024
|
100 |
+
/ 1024
|
101 |
+
/ 1024
|
102 |
+
+ 0.4
|
103 |
+
)
|
104 |
+
)
|
105 |
+
if if_gpu_ok and len(gpu_infos) > 0:
|
106 |
+
gpu_info = "\n".join(gpu_infos)
|
107 |
+
default_batch_size = 1
|
108 |
+
else:
|
109 |
+
gpu_info = i18n("很遗憾您这没有能用的显卡来支持您训练")
|
110 |
+
default_batch_size = 1
|
111 |
+
gpus = "-".join([i[0] for i in gpu_infos])
|
112 |
+
|
113 |
+
|
114 |
+
class ToolButton(gr.Button, gr.components.FormComponent):
|
115 |
+
"""Small button with single emoji as text, fits inside gradio forms"""
|
116 |
+
|
117 |
+
def __init__(self, **kwargs):
|
118 |
+
super().__init__(variant="tool", **kwargs)
|
119 |
+
|
120 |
+
def get_block_name(self):
|
121 |
+
return "button"
|
122 |
+
|
123 |
+
|
124 |
+
hubert_model = None
|
125 |
+
|
126 |
+
|
127 |
+
def load_hubert():
|
128 |
+
global hubert_model
|
129 |
+
models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
|
130 |
+
["hubert_base.pt"],
|
131 |
+
suffix="",
|
132 |
+
)
|
133 |
+
hubert_model = models[0]
|
134 |
+
hubert_model = hubert_model.to(config.device)
|
135 |
+
if config.is_half:
|
136 |
+
hubert_model = hubert_model.half()
|
137 |
+
else:
|
138 |
+
hubert_model = hubert_model.float()
|
139 |
+
hubert_model.eval()
|
140 |
+
|
141 |
+
|
142 |
+
weight_root = "weights"
|
143 |
+
weight_uvr5_root = "uvr5_weights"
|
144 |
+
index_root = "logs"
|
145 |
+
names = []
|
146 |
+
for name in os.listdir(weight_root):
|
147 |
+
if name.endswith(".pth"):
|
148 |
+
names.append(name)
|
149 |
+
index_paths = []
|
150 |
+
for root, dirs, files in os.walk(index_root, topdown=False):
|
151 |
+
for name in files:
|
152 |
+
if name.endswith(".index") and "trained" not in name:
|
153 |
+
index_paths.append("%s/%s" % (root, name))
|
154 |
+
uvr5_names = []
|
155 |
+
for name in os.listdir(weight_uvr5_root):
|
156 |
+
if name.endswith(".pth") or "onnx" in name:
|
157 |
+
uvr5_names.append(name.replace(".pth", ""))
|
158 |
+
|
159 |
+
|
160 |
+
def vc_single(
|
161 |
+
sid,
|
162 |
+
input_audio_path,
|
163 |
+
f0_up_key,
|
164 |
+
f0_file,
|
165 |
+
f0_method,
|
166 |
+
file_index,
|
167 |
+
file_index2,
|
168 |
+
# file_big_npy,
|
169 |
+
index_rate,
|
170 |
+
filter_radius,
|
171 |
+
resample_sr,
|
172 |
+
rms_mix_rate,
|
173 |
+
protect,
|
174 |
+
): # spk_item, input_audio0, vc_transform0,f0_file,f0method0
|
175 |
+
global tgt_sr, net_g, vc, hubert_model, version
|
176 |
+
print(f0_up_key)
|
177 |
+
|
178 |
+
if input_audio_path is None:
|
179 |
+
return "You need to upload an audio", None
|
180 |
+
print("input_audio_path: ", input_audio_path)
|
181 |
+
print("f0_up_key: ", f0_up_key)
|
182 |
+
f0_up_key = int(f0_up_key)
|
183 |
+
try:
|
184 |
+
audio = load_audio(input_audio_path, 16000)
|
185 |
+
audio_max = np.abs(audio).max() / 0.95
|
186 |
+
if audio_max > 1:
|
187 |
+
audio /= audio_max
|
188 |
+
times = [0, 0, 0]
|
189 |
+
if not hubert_model:
|
190 |
+
load_hubert()
|
191 |
+
if_f0 = cpt.get("f0", 1)
|
192 |
+
file_index = (
|
193 |
+
(
|
194 |
+
file_index.strip(" ")
|
195 |
+
.strip('"')
|
196 |
+
.strip("\n")
|
197 |
+
.strip('"')
|
198 |
+
.strip(" ")
|
199 |
+
.replace("trained", "added")
|
200 |
+
)
|
201 |
+
if file_index != ""
|
202 |
+
else file_index2
|
203 |
+
) # 防止小白写错,自动帮他替换掉
|
204 |
+
# file_big_npy = (
|
205 |
+
# file_big_npy.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
206 |
+
# )
|
207 |
+
audio_opt = vc.pipeline(
|
208 |
+
hubert_model,
|
209 |
+
net_g,
|
210 |
+
sid,
|
211 |
+
audio,
|
212 |
+
input_audio_path,
|
213 |
+
times,
|
214 |
+
f0_up_key,
|
215 |
+
f0_method,
|
216 |
+
file_index,
|
217 |
+
# file_big_npy,
|
218 |
+
index_rate,
|
219 |
+
if_f0,
|
220 |
+
filter_radius,
|
221 |
+
tgt_sr,
|
222 |
+
resample_sr,
|
223 |
+
rms_mix_rate,
|
224 |
+
version,
|
225 |
+
protect,
|
226 |
+
f0_file=f0_file,
|
227 |
+
)
|
228 |
+
print(f0_up_key)
|
229 |
+
|
230 |
+
if tgt_sr != resample_sr >= 16000:
|
231 |
+
tgt_sr = resample_sr
|
232 |
+
index_info = (
|
233 |
+
"Using index:%s." % file_index
|
234 |
+
if os.path.exists(file_index)
|
235 |
+
else "Index not used."
|
236 |
+
)
|
237 |
+
return "Success.\n %s\nTime:\n npy:%ss, f0:%ss, infer:%ss" % (
|
238 |
+
index_info,
|
239 |
+
times[0],
|
240 |
+
times[1],
|
241 |
+
times[2],
|
242 |
+
), (tgt_sr, audio_opt)
|
243 |
+
except:
|
244 |
+
info = traceback.format_exc()
|
245 |
+
print(info)
|
246 |
+
return info, (None, None)
|
247 |
+
|
248 |
+
|
249 |
+
def vc_multi(
|
250 |
+
sid,
|
251 |
+
dir_path,
|
252 |
+
opt_root,
|
253 |
+
paths,
|
254 |
+
f0_up_key,
|
255 |
+
f0_method,
|
256 |
+
file_index,
|
257 |
+
file_index2,
|
258 |
+
# file_big_npy,
|
259 |
+
index_rate,
|
260 |
+
filter_radius,
|
261 |
+
resample_sr,
|
262 |
+
rms_mix_rate,
|
263 |
+
protect,
|
264 |
+
format1,
|
265 |
+
):
|
266 |
+
try:
|
267 |
+
print(f0_up_key)
|
268 |
+
|
269 |
+
dir_path = (
|
270 |
+
dir_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
271 |
+
) # 防止小白拷路径头尾带了空格和"和回车
|
272 |
+
opt_root = opt_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
273 |
+
os.makedirs(opt_root, exist_ok=True)
|
274 |
+
try:
|
275 |
+
if dir_path != "":
|
276 |
+
paths = [os.path.join(dir_path, name) for name in os.listdir(dir_path)]
|
277 |
+
else:
|
278 |
+
paths = [path.name for path in paths]
|
279 |
+
except:
|
280 |
+
traceback.print_exc()
|
281 |
+
paths = [path.name for path in paths]
|
282 |
+
infos = []
|
283 |
+
for path in paths:
|
284 |
+
info, opt = vc_single(
|
285 |
+
sid,
|
286 |
+
path,
|
287 |
+
f0_up_key,
|
288 |
+
None,
|
289 |
+
f0_method,
|
290 |
+
file_index,
|
291 |
+
file_index2,
|
292 |
+
# file_big_npy,
|
293 |
+
index_rate,
|
294 |
+
filter_radius,
|
295 |
+
resample_sr,
|
296 |
+
rms_mix_rate,
|
297 |
+
protect,
|
298 |
+
)
|
299 |
+
if "Success" in info:
|
300 |
+
try:
|
301 |
+
tgt_sr, audio_opt = opt
|
302 |
+
if format1 in ["wav", "flac"]:
|
303 |
+
sf.write(
|
304 |
+
"%s/%s.%s" % (opt_root, os.path.basename(path), format1),
|
305 |
+
audio_opt,
|
306 |
+
tgt_sr,
|
307 |
+
)
|
308 |
+
else:
|
309 |
+
path = "%s/%s.wav" % (opt_root, os.path.basename(path))
|
310 |
+
sf.write(
|
311 |
+
path,
|
312 |
+
audio_opt,
|
313 |
+
tgt_sr,
|
314 |
+
)
|
315 |
+
if os.path.exists(path):
|
316 |
+
os.system(
|
317 |
+
"ffmpeg -i %s -vn %s -q:a 2 -y"
|
318 |
+
% (path, path[:-4] + ".%s" % format1)
|
319 |
+
)
|
320 |
+
except:
|
321 |
+
info += traceback.format_exc()
|
322 |
+
infos.append("%s->%s" % (os.path.basename(path), info))
|
323 |
+
yield "\n".join(infos)
|
324 |
+
yield "\n".join(infos)
|
325 |
+
except:
|
326 |
+
yield traceback.format_exc()
|
327 |
+
|
328 |
+
|
329 |
+
def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0):
|
330 |
+
infos = []
|
331 |
+
try:
|
332 |
+
inp_root = inp_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
333 |
+
save_root_vocal = (
|
334 |
+
save_root_vocal.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
335 |
+
)
|
336 |
+
save_root_ins = (
|
337 |
+
save_root_ins.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
|
338 |
+
)
|
339 |
+
if model_name == "onnx_dereverb_By_FoxJoy":
|
340 |
+
pre_fun = MDXNetDereverb(15)
|
341 |
+
else:
|
342 |
+
func = _audio_pre_ if "DeEcho" not in model_name else _audio_pre_new
|
343 |
+
pre_fun = func(
|
344 |
+
agg=int(agg),
|
345 |
+
model_path=os.path.join(weight_uvr5_root, model_name + ".pth"),
|
346 |
+
device=config.device,
|
347 |
+
is_half=config.is_half,
|
348 |
+
)
|
349 |
+
if inp_root != "":
|
350 |
+
paths = [os.path.join(inp_root, name) for name in os.listdir(inp_root)]
|
351 |
+
else:
|
352 |
+
paths = [path.name for path in paths]
|
353 |
+
for path in paths:
|
354 |
+
inp_path = os.path.join(inp_root, path)
|
355 |
+
need_reformat = 1
|
356 |
+
done = 0
|
357 |
+
try:
|
358 |
+
info = ffmpeg.probe(inp_path, cmd="ffprobe")
|
359 |
+
if (
|
360 |
+
info["streams"][0]["channels"] == 2
|
361 |
+
and info["streams"][0]["sample_rate"] == "44100"
|
362 |
+
):
|
363 |
+
need_reformat = 0
|
364 |
+
pre_fun._path_audio_(
|
365 |
+
inp_path, save_root_ins, save_root_vocal, format0
|
366 |
+
)
|
367 |
+
done = 1
|
368 |
+
except:
|
369 |
+
need_reformat = 1
|
370 |
+
traceback.print_exc()
|
371 |
+
if need_reformat == 1:
|
372 |
+
tmp_path = "%s/%s.reformatted.wav" % (tmp, os.path.basename(inp_path))
|
373 |
+
os.system(
|
374 |
+
"ffmpeg -i %s -vn -acodec pcm_s16le -ac 2 -ar 44100 %s -y"
|
375 |
+
% (inp_path, tmp_path)
|
376 |
+
)
|
377 |
+
inp_path = tmp_path
|
378 |
+
try:
|
379 |
+
if done == 0:
|
380 |
+
pre_fun._path_audio_(
|
381 |
+
inp_path, save_root_ins, save_root_vocal, format0
|
382 |
+
)
|
383 |
+
infos.append("%s->Success" % (os.path.basename(inp_path)))
|
384 |
+
yield "\n".join(infos)
|
385 |
+
except:
|
386 |
+
infos.append(
|
387 |
+
"%s->%s" % (os.path.basename(inp_path), traceback.format_exc())
|
388 |
+
)
|
389 |
+
yield "\n".join(infos)
|
390 |
+
except:
|
391 |
+
infos.append(traceback.format_exc())
|
392 |
+
yield "\n".join(infos)
|
393 |
+
finally:
|
394 |
+
try:
|
395 |
+
if model_name == "onnx_dereverb_By_FoxJoy":
|
396 |
+
del pre_fun.pred.model
|
397 |
+
del pre_fun.pred.model_
|
398 |
+
else:
|
399 |
+
del pre_fun.model
|
400 |
+
del pre_fun
|
401 |
+
except:
|
402 |
+
traceback.print_exc()
|
403 |
+
print("clean_empty_cache")
|
404 |
+
if torch.cuda.is_available():
|
405 |
+
torch.cuda.empty_cache()
|
406 |
+
yield "\n".join(infos)
|
407 |
+
|
408 |
+
|
409 |
+
# 一个选项卡全局只能有一个音色
|
410 |
+
def get_vc(sid, to_return_protect0, to_return_protect1):
|
411 |
+
global n_spk, tgt_sr, net_g, vc, cpt, version
|
412 |
+
if sid == "" or sid == []:
|
413 |
+
global hubert_model
|
414 |
+
if hubert_model is not None: # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
|
415 |
+
print("clean_empty_cache")
|
416 |
+
del net_g, n_spk, vc, hubert_model, tgt_sr # ,cpt
|
417 |
+
hubert_model = net_g = n_spk = vc = hubert_model = tgt_sr = None
|
418 |
+
if torch.cuda.is_available():
|
419 |
+
torch.cuda.empty_cache()
|
420 |
+
###楼下不这么折腾清理不干净
|
421 |
+
if_f0 = cpt.get("f0", 1)
|
422 |
+
version = cpt.get("version", "v1")
|
423 |
+
if version == "v1":
|
424 |
+
if if_f0 == 1:
|
425 |
+
net_g = SynthesizerTrnMs256NSFsid(
|
426 |
+
*cpt["config"], is_half=config.is_half
|
427 |
+
)
|
428 |
+
else:
|
429 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
430 |
+
elif version == "v2":
|
431 |
+
if if_f0 == 1:
|
432 |
+
net_g = SynthesizerTrnMs768NSFsid(
|
433 |
+
*cpt["config"], is_half=config.is_half
|
434 |
+
)
|
435 |
+
else:
|
436 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
437 |
+
del net_g, cpt
|
438 |
+
if torch.cuda.is_available():
|
439 |
+
torch.cuda.empty_cache()
|
440 |
+
cpt = None
|
441 |
+
return {"visible": False, "__type__": "update"}
|
442 |
+
person = "%s/%s" % (weight_root, sid)
|
443 |
+
print("loading %s" % person)
|
444 |
+
cpt = torch.load(person, map_location="cpu")
|
445 |
+
tgt_sr = cpt["config"][-1]
|
446 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0] # n_spk
|
447 |
+
if_f0 = cpt.get("f0", 1)
|
448 |
+
if if_f0 == 0:
|
449 |
+
to_return_protect0 = to_return_protect1 = {
|
450 |
+
"visible": False,
|
451 |
+
"value": 0.5,
|
452 |
+
"__type__": "update",
|
453 |
+
}
|
454 |
+
else:
|
455 |
+
to_return_protect0 = {
|
456 |
+
"visible": True,
|
457 |
+
"value": to_return_protect0,
|
458 |
+
"__type__": "update",
|
459 |
+
}
|
460 |
+
to_return_protect1 = {
|
461 |
+
"visible": True,
|
462 |
+
"value": to_return_protect1,
|
463 |
+
"__type__": "update",
|
464 |
+
}
|
465 |
+
version = cpt.get("version", "v1")
|
466 |
+
if version == "v1":
|
467 |
+
if if_f0 == 1:
|
468 |
+
net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
|
469 |
+
else:
|
470 |
+
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
471 |
+
elif version == "v2":
|
472 |
+
if if_f0 == 1:
|
473 |
+
net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
|
474 |
+
else:
|
475 |
+
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
476 |
+
del net_g.enc_q
|
477 |
+
print(net_g.load_state_dict(cpt["weight"], strict=False))
|
478 |
+
net_g.eval().to(config.device)
|
479 |
+
if config.is_half:
|
480 |
+
net_g = net_g.half()
|
481 |
+
else:
|
482 |
+
net_g = net_g.float()
|
483 |
+
vc = VC(tgt_sr, config)
|
484 |
+
n_spk = cpt["config"][-3]
|
485 |
+
return (
|
486 |
+
{"visible": True, "maximum": n_spk, "__type__": "update"},
|
487 |
+
to_return_protect0,
|
488 |
+
to_return_protect1,
|
489 |
+
)
|
490 |
+
|
491 |
+
|
492 |
+
def change_choices():
|
493 |
+
names = []
|
494 |
+
for name in os.listdir(weight_root):
|
495 |
+
if name.endswith(".pth"):
|
496 |
+
names.append(name)
|
497 |
+
index_paths = []
|
498 |
+
for root, dirs, files in os.walk(index_root, topdown=False):
|
499 |
+
for name in files:
|
500 |
+
if name.endswith(".index") and "trained" not in name:
|
501 |
+
index_paths.append("%s/%s" % (root, name))
|
502 |
+
return {"choices": sorted(names), "__type__": "update"}, {
|
503 |
+
"choices": sorted(index_paths),
|
504 |
+
"__type__": "update",
|
505 |
+
}
|
506 |
+
|
507 |
+
|
508 |
+
def clean():
|
509 |
+
return {"value": "", "__type__": "update"}
|
510 |
+
|
511 |
+
|
512 |
+
sr_dict = {
|
513 |
+
"32k": 32000,
|
514 |
+
"40k": 40000,
|
515 |
+
"48k": 48000,
|
516 |
+
}
|
517 |
+
|
518 |
+
|
519 |
+
def if_done(done, p):
|
520 |
+
while 1:
|
521 |
+
if p.poll() is None:
|
522 |
+
sleep(0.5)
|
523 |
+
else:
|
524 |
+
break
|
525 |
+
done[0] = True
|
526 |
+
|
527 |
+
|
528 |
+
def if_done_multi(done, ps):
|
529 |
+
while 1:
|
530 |
+
# poll==None代表进程未结束
|
531 |
+
# 只要有一个进程未结束都不停
|
532 |
+
flag = 1
|
533 |
+
for p in ps:
|
534 |
+
if p.poll() is None:
|
535 |
+
flag = 0
|
536 |
+
sleep(0.5)
|
537 |
+
break
|
538 |
+
if flag == 1:
|
539 |
+
break
|
540 |
+
done[0] = True
|
541 |
+
|
542 |
+
|
543 |
+
def preprocess_dataset(trainset_dir, exp_dir, sr, n_p):
|
544 |
+
sr = sr_dict[sr]
|
545 |
+
os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
|
546 |
+
f = open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "w")
|
547 |
+
f.close()
|
548 |
+
cmd = (
|
549 |
+
config.python_cmd
|
550 |
+
+ " trainset_preprocess_pipeline_print.py %s %s %s %s/logs/%s "
|
551 |
+
% (trainset_dir, sr, n_p, now_dir, exp_dir)
|
552 |
+
+ str(config.noparallel)
|
553 |
+
)
|
554 |
+
print(cmd)
|
555 |
+
p = Popen(cmd, shell=True) # , stdin=PIPE, stdout=PIPE,stderr=PIPE,cwd=now_dir
|
556 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
557 |
+
done = [False]
|
558 |
+
threading.Thread(
|
559 |
+
target=if_done,
|
560 |
+
args=(
|
561 |
+
done,
|
562 |
+
p,
|
563 |
+
),
|
564 |
+
).start()
|
565 |
+
while 1:
|
566 |
+
with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
|
567 |
+
yield (f.read())
|
568 |
+
sleep(1)
|
569 |
+
if done[0]:
|
570 |
+
break
|
571 |
+
with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
|
572 |
+
log = f.read()
|
573 |
+
print(log)
|
574 |
+
yield log
|
575 |
+
|
576 |
+
|
577 |
+
# but2.click(extract_f0,[gpus6,np7,f0method8,if_f0_3,trainset_dir4],[info2])
|
578 |
+
def extract_f0_feature(gpus, n_p, f0method, if_f0, exp_dir, version19):
|
579 |
+
gpus = gpus.split("-")
|
580 |
+
os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
|
581 |
+
f = open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "w")
|
582 |
+
f.close()
|
583 |
+
if if_f0:
|
584 |
+
cmd = config.python_cmd + " extract_f0_print.py %s/logs/%s %s %s" % (
|
585 |
+
now_dir,
|
586 |
+
exp_dir,
|
587 |
+
n_p,
|
588 |
+
f0method,
|
589 |
+
)
|
590 |
+
print(cmd)
|
591 |
+
p = Popen(cmd, shell=True, cwd=now_dir) # , stdin=PIPE, stdout=PIPE,stderr=PIPE
|
592 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
593 |
+
done = [False]
|
594 |
+
threading.Thread(
|
595 |
+
target=if_done,
|
596 |
+
args=(
|
597 |
+
done,
|
598 |
+
p,
|
599 |
+
),
|
600 |
+
).start()
|
601 |
+
while 1:
|
602 |
+
with open(
|
603 |
+
"%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r"
|
604 |
+
) as f:
|
605 |
+
yield (f.read())
|
606 |
+
sleep(1)
|
607 |
+
if done[0]:
|
608 |
+
break
|
609 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
610 |
+
log = f.read()
|
611 |
+
print(log)
|
612 |
+
yield log
|
613 |
+
####对不同part分别开多进程
|
614 |
+
"""
|
615 |
+
n_part=int(sys.argv[1])
|
616 |
+
i_part=int(sys.argv[2])
|
617 |
+
i_gpu=sys.argv[3]
|
618 |
+
exp_dir=sys.argv[4]
|
619 |
+
os.environ["CUDA_VISIBLE_DEVICES"]=str(i_gpu)
|
620 |
+
"""
|
621 |
+
leng = len(gpus)
|
622 |
+
ps = []
|
623 |
+
for idx, n_g in enumerate(gpus):
|
624 |
+
cmd = (
|
625 |
+
config.python_cmd
|
626 |
+
+ " extract_feature_print.py %s %s %s %s %s/logs/%s %s"
|
627 |
+
% (
|
628 |
+
config.device,
|
629 |
+
leng,
|
630 |
+
idx,
|
631 |
+
n_g,
|
632 |
+
now_dir,
|
633 |
+
exp_dir,
|
634 |
+
version19,
|
635 |
+
)
|
636 |
+
)
|
637 |
+
print(cmd)
|
638 |
+
p = Popen(
|
639 |
+
cmd, shell=True, cwd=now_dir
|
640 |
+
) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
|
641 |
+
ps.append(p)
|
642 |
+
###煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
|
643 |
+
done = [False]
|
644 |
+
threading.Thread(
|
645 |
+
target=if_done_multi,
|
646 |
+
args=(
|
647 |
+
done,
|
648 |
+
ps,
|
649 |
+
),
|
650 |
+
).start()
|
651 |
+
while 1:
|
652 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
653 |
+
yield (f.read())
|
654 |
+
sleep(1)
|
655 |
+
if done[0]:
|
656 |
+
break
|
657 |
+
with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
|
658 |
+
log = f.read()
|
659 |
+
print(log)
|
660 |
+
yield log
|
661 |
+
|
662 |
+
|
663 |
+
def change_sr2(sr2, if_f0_3, version19):
|
664 |
+
path_str = "" if version19 == "v1" else "_v2"
|
665 |
+
f0_str = "f0" if if_f0_3 else ""
|
666 |
+
if_pretrained_generator_exist = os.access(
|
667 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
668 |
+
)
|
669 |
+
if_pretrained_discriminator_exist = os.access(
|
670 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
671 |
+
)
|
672 |
+
if not if_pretrained_generator_exist:
|
673 |
+
print(
|
674 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2),
|
675 |
+
"not exist, will not use pretrained model",
|
676 |
+
)
|
677 |
+
if not if_pretrained_discriminator_exist:
|
678 |
+
print(
|
679 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2),
|
680 |
+
"not exist, will not use pretrained model",
|
681 |
+
)
|
682 |
+
return (
|
683 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2)
|
684 |
+
if if_pretrained_generator_exist
|
685 |
+
else "",
|
686 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2)
|
687 |
+
if if_pretrained_discriminator_exist
|
688 |
+
else "",
|
689 |
+
)
|
690 |
+
|
691 |
+
|
692 |
+
def change_version19(sr2, if_f0_3, version19):
|
693 |
+
path_str = "" if version19 == "v1" else "_v2"
|
694 |
+
if sr2 == "32k" and version19 == "v1":
|
695 |
+
sr2 = "40k"
|
696 |
+
to_return_sr2 = (
|
697 |
+
{"choices": ["40k", "48k"], "__type__": "update", "value": sr2}
|
698 |
+
if version19 == "v1"
|
699 |
+
else {"choices": ["40k", "48k", "32k"], "__type__": "update", "value": sr2}
|
700 |
+
)
|
701 |
+
f0_str = "f0" if if_f0_3 else ""
|
702 |
+
if_pretrained_generator_exist = os.access(
|
703 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
704 |
+
)
|
705 |
+
if_pretrained_discriminator_exist = os.access(
|
706 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2), os.F_OK
|
707 |
+
)
|
708 |
+
if not if_pretrained_generator_exist:
|
709 |
+
print(
|
710 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2),
|
711 |
+
"not exist, will not use pretrained model",
|
712 |
+
)
|
713 |
+
if not if_pretrained_discriminator_exist:
|
714 |
+
print(
|
715 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2),
|
716 |
+
"not exist, will not use pretrained model",
|
717 |
+
)
|
718 |
+
return (
|
719 |
+
"pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2)
|
720 |
+
if if_pretrained_generator_exist
|
721 |
+
else "",
|
722 |
+
"pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2)
|
723 |
+
if if_pretrained_discriminator_exist
|
724 |
+
else "",
|
725 |
+
to_return_sr2,
|
726 |
+
)
|
727 |
+
|
728 |
+
|
729 |
+
def change_f0(if_f0_3, sr2, version19): # f0method8,pretrained_G14,pretrained_D15
|
730 |
+
path_str = "" if version19 == "v1" else "_v2"
|
731 |
+
if_pretrained_generator_exist = os.access(
|
732 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2), os.F_OK
|
733 |
+
)
|
734 |
+
if_pretrained_discriminator_exist = os.access(
|
735 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2), os.F_OK
|
736 |
+
)
|
737 |
+
if not if_pretrained_generator_exist:
|
738 |
+
print(
|
739 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2),
|
740 |
+
"not exist, will not use pretrained model",
|
741 |
+
)
|
742 |
+
if not if_pretrained_discriminator_exist:
|
743 |
+
print(
|
744 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2),
|
745 |
+
"not exist, will not use pretrained model",
|
746 |
+
)
|
747 |
+
if if_f0_3:
|
748 |
+
return (
|
749 |
+
{"visible": True, "__type__": "update"},
|
750 |
+
"pretrained%s/f0G%s.pth" % (path_str, sr2)
|
751 |
+
if if_pretrained_generator_exist
|
752 |
+
else "",
|
753 |
+
"pretrained%s/f0D%s.pth" % (path_str, sr2)
|
754 |
+
if if_pretrained_discriminator_exist
|
755 |
+
else "",
|
756 |
+
)
|
757 |
+
return (
|
758 |
+
{"visible": False, "__type__": "update"},
|
759 |
+
("pretrained%s/G%s.pth" % (path_str, sr2))
|
760 |
+
if if_pretrained_generator_exist
|
761 |
+
else "",
|
762 |
+
("pretrained%s/D%s.pth" % (path_str, sr2))
|
763 |
+
if if_pretrained_discriminator_exist
|
764 |
+
else "",
|
765 |
+
)
|
766 |
+
|
767 |
+
|
768 |
+
# but3.click(click_train,[exp_dir1,sr2,if_f0_3,save_epoch10,total_epoch11,batch_size12,if_save_latest13,pretrained_G14,pretrained_D15,gpus16])
|
769 |
+
def click_train(
|
770 |
+
exp_dir1,
|
771 |
+
sr2,
|
772 |
+
if_f0_3,
|
773 |
+
spk_id5,
|
774 |
+
save_epoch10,
|
775 |
+
total_epoch11,
|
776 |
+
batch_size12,
|
777 |
+
if_save_latest13,
|
778 |
+
pretrained_G14,
|
779 |
+
pretrained_D15,
|
780 |
+
gpus16,
|
781 |
+
if_cache_gpu17,
|
782 |
+
if_save_every_weights18,
|
783 |
+
version19,
|
784 |
+
):
|
785 |
+
# 生成filelist
|
786 |
+
exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
787 |
+
os.makedirs(exp_dir, exist_ok=True)
|
788 |
+
gt_wavs_dir = "%s/0_gt_wavs" % (exp_dir)
|
789 |
+
feature_dir = (
|
790 |
+
"%s/3_feature256" % (exp_dir)
|
791 |
+
if version19 == "v1"
|
792 |
+
else "%s/3_feature768" % (exp_dir)
|
793 |
+
)
|
794 |
+
if if_f0_3:
|
795 |
+
f0_dir = "%s/2a_f0" % (exp_dir)
|
796 |
+
f0nsf_dir = "%s/2b-f0nsf" % (exp_dir)
|
797 |
+
names = (
|
798 |
+
set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)])
|
799 |
+
& set([name.split(".")[0] for name in os.listdir(feature_dir)])
|
800 |
+
& set([name.split(".")[0] for name in os.listdir(f0_dir)])
|
801 |
+
& set([name.split(".")[0] for name in os.listdir(f0nsf_dir)])
|
802 |
+
)
|
803 |
+
else:
|
804 |
+
names = set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)]) & set(
|
805 |
+
[name.split(".")[0] for name in os.listdir(feature_dir)]
|
806 |
+
)
|
807 |
+
opt = []
|
808 |
+
for name in names:
|
809 |
+
if if_f0_3:
|
810 |
+
opt.append(
|
811 |
+
"%s/%s.wav|%s/%s.npy|%s/%s.wav.npy|%s/%s.wav.npy|%s"
|
812 |
+
% (
|
813 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
814 |
+
name,
|
815 |
+
feature_dir.replace("\\", "\\\\"),
|
816 |
+
name,
|
817 |
+
f0_dir.replace("\\", "\\\\"),
|
818 |
+
name,
|
819 |
+
f0nsf_dir.replace("\\", "\\\\"),
|
820 |
+
name,
|
821 |
+
spk_id5,
|
822 |
+
)
|
823 |
+
)
|
824 |
+
else:
|
825 |
+
opt.append(
|
826 |
+
"%s/%s.wav|%s/%s.npy|%s"
|
827 |
+
% (
|
828 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
829 |
+
name,
|
830 |
+
feature_dir.replace("\\", "\\\\"),
|
831 |
+
name,
|
832 |
+
spk_id5,
|
833 |
+
)
|
834 |
+
)
|
835 |
+
fea_dim = 256 if version19 == "v1" else 768
|
836 |
+
if if_f0_3:
|
837 |
+
for _ in range(2):
|
838 |
+
opt.append(
|
839 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s/logs/mute/2a_f0/mute.wav.npy|%s/logs/mute/2b-f0nsf/mute.wav.npy|%s"
|
840 |
+
% (now_dir, sr2, now_dir, fea_dim, now_dir, now_dir, spk_id5)
|
841 |
+
)
|
842 |
+
else:
|
843 |
+
for _ in range(2):
|
844 |
+
opt.append(
|
845 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s"
|
846 |
+
% (now_dir, sr2, now_dir, fea_dim, spk_id5)
|
847 |
+
)
|
848 |
+
shuffle(opt)
|
849 |
+
with open("%s/filelist.txt" % exp_dir, "w") as f:
|
850 |
+
f.write("\n".join(opt))
|
851 |
+
print("write filelist done")
|
852 |
+
# 生成config#无需生成config
|
853 |
+
# cmd = python_cmd + " train_nsf_sim_cache_sid_load_pretrain.py -e mi-test -sr 40k -f0 1 -bs 4 -g 0 -te 10 -se 5 -pg pretrained/f0G40k.pth -pd pretrained/f0D40k.pth -l 1 -c 0"
|
854 |
+
print("use gpus:", gpus16)
|
855 |
+
if pretrained_G14 == "":
|
856 |
+
print("no pretrained Generator")
|
857 |
+
if pretrained_D15 == "":
|
858 |
+
print("no pretrained Discriminator")
|
859 |
+
if gpus16:
|
860 |
+
cmd = (
|
861 |
+
config.python_cmd
|
862 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -g %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
863 |
+
% (
|
864 |
+
exp_dir1,
|
865 |
+
sr2,
|
866 |
+
1 if if_f0_3 else 0,
|
867 |
+
batch_size12,
|
868 |
+
gpus16,
|
869 |
+
total_epoch11,
|
870 |
+
save_epoch10,
|
871 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
872 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
873 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
874 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
875 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
876 |
+
version19,
|
877 |
+
)
|
878 |
+
)
|
879 |
+
else:
|
880 |
+
cmd = (
|
881 |
+
config.python_cmd
|
882 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
883 |
+
% (
|
884 |
+
exp_dir1,
|
885 |
+
sr2,
|
886 |
+
1 if if_f0_3 else 0,
|
887 |
+
batch_size12,
|
888 |
+
total_epoch11,
|
889 |
+
save_epoch10,
|
890 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "\b",
|
891 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "\b",
|
892 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
893 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
894 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
895 |
+
version19,
|
896 |
+
)
|
897 |
+
)
|
898 |
+
print(cmd)
|
899 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
900 |
+
p.wait()
|
901 |
+
return "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log"
|
902 |
+
|
903 |
+
|
904 |
+
# but4.click(train_index, [exp_dir1], info3)
|
905 |
+
def train_index(exp_dir1, version19):
|
906 |
+
exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
907 |
+
os.makedirs(exp_dir, exist_ok=True)
|
908 |
+
feature_dir = (
|
909 |
+
"%s/3_feature256" % (exp_dir)
|
910 |
+
if version19 == "v1"
|
911 |
+
else "%s/3_feature768" % (exp_dir)
|
912 |
+
)
|
913 |
+
if not os.path.exists(feature_dir):
|
914 |
+
return "请先进行特征提取!"
|
915 |
+
listdir_res = list(os.listdir(feature_dir))
|
916 |
+
if len(listdir_res) == 0:
|
917 |
+
return "请先进行特征提取!"
|
918 |
+
infos = []
|
919 |
+
npys = []
|
920 |
+
for name in sorted(listdir_res):
|
921 |
+
phone = np.load("%s/%s" % (feature_dir, name))
|
922 |
+
npys.append(phone)
|
923 |
+
big_npy = np.concatenate(npys, 0)
|
924 |
+
big_npy_idx = np.arange(big_npy.shape[0])
|
925 |
+
np.random.shuffle(big_npy_idx)
|
926 |
+
big_npy = big_npy[big_npy_idx]
|
927 |
+
if big_npy.shape[0] > 2e5:
|
928 |
+
# if(1):
|
929 |
+
infos.append("Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0])
|
930 |
+
yield "\n".join(infos)
|
931 |
+
try:
|
932 |
+
big_npy = (
|
933 |
+
MiniBatchKMeans(
|
934 |
+
n_clusters=10000,
|
935 |
+
verbose=True,
|
936 |
+
batch_size=256 * config.n_cpu,
|
937 |
+
compute_labels=False,
|
938 |
+
init="random",
|
939 |
+
)
|
940 |
+
.fit(big_npy)
|
941 |
+
.cluster_centers_
|
942 |
+
)
|
943 |
+
except:
|
944 |
+
info = traceback.format_exc()
|
945 |
+
print(info)
|
946 |
+
infos.append(info)
|
947 |
+
yield "\n".join(infos)
|
948 |
+
|
949 |
+
np.save("%s/total_fea.npy" % exp_dir, big_npy)
|
950 |
+
n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
|
951 |
+
infos.append("%s,%s" % (big_npy.shape, n_ivf))
|
952 |
+
yield "\n".join(infos)
|
953 |
+
index = faiss.index_factory(256 if version19 == "v1" else 768, "IVF%s,Flat" % n_ivf)
|
954 |
+
# index = faiss.index_factory(256if version19=="v1"else 768, "IVF%s,PQ128x4fs,RFlat"%n_ivf)
|
955 |
+
infos.append("training")
|
956 |
+
yield "\n".join(infos)
|
957 |
+
index_ivf = faiss.extract_index_ivf(index) #
|
958 |
+
index_ivf.nprobe = 1
|
959 |
+
index.train(big_npy)
|
960 |
+
faiss.write_index(
|
961 |
+
index,
|
962 |
+
"%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
963 |
+
% (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
964 |
+
)
|
965 |
+
# faiss.write_index(index, '%s/trained_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19))
|
966 |
+
infos.append("adding")
|
967 |
+
yield "\n".join(infos)
|
968 |
+
batch_size_add = 8192
|
969 |
+
for i in range(0, big_npy.shape[0], batch_size_add):
|
970 |
+
index.add(big_npy[i : i + batch_size_add])
|
971 |
+
faiss.write_index(
|
972 |
+
index,
|
973 |
+
"%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
974 |
+
% (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
975 |
+
)
|
976 |
+
infos.append(
|
977 |
+
"成功构建索引,added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
978 |
+
% (n_ivf, index_ivf.nprobe, exp_dir1, version19)
|
979 |
+
)
|
980 |
+
# faiss.write_index(index, '%s/added_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19))
|
981 |
+
# infos.append("成功构建索引,added_IVF%s_Flat_FastScan_%s.index"%(n_ivf,version19))
|
982 |
+
yield "\n".join(infos)
|
983 |
+
|
984 |
+
|
985 |
+
# but5.click(train1key, [exp_dir1, sr2, if_f0_3, trainset_dir4, spk_id5, gpus6, np7, f0method8, save_epoch10, total_epoch11, batch_size12, if_save_latest13, pretrained_G14, pretrained_D15, gpus16, if_cache_gpu17], info3)
|
986 |
+
def train1key(
|
987 |
+
exp_dir1,
|
988 |
+
sr2,
|
989 |
+
if_f0_3,
|
990 |
+
trainset_dir4,
|
991 |
+
spk_id5,
|
992 |
+
np7,
|
993 |
+
f0method8,
|
994 |
+
save_epoch10,
|
995 |
+
total_epoch11,
|
996 |
+
batch_size12,
|
997 |
+
if_save_latest13,
|
998 |
+
pretrained_G14,
|
999 |
+
pretrained_D15,
|
1000 |
+
gpus16,
|
1001 |
+
if_cache_gpu17,
|
1002 |
+
if_save_every_weights18,
|
1003 |
+
version19,
|
1004 |
+
):
|
1005 |
+
infos = []
|
1006 |
+
|
1007 |
+
def get_info_str(strr):
|
1008 |
+
infos.append(strr)
|
1009 |
+
return "\n".join(infos)
|
1010 |
+
|
1011 |
+
model_log_dir = "%s/logs/%s" % (now_dir, exp_dir1)
|
1012 |
+
preprocess_log_path = "%s/preprocess.log" % model_log_dir
|
1013 |
+
extract_f0_feature_log_path = "%s/extract_f0_feature.log" % model_log_dir
|
1014 |
+
gt_wavs_dir = "%s/0_gt_wavs" % model_log_dir
|
1015 |
+
feature_dir = (
|
1016 |
+
"%s/3_feature256" % model_log_dir
|
1017 |
+
if version19 == "v1"
|
1018 |
+
else "%s/3_feature768" % model_log_dir
|
1019 |
+
)
|
1020 |
+
|
1021 |
+
os.makedirs(model_log_dir, exist_ok=True)
|
1022 |
+
#########step1:处理数据
|
1023 |
+
open(preprocess_log_path, "w").close()
|
1024 |
+
cmd = (
|
1025 |
+
config.python_cmd
|
1026 |
+
+ " trainset_preprocess_pipeline_print.py %s %s %s %s "
|
1027 |
+
% (trainset_dir4, sr_dict[sr2], np7, model_log_dir)
|
1028 |
+
+ str(config.noparallel)
|
1029 |
+
)
|
1030 |
+
yield get_info_str(i18n("step1:正在处理数据"))
|
1031 |
+
yield get_info_str(cmd)
|
1032 |
+
p = Popen(cmd, shell=True)
|
1033 |
+
p.wait()
|
1034 |
+
with open(preprocess_log_path, "r") as f:
|
1035 |
+
print(f.read())
|
1036 |
+
#########step2a:提取音高
|
1037 |
+
open(extract_f0_feature_log_path, "w")
|
1038 |
+
if if_f0_3:
|
1039 |
+
yield get_info_str("step2a:正在提取音高")
|
1040 |
+
cmd = config.python_cmd + " extract_f0_print.py %s %s %s" % (
|
1041 |
+
model_log_dir,
|
1042 |
+
np7,
|
1043 |
+
f0method8,
|
1044 |
+
)
|
1045 |
+
yield get_info_str(cmd)
|
1046 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
1047 |
+
p.wait()
|
1048 |
+
with open(extract_f0_feature_log_path, "r") as f:
|
1049 |
+
print(f.read())
|
1050 |
+
else:
|
1051 |
+
yield get_info_str(i18n("step2a:无需提取音高"))
|
1052 |
+
#######step2b:提取特征
|
1053 |
+
yield get_info_str(i18n("step2b:正在提取特征"))
|
1054 |
+
gpus = gpus16.split("-")
|
1055 |
+
leng = len(gpus)
|
1056 |
+
ps = []
|
1057 |
+
for idx, n_g in enumerate(gpus):
|
1058 |
+
cmd = config.python_cmd + " extract_feature_print.py %s %s %s %s %s %s" % (
|
1059 |
+
config.device,
|
1060 |
+
leng,
|
1061 |
+
idx,
|
1062 |
+
n_g,
|
1063 |
+
model_log_dir,
|
1064 |
+
version19,
|
1065 |
+
)
|
1066 |
+
yield get_info_str(cmd)
|
1067 |
+
p = Popen(
|
1068 |
+
cmd, shell=True, cwd=now_dir
|
1069 |
+
) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
|
1070 |
+
ps.append(p)
|
1071 |
+
for p in ps:
|
1072 |
+
p.wait()
|
1073 |
+
with open(extract_f0_feature_log_path, "r") as f:
|
1074 |
+
print(f.read())
|
1075 |
+
#######step3a:训练模型
|
1076 |
+
yield get_info_str(i18n("step3a:正在训练模型"))
|
1077 |
+
# 生成filelist
|
1078 |
+
if if_f0_3:
|
1079 |
+
f0_dir = "%s/2a_f0" % model_log_dir
|
1080 |
+
f0nsf_dir = "%s/2b-f0nsf" % model_log_dir
|
1081 |
+
names = (
|
1082 |
+
set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)])
|
1083 |
+
& set([name.split(".")[0] for name in os.listdir(feature_dir)])
|
1084 |
+
& set([name.split(".")[0] for name in os.listdir(f0_dir)])
|
1085 |
+
& set([name.split(".")[0] for name in os.listdir(f0nsf_dir)])
|
1086 |
+
)
|
1087 |
+
else:
|
1088 |
+
names = set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)]) & set(
|
1089 |
+
[name.split(".")[0] for name in os.listdir(feature_dir)]
|
1090 |
+
)
|
1091 |
+
opt = []
|
1092 |
+
for name in names:
|
1093 |
+
if if_f0_3:
|
1094 |
+
opt.append(
|
1095 |
+
"%s/%s.wav|%s/%s.npy|%s/%s.wav.npy|%s/%s.wav.npy|%s"
|
1096 |
+
% (
|
1097 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
1098 |
+
name,
|
1099 |
+
feature_dir.replace("\\", "\\\\"),
|
1100 |
+
name,
|
1101 |
+
f0_dir.replace("\\", "\\\\"),
|
1102 |
+
name,
|
1103 |
+
f0nsf_dir.replace("\\", "\\\\"),
|
1104 |
+
name,
|
1105 |
+
spk_id5,
|
1106 |
+
)
|
1107 |
+
)
|
1108 |
+
else:
|
1109 |
+
opt.append(
|
1110 |
+
"%s/%s.wav|%s/%s.npy|%s"
|
1111 |
+
% (
|
1112 |
+
gt_wavs_dir.replace("\\", "\\\\"),
|
1113 |
+
name,
|
1114 |
+
feature_dir.replace("\\", "\\\\"),
|
1115 |
+
name,
|
1116 |
+
spk_id5,
|
1117 |
+
)
|
1118 |
+
)
|
1119 |
+
fea_dim = 256 if version19 == "v1" else 768
|
1120 |
+
if if_f0_3:
|
1121 |
+
for _ in range(2):
|
1122 |
+
opt.append(
|
1123 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s/logs/mute/2a_f0/mute.wav.npy|%s/logs/mute/2b-f0nsf/mute.wav.npy|%s"
|
1124 |
+
% (now_dir, sr2, now_dir, fea_dim, now_dir, now_dir, spk_id5)
|
1125 |
+
)
|
1126 |
+
else:
|
1127 |
+
for _ in range(2):
|
1128 |
+
opt.append(
|
1129 |
+
"%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s"
|
1130 |
+
% (now_dir, sr2, now_dir, fea_dim, spk_id5)
|
1131 |
+
)
|
1132 |
+
shuffle(opt)
|
1133 |
+
with open("%s/filelist.txt" % model_log_dir, "w") as f:
|
1134 |
+
f.write("\n".join(opt))
|
1135 |
+
yield get_info_str("write filelist done")
|
1136 |
+
if gpus16:
|
1137 |
+
cmd = (
|
1138 |
+
config.python_cmd
|
1139 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -g %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
1140 |
+
% (
|
1141 |
+
exp_dir1,
|
1142 |
+
sr2,
|
1143 |
+
1 if if_f0_3 else 0,
|
1144 |
+
batch_size12,
|
1145 |
+
gpus16,
|
1146 |
+
total_epoch11,
|
1147 |
+
save_epoch10,
|
1148 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
1149 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
1150 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
1151 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
1152 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
1153 |
+
version19,
|
1154 |
+
)
|
1155 |
+
)
|
1156 |
+
else:
|
1157 |
+
cmd = (
|
1158 |
+
config.python_cmd
|
1159 |
+
+ " train_nsf_sim_cache_sid_load_pretrain.py -e %s -sr %s -f0 %s -bs %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s"
|
1160 |
+
% (
|
1161 |
+
exp_dir1,
|
1162 |
+
sr2,
|
1163 |
+
1 if if_f0_3 else 0,
|
1164 |
+
batch_size12,
|
1165 |
+
total_epoch11,
|
1166 |
+
save_epoch10,
|
1167 |
+
"-pg %s" % pretrained_G14 if pretrained_G14 != "" else "",
|
1168 |
+
"-pd %s" % pretrained_D15 if pretrained_D15 != "" else "",
|
1169 |
+
1 if if_save_latest13 == i18n("是") else 0,
|
1170 |
+
1 if if_cache_gpu17 == i18n("是") else 0,
|
1171 |
+
1 if if_save_every_weights18 == i18n("是") else 0,
|
1172 |
+
version19,
|
1173 |
+
)
|
1174 |
+
)
|
1175 |
+
yield get_info_str(cmd)
|
1176 |
+
p = Popen(cmd, shell=True, cwd=now_dir)
|
1177 |
+
p.wait()
|
1178 |
+
yield get_info_str(i18n("训练结束, 您可查看控制台训练日志或实验文件夹下的train.log"))
|
1179 |
+
#######step3b:训练索引
|
1180 |
+
npys = []
|
1181 |
+
listdir_res = list(os.listdir(feature_dir))
|
1182 |
+
for name in sorted(listdir_res):
|
1183 |
+
phone = np.load("%s/%s" % (feature_dir, name))
|
1184 |
+
npys.append(phone)
|
1185 |
+
big_npy = np.concatenate(npys, 0)
|
1186 |
+
|
1187 |
+
big_npy_idx = np.arange(big_npy.shape[0])
|
1188 |
+
np.random.shuffle(big_npy_idx)
|
1189 |
+
big_npy = big_npy[big_npy_idx]
|
1190 |
+
|
1191 |
+
if big_npy.shape[0] > 2e5:
|
1192 |
+
# if(1):
|
1193 |
+
info = "Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0]
|
1194 |
+
print(info)
|
1195 |
+
yield get_info_str(info)
|
1196 |
+
try:
|
1197 |
+
big_npy = (
|
1198 |
+
MiniBatchKMeans(
|
1199 |
+
n_clusters=10000,
|
1200 |
+
verbose=True,
|
1201 |
+
batch_size=256 * config.n_cpu,
|
1202 |
+
compute_labels=False,
|
1203 |
+
init="random",
|
1204 |
+
)
|
1205 |
+
.fit(big_npy)
|
1206 |
+
.cluster_centers_
|
1207 |
+
)
|
1208 |
+
except:
|
1209 |
+
info = traceback.format_exc()
|
1210 |
+
print(info)
|
1211 |
+
yield get_info_str(info)
|
1212 |
+
|
1213 |
+
np.save("%s/total_fea.npy" % model_log_dir, big_npy)
|
1214 |
+
|
1215 |
+
# n_ivf = big_npy.shape[0] // 39
|
1216 |
+
n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
|
1217 |
+
yield get_info_str("%s,%s" % (big_npy.shape, n_ivf))
|
1218 |
+
index = faiss.index_factory(256 if version19 == "v1" else 768, "IVF%s,Flat" % n_ivf)
|
1219 |
+
yield get_info_str("training index")
|
1220 |
+
index_ivf = faiss.extract_index_ivf(index) #
|
1221 |
+
index_ivf.nprobe = 1
|
1222 |
+
index.train(big_npy)
|
1223 |
+
faiss.write_index(
|
1224 |
+
index,
|
1225 |
+
"%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1226 |
+
% (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
1227 |
+
)
|
1228 |
+
yield get_info_str("adding index")
|
1229 |
+
batch_size_add = 8192
|
1230 |
+
for i in range(0, big_npy.shape[0], batch_size_add):
|
1231 |
+
index.add(big_npy[i : i + batch_size_add])
|
1232 |
+
faiss.write_index(
|
1233 |
+
index,
|
1234 |
+
"%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1235 |
+
% (model_log_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
|
1236 |
+
)
|
1237 |
+
yield get_info_str(
|
1238 |
+
"成功构建索引, added_IVF%s_Flat_nprobe_%s_%s_%s.index"
|
1239 |
+
% (n_ivf, index_ivf.nprobe, exp_dir1, version19)
|
1240 |
+
)
|
1241 |
+
yield get_info_str(i18n("全流程结束!"))
|
1242 |
+
|
1243 |
+
|
1244 |
+
# ckpt_path2.change(change_info_,[ckpt_path2],[sr__,if_f0__])
|
1245 |
+
def change_info_(ckpt_path):
|
1246 |
+
if not os.path.exists(ckpt_path.replace(os.path.basename(ckpt_path), "train.log")):
|
1247 |
+
return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
|
1248 |
+
try:
|
1249 |
+
with open(
|
1250 |
+
ckpt_path.replace(os.path.basename(ckpt_path), "train.log"), "r"
|
1251 |
+
) as f:
|
1252 |
+
info = eval(f.read().strip("\n").split("\n")[0].split("\t")[-1])
|
1253 |
+
sr, f0 = info["sample_rate"], info["if_f0"]
|
1254 |
+
version = "v2" if ("version" in info and info["version"] == "v2") else "v1"
|
1255 |
+
return sr, str(f0), version
|
1256 |
+
except:
|
1257 |
+
traceback.print_exc()
|
1258 |
+
return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
|
1259 |
+
|
1260 |
+
|
1261 |
+
def export_onnx(ModelPath, ExportedPath):
|
1262 |
+
cpt = torch.load(ModelPath, map_location="cpu")
|
1263 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
|
1264 |
+
vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768
|
1265 |
+
|
1266 |
+
test_phone = torch.rand(1, 200, vec_channels) # hidden unit
|
1267 |
+
test_phone_lengths = torch.tensor([200]).long() # hidden unit 长度(貌似没啥用)
|
1268 |
+
test_pitch = torch.randint(size=(1, 200), low=5, high=255) # 基频(单位赫兹)
|
1269 |
+
test_pitchf = torch.rand(1, 200) # nsf基频
|
1270 |
+
test_ds = torch.LongTensor([0]) # 说话人ID
|
1271 |
+
test_rnd = torch.rand(1, 192, 200) # 噪声(加入随机因子)
|
1272 |
+
|
1273 |
+
device = "cpu" # 导出时设备(不影响使用模型)
|
1274 |
+
|
1275 |
+
net_g = SynthesizerTrnMsNSFsidM(
|
1276 |
+
*cpt["config"], is_half=False, version=cpt.get("version", "v1")
|
1277 |
+
) # fp32导出(C++要支持fp16必须手动将内存重新排列所以暂时不用fp16)
|
1278 |
+
net_g.load_state_dict(cpt["weight"], strict=False)
|
1279 |
+
input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
|
1280 |
+
output_names = [
|
1281 |
+
"audio",
|
1282 |
+
]
|
1283 |
+
# net_g.construct_spkmixmap(n_speaker) 多角色混合轨道导出
|
1284 |
+
torch.onnx.export(
|
1285 |
+
net_g,
|
1286 |
+
(
|
1287 |
+
test_phone.to(device),
|
1288 |
+
test_phone_lengths.to(device),
|
1289 |
+
test_pitch.to(device),
|
1290 |
+
test_pitchf.to(device),
|
1291 |
+
test_ds.to(device),
|
1292 |
+
test_rnd.to(device),
|
1293 |
+
),
|
1294 |
+
ExportedPath,
|
1295 |
+
dynamic_axes={
|
1296 |
+
"phone": [1],
|
1297 |
+
"pitch": [1],
|
1298 |
+
"pitchf": [1],
|
1299 |
+
"rnd": [2],
|
1300 |
+
},
|
1301 |
+
do_constant_folding=False,
|
1302 |
+
opset_version=13,
|
1303 |
+
verbose=False,
|
1304 |
+
input_names=input_names,
|
1305 |
+
output_names=output_names,
|
1306 |
+
)
|
1307 |
+
return "Finished"
|
1308 |
+
|
1309 |
+
# sid0: Inferencing voice/ model name
|
1310 |
+
# f0_up_key: Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12):
|
1311 |
+
# opt_root: output folder path
|
1312 |
+
# f0method: pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive)
|
1313 |
+
# filter_radius: If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.
|
1314 |
+
# file_index: Path to the feature index file.:
|
1315 |
+
# Auto-detect index path
|
1316 |
+
# index_rate: Search feature ratio (controls accent strength, too high has artifacting): minimum=0, maximum=1
|
1317 |
+
# resample_sr: Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling (time consuming)
|
1318 |
+
# rms_mix_rate: Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume:
|
1319 |
+
# protect: Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy
|
1320 |
+
# dir_path: Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)
|
1321 |
+
# format1: choices=["wav", "flac", "mp3", "m4a"]
|
1322 |
+
|
1323 |
+
def run(sid0, paths, dir_path=None, f0_up_key=0, opt_root="opt", f0_method="pm", filter_radius=3, file_index="", file_index2=None, index_rate=1, resample_sr=0, rms_mix_rate=1, protect=0.33, format1="wav"):
|
1324 |
+
if (dir_path=='' or dir_path==None) and (paths == '' or paths==None):
|
1325 |
+
return "must provide either dir_input or file path"
|
1326 |
+
if paths != None or paths != '':
|
1327 |
+
tempfile = [file_to_tempfile(paths)]
|
1328 |
+
print(paths)
|
1329 |
+
print(protect)
|
1330 |
+
get_vc(sid0, protect, protect)
|
1331 |
+
|
1332 |
+
vc_output3 = vc_multi(
|
1333 |
+
0, # sid: 0
|
1334 |
+
dir_path, # dir_path:
|
1335 |
+
opt_root, # opt_root: opt
|
1336 |
+
tempfile, # paths: [<tempfile._TemporaryFileWrapper object at 0x7f42c7dbb970>]
|
1337 |
+
f0_up_key, # f0_up_key: -12.0
|
1338 |
+
f0_method, # f0_method: pm
|
1339 |
+
file_index, # file_index:
|
1340 |
+
file_index2, # file_index2:
|
1341 |
+
index_rate, # index_rate: 1
|
1342 |
+
filter_radius, # filter_radius: 3
|
1343 |
+
resample_sr, # resample_sr: 0
|
1344 |
+
rms_mix_rate, # rms_mix_rate: 1
|
1345 |
+
protect, # protect: 0.33
|
1346 |
+
format1 # format1: wav
|
1347 |
+
)
|
1348 |
+
out_path = paths
|
1349 |
+
wavfile.write(out_path, tgt_sr, vc_output3)
|
1350 |
+
|
1351 |
+
return vc_output3
|
1352 |
+
|
1353 |
+
def get_models():
|
1354 |
+
return names
|
1355 |
+
|
1356 |
+
|
1357 |
+
def file_to_tempfile(file_path):
|
1358 |
+
with open(file_path, 'rb') as file:
|
1359 |
+
temp_file = tempfile.TemporaryFile()
|
1360 |
+
temp_file.write(file.read())
|
1361 |
+
temp_file.seek(0)
|
1362 |
+
return temp_file
|
1363 |
+
|
1364 |
+
print(run('mymodelimran.pth', '/home/teewhy/Desktop/RVC/Retrieval-based-Voice-Conversion-WebUI/opt/abcxot47ylz.mp3.mp3'))
|
1365 |
+
|
1366 |
+
|
1367 |
+
|
1368 |
+
|
1369 |
+
|
1370 |
+
|
1371 |
+
|
1372 |
+
|
1373 |
+
|
1374 |
+
|
1375 |
+
|
1376 |
+
|
1377 |
+
# sid: 0
|
1378 |
+
# dir_path:
|
1379 |
+
# opt_root: opt
|
1380 |
+
# paths: [<tempfile._TemporaryFileWrapper object at 0x7f42c7dbb970>]
|
1381 |
+
# f0_up_key: -12.0
|
1382 |
+
# f0_method: pm
|
1383 |
+
# file_index:
|
1384 |
+
# file_index2:
|
1385 |
+
# index_rate: 1
|
1386 |
+
# filter_radius: 3
|
1387 |
+
# resample_sr: 0
|
1388 |
+
# rms_mix_rate: 1
|
1389 |
+
# protect: 0.33
|
1390 |
+
# format1: wav
|
1391 |
+
# -12.0
|
1392 |
+
# input_audio_path: /tmp/vocal_7x2a5ym2vczt3tom5.mp3_10.wav
|
test.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import RVC_class
|
2 |
+
converter = RVC_class.VoiceConverter()
|
3 |
+
result = converter.single_run('https://tmpfiles.org/dl/1669357/recordonline-voice-recorder.com2.wav', 'imran_khan.pth', -12, False)
|
4 |
+
|
5 |
+
print(result)
|
train_nsf_sim_cache_sid_load_pretrain.py
ADDED
@@ -0,0 +1,595 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys, os
|
2 |
+
|
3 |
+
now_dir = os.getcwd()
|
4 |
+
sys.path.append(os.path.join(now_dir))
|
5 |
+
sys.path.append(os.path.join(now_dir, "train"))
|
6 |
+
import utils
|
7 |
+
import datetime
|
8 |
+
|
9 |
+
hps = utils.get_hparams()
|
10 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
|
11 |
+
n_gpus = len(hps.gpus.split("-"))
|
12 |
+
from random import shuffle, randint
|
13 |
+
import traceback, json, argparse, itertools, math, torch, pdb
|
14 |
+
|
15 |
+
torch.backends.cudnn.deterministic = False
|
16 |
+
torch.backends.cudnn.benchmark = False
|
17 |
+
from torch import nn, optim
|
18 |
+
from torch.nn import functional as F
|
19 |
+
from torch.utils.data import DataLoader
|
20 |
+
from torch.utils.tensorboard import SummaryWriter
|
21 |
+
import torch.multiprocessing as mp
|
22 |
+
import torch.distributed as dist
|
23 |
+
from torch.nn.parallel import DistributedDataParallel as DDP
|
24 |
+
from torch.cuda.amp import autocast, GradScaler
|
25 |
+
from infer_pack import commons
|
26 |
+
from time import sleep
|
27 |
+
from time import time as ttime
|
28 |
+
from data_utils import (
|
29 |
+
TextAudioLoaderMultiNSFsid,
|
30 |
+
TextAudioLoader,
|
31 |
+
TextAudioCollateMultiNSFsid,
|
32 |
+
TextAudioCollate,
|
33 |
+
DistributedBucketSampler,
|
34 |
+
)
|
35 |
+
|
36 |
+
if hps.version == "v1":
|
37 |
+
from infer_pack.models import (
|
38 |
+
SynthesizerTrnMs256NSFsid as RVC_Model_f0,
|
39 |
+
SynthesizerTrnMs256NSFsid_nono as RVC_Model_nof0,
|
40 |
+
MultiPeriodDiscriminator,
|
41 |
+
)
|
42 |
+
else:
|
43 |
+
from infer_pack.models import (
|
44 |
+
SynthesizerTrnMs768NSFsid as RVC_Model_f0,
|
45 |
+
SynthesizerTrnMs768NSFsid_nono as RVC_Model_nof0,
|
46 |
+
MultiPeriodDiscriminatorV2 as MultiPeriodDiscriminator,
|
47 |
+
)
|
48 |
+
from losses import generator_loss, discriminator_loss, feature_loss, kl_loss
|
49 |
+
from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
|
50 |
+
from process_ckpt import savee
|
51 |
+
|
52 |
+
global_step = 0
|
53 |
+
|
54 |
+
|
55 |
+
class EpochRecorder:
|
56 |
+
def __init__(self):
|
57 |
+
self.last_time = ttime()
|
58 |
+
|
59 |
+
def record(self):
|
60 |
+
now_time = ttime()
|
61 |
+
elapsed_time = now_time - self.last_time
|
62 |
+
self.last_time = now_time
|
63 |
+
elapsed_time_str = str(datetime.timedelta(seconds=elapsed_time))
|
64 |
+
current_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
|
65 |
+
return f"[{current_time}] | ({elapsed_time_str})"
|
66 |
+
|
67 |
+
|
68 |
+
def main():
|
69 |
+
n_gpus = torch.cuda.device_count()
|
70 |
+
if torch.cuda.is_available() == False and torch.backends.mps.is_available() == True:
|
71 |
+
n_gpus = 1
|
72 |
+
os.environ["MASTER_ADDR"] = "localhost"
|
73 |
+
os.environ["MASTER_PORT"] = str(randint(20000, 55555))
|
74 |
+
children = []
|
75 |
+
for i in range(n_gpus):
|
76 |
+
subproc = mp.Process(
|
77 |
+
target=run,
|
78 |
+
args=(
|
79 |
+
i,
|
80 |
+
n_gpus,
|
81 |
+
hps,
|
82 |
+
),
|
83 |
+
)
|
84 |
+
children.append(subproc)
|
85 |
+
subproc.start()
|
86 |
+
|
87 |
+
for i in range(n_gpus):
|
88 |
+
children[i].join()
|
89 |
+
|
90 |
+
|
91 |
+
def run(rank, n_gpus, hps):
|
92 |
+
global global_step
|
93 |
+
if rank == 0:
|
94 |
+
logger = utils.get_logger(hps.model_dir)
|
95 |
+
logger.info(hps)
|
96 |
+
# utils.check_git_hash(hps.model_dir)
|
97 |
+
writer = SummaryWriter(log_dir=hps.model_dir)
|
98 |
+
writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
|
99 |
+
|
100 |
+
dist.init_process_group(
|
101 |
+
backend="gloo", init_method="env://", world_size=n_gpus, rank=rank
|
102 |
+
)
|
103 |
+
torch.manual_seed(hps.train.seed)
|
104 |
+
if torch.cuda.is_available():
|
105 |
+
torch.cuda.set_device(rank)
|
106 |
+
|
107 |
+
if hps.if_f0 == 1:
|
108 |
+
train_dataset = TextAudioLoaderMultiNSFsid(hps.data.training_files, hps.data)
|
109 |
+
else:
|
110 |
+
train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
|
111 |
+
train_sampler = DistributedBucketSampler(
|
112 |
+
train_dataset,
|
113 |
+
hps.train.batch_size * n_gpus,
|
114 |
+
# [100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200,1400], # 16s
|
115 |
+
[100, 200, 300, 400, 500, 600, 700, 800, 900], # 16s
|
116 |
+
num_replicas=n_gpus,
|
117 |
+
rank=rank,
|
118 |
+
shuffle=True,
|
119 |
+
)
|
120 |
+
# It is possible that dataloader's workers are out of shared memory. Please try to raise your shared memory limit.
|
121 |
+
# num_workers=8 -> num_workers=4
|
122 |
+
if hps.if_f0 == 1:
|
123 |
+
collate_fn = TextAudioCollateMultiNSFsid()
|
124 |
+
else:
|
125 |
+
collate_fn = TextAudioCollate()
|
126 |
+
train_loader = DataLoader(
|
127 |
+
train_dataset,
|
128 |
+
num_workers=4,
|
129 |
+
shuffle=False,
|
130 |
+
pin_memory=True,
|
131 |
+
collate_fn=collate_fn,
|
132 |
+
batch_sampler=train_sampler,
|
133 |
+
persistent_workers=True,
|
134 |
+
prefetch_factor=8,
|
135 |
+
)
|
136 |
+
if hps.if_f0 == 1:
|
137 |
+
net_g = RVC_Model_f0(
|
138 |
+
hps.data.filter_length // 2 + 1,
|
139 |
+
hps.train.segment_size // hps.data.hop_length,
|
140 |
+
**hps.model,
|
141 |
+
is_half=hps.train.fp16_run,
|
142 |
+
sr=hps.sample_rate,
|
143 |
+
)
|
144 |
+
else:
|
145 |
+
net_g = RVC_Model_nof0(
|
146 |
+
hps.data.filter_length // 2 + 1,
|
147 |
+
hps.train.segment_size // hps.data.hop_length,
|
148 |
+
**hps.model,
|
149 |
+
is_half=hps.train.fp16_run,
|
150 |
+
)
|
151 |
+
if torch.cuda.is_available():
|
152 |
+
net_g = net_g.cuda(rank)
|
153 |
+
net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm)
|
154 |
+
if torch.cuda.is_available():
|
155 |
+
net_d = net_d.cuda(rank)
|
156 |
+
optim_g = torch.optim.AdamW(
|
157 |
+
net_g.parameters(),
|
158 |
+
hps.train.learning_rate,
|
159 |
+
betas=hps.train.betas,
|
160 |
+
eps=hps.train.eps,
|
161 |
+
)
|
162 |
+
optim_d = torch.optim.AdamW(
|
163 |
+
net_d.parameters(),
|
164 |
+
hps.train.learning_rate,
|
165 |
+
betas=hps.train.betas,
|
166 |
+
eps=hps.train.eps,
|
167 |
+
)
|
168 |
+
# net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
|
169 |
+
# net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
|
170 |
+
if torch.cuda.is_available():
|
171 |
+
net_g = DDP(net_g, device_ids=[rank])
|
172 |
+
net_d = DDP(net_d, device_ids=[rank])
|
173 |
+
else:
|
174 |
+
net_g = DDP(net_g)
|
175 |
+
net_d = DDP(net_d)
|
176 |
+
|
177 |
+
try: # 如果能加载自动resume
|
178 |
+
_, _, _, epoch_str = utils.load_checkpoint(
|
179 |
+
utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d
|
180 |
+
) # D多半加载没事
|
181 |
+
if rank == 0:
|
182 |
+
logger.info("loaded D")
|
183 |
+
# _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g,load_opt=0)
|
184 |
+
_, _, _, epoch_str = utils.load_checkpoint(
|
185 |
+
utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g
|
186 |
+
)
|
187 |
+
global_step = (epoch_str - 1) * len(train_loader)
|
188 |
+
# epoch_str = 1
|
189 |
+
# global_step = 0
|
190 |
+
except: # 如果首次不能加载,加载pretrain
|
191 |
+
# traceback.print_exc()
|
192 |
+
epoch_str = 1
|
193 |
+
global_step = 0
|
194 |
+
if hps.pretrainG != "":
|
195 |
+
if rank == 0:
|
196 |
+
logger.info("loaded pretrained %s" % (hps.pretrainG))
|
197 |
+
print(
|
198 |
+
net_g.module.load_state_dict(
|
199 |
+
torch.load(hps.pretrainG, map_location="cpu")["model"]
|
200 |
+
)
|
201 |
+
) ##测试不加载优化器
|
202 |
+
if hps.pretrainD != "":
|
203 |
+
if rank == 0:
|
204 |
+
logger.info("loaded pretrained %s" % (hps.pretrainD))
|
205 |
+
print(
|
206 |
+
net_d.module.load_state_dict(
|
207 |
+
torch.load(hps.pretrainD, map_location="cpu")["model"]
|
208 |
+
)
|
209 |
+
)
|
210 |
+
|
211 |
+
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
|
212 |
+
optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
|
213 |
+
)
|
214 |
+
scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
|
215 |
+
optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
|
216 |
+
)
|
217 |
+
|
218 |
+
scaler = GradScaler(enabled=hps.train.fp16_run)
|
219 |
+
|
220 |
+
cache = []
|
221 |
+
for epoch in range(epoch_str, hps.train.epochs + 1):
|
222 |
+
if rank == 0:
|
223 |
+
train_and_evaluate(
|
224 |
+
rank,
|
225 |
+
epoch,
|
226 |
+
hps,
|
227 |
+
[net_g, net_d],
|
228 |
+
[optim_g, optim_d],
|
229 |
+
[scheduler_g, scheduler_d],
|
230 |
+
scaler,
|
231 |
+
[train_loader, None],
|
232 |
+
logger,
|
233 |
+
[writer, writer_eval],
|
234 |
+
cache,
|
235 |
+
)
|
236 |
+
else:
|
237 |
+
train_and_evaluate(
|
238 |
+
rank,
|
239 |
+
epoch,
|
240 |
+
hps,
|
241 |
+
[net_g, net_d],
|
242 |
+
[optim_g, optim_d],
|
243 |
+
[scheduler_g, scheduler_d],
|
244 |
+
scaler,
|
245 |
+
[train_loader, None],
|
246 |
+
None,
|
247 |
+
None,
|
248 |
+
cache,
|
249 |
+
)
|
250 |
+
scheduler_g.step()
|
251 |
+
scheduler_d.step()
|
252 |
+
|
253 |
+
|
254 |
+
def train_and_evaluate(
|
255 |
+
rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers, cache
|
256 |
+
):
|
257 |
+
net_g, net_d = nets
|
258 |
+
optim_g, optim_d = optims
|
259 |
+
train_loader, eval_loader = loaders
|
260 |
+
if writers is not None:
|
261 |
+
writer, writer_eval = writers
|
262 |
+
|
263 |
+
train_loader.batch_sampler.set_epoch(epoch)
|
264 |
+
global global_step
|
265 |
+
|
266 |
+
net_g.train()
|
267 |
+
net_d.train()
|
268 |
+
|
269 |
+
# Prepare data iterator
|
270 |
+
if hps.if_cache_data_in_gpu == True:
|
271 |
+
# Use Cache
|
272 |
+
data_iterator = cache
|
273 |
+
if cache == []:
|
274 |
+
# Make new cache
|
275 |
+
for batch_idx, info in enumerate(train_loader):
|
276 |
+
# Unpack
|
277 |
+
if hps.if_f0 == 1:
|
278 |
+
(
|
279 |
+
phone,
|
280 |
+
phone_lengths,
|
281 |
+
pitch,
|
282 |
+
pitchf,
|
283 |
+
spec,
|
284 |
+
spec_lengths,
|
285 |
+
wave,
|
286 |
+
wave_lengths,
|
287 |
+
sid,
|
288 |
+
) = info
|
289 |
+
else:
|
290 |
+
(
|
291 |
+
phone,
|
292 |
+
phone_lengths,
|
293 |
+
spec,
|
294 |
+
spec_lengths,
|
295 |
+
wave,
|
296 |
+
wave_lengths,
|
297 |
+
sid,
|
298 |
+
) = info
|
299 |
+
# Load on CUDA
|
300 |
+
if torch.cuda.is_available():
|
301 |
+
phone = phone.cuda(rank, non_blocking=True)
|
302 |
+
phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
|
303 |
+
if hps.if_f0 == 1:
|
304 |
+
pitch = pitch.cuda(rank, non_blocking=True)
|
305 |
+
pitchf = pitchf.cuda(rank, non_blocking=True)
|
306 |
+
sid = sid.cuda(rank, non_blocking=True)
|
307 |
+
spec = spec.cuda(rank, non_blocking=True)
|
308 |
+
spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
|
309 |
+
wave = wave.cuda(rank, non_blocking=True)
|
310 |
+
wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
|
311 |
+
# Cache on list
|
312 |
+
if hps.if_f0 == 1:
|
313 |
+
cache.append(
|
314 |
+
(
|
315 |
+
batch_idx,
|
316 |
+
(
|
317 |
+
phone,
|
318 |
+
phone_lengths,
|
319 |
+
pitch,
|
320 |
+
pitchf,
|
321 |
+
spec,
|
322 |
+
spec_lengths,
|
323 |
+
wave,
|
324 |
+
wave_lengths,
|
325 |
+
sid,
|
326 |
+
),
|
327 |
+
)
|
328 |
+
)
|
329 |
+
else:
|
330 |
+
cache.append(
|
331 |
+
(
|
332 |
+
batch_idx,
|
333 |
+
(
|
334 |
+
phone,
|
335 |
+
phone_lengths,
|
336 |
+
spec,
|
337 |
+
spec_lengths,
|
338 |
+
wave,
|
339 |
+
wave_lengths,
|
340 |
+
sid,
|
341 |
+
),
|
342 |
+
)
|
343 |
+
)
|
344 |
+
else:
|
345 |
+
# Load shuffled cache
|
346 |
+
shuffle(cache)
|
347 |
+
else:
|
348 |
+
# Loader
|
349 |
+
data_iterator = enumerate(train_loader)
|
350 |
+
|
351 |
+
# Run steps
|
352 |
+
epoch_recorder = EpochRecorder()
|
353 |
+
for batch_idx, info in data_iterator:
|
354 |
+
# Data
|
355 |
+
## Unpack
|
356 |
+
if hps.if_f0 == 1:
|
357 |
+
(
|
358 |
+
phone,
|
359 |
+
phone_lengths,
|
360 |
+
pitch,
|
361 |
+
pitchf,
|
362 |
+
spec,
|
363 |
+
spec_lengths,
|
364 |
+
wave,
|
365 |
+
wave_lengths,
|
366 |
+
sid,
|
367 |
+
) = info
|
368 |
+
else:
|
369 |
+
phone, phone_lengths, spec, spec_lengths, wave, wave_lengths, sid = info
|
370 |
+
## Load on CUDA
|
371 |
+
if (hps.if_cache_data_in_gpu == False) and torch.cuda.is_available():
|
372 |
+
phone = phone.cuda(rank, non_blocking=True)
|
373 |
+
phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
|
374 |
+
if hps.if_f0 == 1:
|
375 |
+
pitch = pitch.cuda(rank, non_blocking=True)
|
376 |
+
pitchf = pitchf.cuda(rank, non_blocking=True)
|
377 |
+
sid = sid.cuda(rank, non_blocking=True)
|
378 |
+
spec = spec.cuda(rank, non_blocking=True)
|
379 |
+
spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
|
380 |
+
wave = wave.cuda(rank, non_blocking=True)
|
381 |
+
# wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
|
382 |
+
|
383 |
+
# Calculate
|
384 |
+
with autocast(enabled=hps.train.fp16_run):
|
385 |
+
if hps.if_f0 == 1:
|
386 |
+
(
|
387 |
+
y_hat,
|
388 |
+
ids_slice,
|
389 |
+
x_mask,
|
390 |
+
z_mask,
|
391 |
+
(z, z_p, m_p, logs_p, m_q, logs_q),
|
392 |
+
) = net_g(phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid)
|
393 |
+
else:
|
394 |
+
(
|
395 |
+
y_hat,
|
396 |
+
ids_slice,
|
397 |
+
x_mask,
|
398 |
+
z_mask,
|
399 |
+
(z, z_p, m_p, logs_p, m_q, logs_q),
|
400 |
+
) = net_g(phone, phone_lengths, spec, spec_lengths, sid)
|
401 |
+
mel = spec_to_mel_torch(
|
402 |
+
spec,
|
403 |
+
hps.data.filter_length,
|
404 |
+
hps.data.n_mel_channels,
|
405 |
+
hps.data.sampling_rate,
|
406 |
+
hps.data.mel_fmin,
|
407 |
+
hps.data.mel_fmax,
|
408 |
+
)
|
409 |
+
y_mel = commons.slice_segments(
|
410 |
+
mel, ids_slice, hps.train.segment_size // hps.data.hop_length
|
411 |
+
)
|
412 |
+
with autocast(enabled=False):
|
413 |
+
y_hat_mel = mel_spectrogram_torch(
|
414 |
+
y_hat.float().squeeze(1),
|
415 |
+
hps.data.filter_length,
|
416 |
+
hps.data.n_mel_channels,
|
417 |
+
hps.data.sampling_rate,
|
418 |
+
hps.data.hop_length,
|
419 |
+
hps.data.win_length,
|
420 |
+
hps.data.mel_fmin,
|
421 |
+
hps.data.mel_fmax,
|
422 |
+
)
|
423 |
+
if hps.train.fp16_run == True:
|
424 |
+
y_hat_mel = y_hat_mel.half()
|
425 |
+
wave = commons.slice_segments(
|
426 |
+
wave, ids_slice * hps.data.hop_length, hps.train.segment_size
|
427 |
+
) # slice
|
428 |
+
|
429 |
+
# Discriminator
|
430 |
+
y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
|
431 |
+
with autocast(enabled=False):
|
432 |
+
loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
|
433 |
+
y_d_hat_r, y_d_hat_g
|
434 |
+
)
|
435 |
+
optim_d.zero_grad()
|
436 |
+
scaler.scale(loss_disc).backward()
|
437 |
+
scaler.unscale_(optim_d)
|
438 |
+
grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
|
439 |
+
scaler.step(optim_d)
|
440 |
+
|
441 |
+
with autocast(enabled=hps.train.fp16_run):
|
442 |
+
# Generator
|
443 |
+
y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
|
444 |
+
with autocast(enabled=False):
|
445 |
+
loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
|
446 |
+
loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
|
447 |
+
loss_fm = feature_loss(fmap_r, fmap_g)
|
448 |
+
loss_gen, losses_gen = generator_loss(y_d_hat_g)
|
449 |
+
loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
|
450 |
+
optim_g.zero_grad()
|
451 |
+
scaler.scale(loss_gen_all).backward()
|
452 |
+
scaler.unscale_(optim_g)
|
453 |
+
grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
|
454 |
+
scaler.step(optim_g)
|
455 |
+
scaler.update()
|
456 |
+
|
457 |
+
if rank == 0:
|
458 |
+
if global_step % hps.train.log_interval == 0:
|
459 |
+
lr = optim_g.param_groups[0]["lr"]
|
460 |
+
logger.info(
|
461 |
+
"Train Epoch: {} [{:.0f}%]".format(
|
462 |
+
epoch, 100.0 * batch_idx / len(train_loader)
|
463 |
+
)
|
464 |
+
)
|
465 |
+
# Amor For Tensorboard display
|
466 |
+
if loss_mel > 75:
|
467 |
+
loss_mel = 75
|
468 |
+
if loss_kl > 9:
|
469 |
+
loss_kl = 9
|
470 |
+
|
471 |
+
logger.info([global_step, lr])
|
472 |
+
logger.info(
|
473 |
+
f"loss_disc={loss_disc:.3f}, loss_gen={loss_gen:.3f}, loss_fm={loss_fm:.3f},loss_mel={loss_mel:.3f}, loss_kl={loss_kl:.3f}"
|
474 |
+
)
|
475 |
+
scalar_dict = {
|
476 |
+
"loss/g/total": loss_gen_all,
|
477 |
+
"loss/d/total": loss_disc,
|
478 |
+
"learning_rate": lr,
|
479 |
+
"grad_norm_d": grad_norm_d,
|
480 |
+
"grad_norm_g": grad_norm_g,
|
481 |
+
}
|
482 |
+
scalar_dict.update(
|
483 |
+
{
|
484 |
+
"loss/g/fm": loss_fm,
|
485 |
+
"loss/g/mel": loss_mel,
|
486 |
+
"loss/g/kl": loss_kl,
|
487 |
+
}
|
488 |
+
)
|
489 |
+
|
490 |
+
scalar_dict.update(
|
491 |
+
{"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
|
492 |
+
)
|
493 |
+
scalar_dict.update(
|
494 |
+
{"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
|
495 |
+
)
|
496 |
+
scalar_dict.update(
|
497 |
+
{"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
|
498 |
+
)
|
499 |
+
image_dict = {
|
500 |
+
"slice/mel_org": utils.plot_spectrogram_to_numpy(
|
501 |
+
y_mel[0].data.cpu().numpy()
|
502 |
+
),
|
503 |
+
"slice/mel_gen": utils.plot_spectrogram_to_numpy(
|
504 |
+
y_hat_mel[0].data.cpu().numpy()
|
505 |
+
),
|
506 |
+
"all/mel": utils.plot_spectrogram_to_numpy(
|
507 |
+
mel[0].data.cpu().numpy()
|
508 |
+
),
|
509 |
+
}
|
510 |
+
utils.summarize(
|
511 |
+
writer=writer,
|
512 |
+
global_step=global_step,
|
513 |
+
images=image_dict,
|
514 |
+
scalars=scalar_dict,
|
515 |
+
)
|
516 |
+
global_step += 1
|
517 |
+
# /Run steps
|
518 |
+
|
519 |
+
if epoch % hps.save_every_epoch == 0 and rank == 0:
|
520 |
+
if hps.if_latest == 0:
|
521 |
+
utils.save_checkpoint(
|
522 |
+
net_g,
|
523 |
+
optim_g,
|
524 |
+
hps.train.learning_rate,
|
525 |
+
epoch,
|
526 |
+
os.path.join(hps.model_dir, "G_{}.pth".format(global_step)),
|
527 |
+
)
|
528 |
+
utils.save_checkpoint(
|
529 |
+
net_d,
|
530 |
+
optim_d,
|
531 |
+
hps.train.learning_rate,
|
532 |
+
epoch,
|
533 |
+
os.path.join(hps.model_dir, "D_{}.pth".format(global_step)),
|
534 |
+
)
|
535 |
+
else:
|
536 |
+
utils.save_checkpoint(
|
537 |
+
net_g,
|
538 |
+
optim_g,
|
539 |
+
hps.train.learning_rate,
|
540 |
+
epoch,
|
541 |
+
os.path.join(hps.model_dir, "G_{}.pth".format(2333333)),
|
542 |
+
)
|
543 |
+
utils.save_checkpoint(
|
544 |
+
net_d,
|
545 |
+
optim_d,
|
546 |
+
hps.train.learning_rate,
|
547 |
+
epoch,
|
548 |
+
os.path.join(hps.model_dir, "D_{}.pth".format(2333333)),
|
549 |
+
)
|
550 |
+
if rank == 0 and hps.save_every_weights == "1":
|
551 |
+
if hasattr(net_g, "module"):
|
552 |
+
ckpt = net_g.module.state_dict()
|
553 |
+
else:
|
554 |
+
ckpt = net_g.state_dict()
|
555 |
+
logger.info(
|
556 |
+
"saving ckpt %s_e%s:%s"
|
557 |
+
% (
|
558 |
+
hps.name,
|
559 |
+
epoch,
|
560 |
+
savee(
|
561 |
+
ckpt,
|
562 |
+
hps.sample_rate,
|
563 |
+
hps.if_f0,
|
564 |
+
hps.name + "_e%s_s%s" % (epoch, global_step),
|
565 |
+
epoch,
|
566 |
+
hps.version,
|
567 |
+
hps,
|
568 |
+
),
|
569 |
+
)
|
570 |
+
)
|
571 |
+
|
572 |
+
if rank == 0:
|
573 |
+
logger.info("====> Epoch: {} {}".format(epoch, epoch_recorder.record()))
|
574 |
+
if epoch >= hps.total_epoch and rank == 0:
|
575 |
+
logger.info("Training is done. The program is closed.")
|
576 |
+
|
577 |
+
if hasattr(net_g, "module"):
|
578 |
+
ckpt = net_g.module.state_dict()
|
579 |
+
else:
|
580 |
+
ckpt = net_g.state_dict()
|
581 |
+
logger.info(
|
582 |
+
"saving final ckpt:%s"
|
583 |
+
% (
|
584 |
+
savee(
|
585 |
+
ckpt, hps.sample_rate, hps.if_f0, hps.name, epoch, hps.version, hps
|
586 |
+
)
|
587 |
+
)
|
588 |
+
)
|
589 |
+
sleep(1)
|
590 |
+
os._exit(2333333)
|
591 |
+
|
592 |
+
|
593 |
+
if __name__ == "__main__":
|
594 |
+
torch.multiprocessing.set_start_method("spawn")
|
595 |
+
main()
|
trainset_preprocess_pipeline_print.py
ADDED
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys, os, multiprocessing
|
2 |
+
from scipy import signal
|
3 |
+
|
4 |
+
now_dir = os.getcwd()
|
5 |
+
sys.path.append(now_dir)
|
6 |
+
|
7 |
+
inp_root = sys.argv[1]
|
8 |
+
sr = int(sys.argv[2])
|
9 |
+
n_p = int(sys.argv[3])
|
10 |
+
exp_dir = sys.argv[4]
|
11 |
+
noparallel = sys.argv[5] == "True"
|
12 |
+
import numpy as np, os, traceback
|
13 |
+
from slicer2 import Slicer
|
14 |
+
import librosa, traceback
|
15 |
+
from scipy.io import wavfile
|
16 |
+
import multiprocessing
|
17 |
+
from my_utils import load_audio
|
18 |
+
|
19 |
+
mutex = multiprocessing.Lock()
|
20 |
+
f = open("%s/preprocess.log" % exp_dir, "a+")
|
21 |
+
|
22 |
+
|
23 |
+
def println(strr):
|
24 |
+
mutex.acquire()
|
25 |
+
print(strr)
|
26 |
+
f.write("%s\n" % strr)
|
27 |
+
f.flush()
|
28 |
+
mutex.release()
|
29 |
+
|
30 |
+
|
31 |
+
class PreProcess:
|
32 |
+
def __init__(self, sr, exp_dir):
|
33 |
+
self.slicer = Slicer(
|
34 |
+
sr=sr,
|
35 |
+
threshold=-42,
|
36 |
+
min_length=1500,
|
37 |
+
min_interval=400,
|
38 |
+
hop_size=15,
|
39 |
+
max_sil_kept=500,
|
40 |
+
)
|
41 |
+
self.sr = sr
|
42 |
+
self.bh, self.ah = signal.butter(N=5, Wn=48, btype="high", fs=self.sr)
|
43 |
+
self.per = 3.0
|
44 |
+
self.overlap = 0.3
|
45 |
+
self.tail = self.per + self.overlap
|
46 |
+
self.max = 0.9
|
47 |
+
self.alpha = 0.75
|
48 |
+
self.exp_dir = exp_dir
|
49 |
+
self.gt_wavs_dir = "%s/0_gt_wavs" % exp_dir
|
50 |
+
self.wavs16k_dir = "%s/1_16k_wavs" % exp_dir
|
51 |
+
os.makedirs(self.exp_dir, exist_ok=True)
|
52 |
+
os.makedirs(self.gt_wavs_dir, exist_ok=True)
|
53 |
+
os.makedirs(self.wavs16k_dir, exist_ok=True)
|
54 |
+
|
55 |
+
def norm_write(self, tmp_audio, idx0, idx1):
|
56 |
+
tmp_max = np.abs(tmp_audio).max()
|
57 |
+
if tmp_max > 2.5:
|
58 |
+
print("%s-%s-%s-filtered" % (idx0, idx1, tmp_max))
|
59 |
+
return
|
60 |
+
tmp_audio = (tmp_audio / tmp_max * (self.max * self.alpha)) + (
|
61 |
+
1 - self.alpha
|
62 |
+
) * tmp_audio
|
63 |
+
wavfile.write(
|
64 |
+
"%s/%s_%s.wav" % (self.gt_wavs_dir, idx0, idx1),
|
65 |
+
self.sr,
|
66 |
+
tmp_audio.astype(np.float32),
|
67 |
+
)
|
68 |
+
tmp_audio = librosa.resample(
|
69 |
+
tmp_audio, orig_sr=self.sr, target_sr=16000
|
70 |
+
) # , res_type="soxr_vhq"
|
71 |
+
wavfile.write(
|
72 |
+
"%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1),
|
73 |
+
16000,
|
74 |
+
tmp_audio.astype(np.float32),
|
75 |
+
)
|
76 |
+
|
77 |
+
def pipeline(self, path, idx0):
|
78 |
+
try:
|
79 |
+
audio = load_audio(path, self.sr)
|
80 |
+
# zero phased digital filter cause pre-ringing noise...
|
81 |
+
# audio = signal.filtfilt(self.bh, self.ah, audio)
|
82 |
+
audio = signal.lfilter(self.bh, self.ah, audio)
|
83 |
+
|
84 |
+
idx1 = 0
|
85 |
+
for audio in self.slicer.slice(audio):
|
86 |
+
i = 0
|
87 |
+
while 1:
|
88 |
+
start = int(self.sr * (self.per - self.overlap) * i)
|
89 |
+
i += 1
|
90 |
+
if len(audio[start:]) > self.tail * self.sr:
|
91 |
+
tmp_audio = audio[start : start + int(self.per * self.sr)]
|
92 |
+
self.norm_write(tmp_audio, idx0, idx1)
|
93 |
+
idx1 += 1
|
94 |
+
else:
|
95 |
+
tmp_audio = audio[start:]
|
96 |
+
idx1 += 1
|
97 |
+
break
|
98 |
+
self.norm_write(tmp_audio, idx0, idx1)
|
99 |
+
println("%s->Suc." % path)
|
100 |
+
except:
|
101 |
+
println("%s->%s" % (path, traceback.format_exc()))
|
102 |
+
|
103 |
+
def pipeline_mp(self, infos):
|
104 |
+
for path, idx0 in infos:
|
105 |
+
self.pipeline(path, idx0)
|
106 |
+
|
107 |
+
def pipeline_mp_inp_dir(self, inp_root, n_p):
|
108 |
+
try:
|
109 |
+
infos = [
|
110 |
+
("%s/%s" % (inp_root, name), idx)
|
111 |
+
for idx, name in enumerate(sorted(list(os.listdir(inp_root))))
|
112 |
+
]
|
113 |
+
if noparallel:
|
114 |
+
for i in range(n_p):
|
115 |
+
self.pipeline_mp(infos[i::n_p])
|
116 |
+
else:
|
117 |
+
ps = []
|
118 |
+
for i in range(n_p):
|
119 |
+
p = multiprocessing.Process(
|
120 |
+
target=self.pipeline_mp, args=(infos[i::n_p],)
|
121 |
+
)
|
122 |
+
ps.append(p)
|
123 |
+
p.start()
|
124 |
+
for i in range(n_p):
|
125 |
+
ps[i].join()
|
126 |
+
except:
|
127 |
+
println("Fail. %s" % traceback.format_exc())
|
128 |
+
|
129 |
+
|
130 |
+
def preprocess_trainset(inp_root, sr, n_p, exp_dir):
|
131 |
+
pp = PreProcess(sr, exp_dir)
|
132 |
+
println("start preprocess")
|
133 |
+
println(sys.argv)
|
134 |
+
pp.pipeline_mp_inp_dir(inp_root, n_p)
|
135 |
+
println("end preprocess")
|
136 |
+
|
137 |
+
|
138 |
+
if __name__ == "__main__":
|
139 |
+
preprocess_trainset(inp_root, sr, n_p, exp_dir)
|
vc_infer_pipeline.py
ADDED
@@ -0,0 +1,449 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np, parselmouth, torch, pdb
|
2 |
+
from time import time as ttime
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import scipy.signal as signal
|
5 |
+
import pyworld, os, traceback, faiss, librosa, torchcrepe
|
6 |
+
from scipy import signal
|
7 |
+
from functools import lru_cache
|
8 |
+
import logging
|
9 |
+
bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
|
10 |
+
|
11 |
+
input_audio_path2wav = {}
|
12 |
+
|
13 |
+
|
14 |
+
@lru_cache
|
15 |
+
def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
|
16 |
+
audio = input_audio_path2wav[input_audio_path]
|
17 |
+
f0, t = pyworld.harvest(
|
18 |
+
audio,
|
19 |
+
fs=fs,
|
20 |
+
f0_ceil=f0max,
|
21 |
+
f0_floor=f0min,
|
22 |
+
frame_period=frame_period,
|
23 |
+
)
|
24 |
+
f0 = pyworld.stonemask(audio, f0, t, fs)
|
25 |
+
return f0
|
26 |
+
|
27 |
+
|
28 |
+
def change_rms(data1, sr1, data2, sr2, rate): # 1是输入音频,2是输出音频,rate是2的占比
|
29 |
+
# print(data1.max(),data2.max())
|
30 |
+
rms1 = librosa.feature.rms(
|
31 |
+
y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2
|
32 |
+
) # 每半秒一个点
|
33 |
+
rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
|
34 |
+
rms1 = torch.from_numpy(rms1)
|
35 |
+
rms1 = F.interpolate(
|
36 |
+
rms1.unsqueeze(0), size=data2.shape[0], mode="linear"
|
37 |
+
).squeeze()
|
38 |
+
rms2 = torch.from_numpy(rms2)
|
39 |
+
rms2 = F.interpolate(
|
40 |
+
rms2.unsqueeze(0), size=data2.shape[0], mode="linear"
|
41 |
+
).squeeze()
|
42 |
+
rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
|
43 |
+
data2 *= (
|
44 |
+
torch.pow(rms1, torch.tensor(1 - rate))
|
45 |
+
* torch.pow(rms2, torch.tensor(rate - 1))
|
46 |
+
).numpy()
|
47 |
+
return data2
|
48 |
+
|
49 |
+
|
50 |
+
class VC(object):
|
51 |
+
def __init__(self, tgt_sr, config):
|
52 |
+
self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
|
53 |
+
config.x_pad,
|
54 |
+
config.x_query,
|
55 |
+
config.x_center,
|
56 |
+
config.x_max,
|
57 |
+
config.is_half,
|
58 |
+
)
|
59 |
+
self.sr = 16000 # hubert输入采样率
|
60 |
+
self.window = 160 # 每帧点数
|
61 |
+
self.t_pad = self.sr * self.x_pad # 每条前后pad时间
|
62 |
+
self.t_pad_tgt = tgt_sr * self.x_pad
|
63 |
+
self.t_pad2 = self.t_pad * 2
|
64 |
+
self.t_query = self.sr * self.x_query # 查询切点前后查询时间
|
65 |
+
self.t_center = self.sr * self.x_center # 查询切点位置
|
66 |
+
self.t_max = self.sr * self.x_max # 免查询时长阈值
|
67 |
+
self.device = config.device
|
68 |
+
|
69 |
+
def get_f0(
|
70 |
+
self,
|
71 |
+
input_audio_path,
|
72 |
+
x,
|
73 |
+
p_len,
|
74 |
+
f0_up_key,
|
75 |
+
f0_method,
|
76 |
+
filter_radius,
|
77 |
+
inp_f0=None,
|
78 |
+
):
|
79 |
+
logging.info("start get_f0")
|
80 |
+
global input_audio_path2wav
|
81 |
+
time_step = self.window / self.sr * 1000
|
82 |
+
f0_min = 50
|
83 |
+
f0_max = 1100
|
84 |
+
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
|
85 |
+
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
|
86 |
+
logging.info("before if_method == pm")
|
87 |
+
if f0_method == "pm":
|
88 |
+
f0 = (
|
89 |
+
parselmouth.Sound(x, self.sr)
|
90 |
+
.to_pitch_ac(
|
91 |
+
time_step=time_step / 1000,
|
92 |
+
voicing_threshold=0.6,
|
93 |
+
pitch_floor=f0_min,
|
94 |
+
pitch_ceiling=f0_max,
|
95 |
+
)
|
96 |
+
.selected_array["frequency"]
|
97 |
+
)
|
98 |
+
pad_size = (p_len - len(f0) + 1) // 2
|
99 |
+
if pad_size > 0 or p_len - len(f0) - pad_size > 0:
|
100 |
+
f0 = np.pad(
|
101 |
+
f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
|
102 |
+
)
|
103 |
+
elif f0_method == "harvest":
|
104 |
+
input_audio_path2wav[input_audio_path] = x.astype(np.double)
|
105 |
+
f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10)
|
106 |
+
if filter_radius > 2:
|
107 |
+
f0 = signal.medfilt(f0, 3)
|
108 |
+
elif f0_method == "crepe":
|
109 |
+
logging.info("inside elif if_method == crepe")
|
110 |
+
model = "full"
|
111 |
+
# Pick a batch size that doesn't cause memory errors on your gpu
|
112 |
+
batch_size = 512
|
113 |
+
# Compute pitch using first gpu
|
114 |
+
audio = torch.tensor(np.copy(x))[None].float()
|
115 |
+
logging.info("before torchcrepe.predict")
|
116 |
+
f0, pd = torchcrepe.predict(
|
117 |
+
audio,
|
118 |
+
self.sr,
|
119 |
+
self.window,
|
120 |
+
f0_min,
|
121 |
+
f0_max,
|
122 |
+
model,
|
123 |
+
batch_size=batch_size,
|
124 |
+
device=self.device,
|
125 |
+
return_periodicity=True,
|
126 |
+
)
|
127 |
+
logging.info("after torchcrepe.predict")
|
128 |
+
pd = torchcrepe.filter.median(pd, 3)
|
129 |
+
f0 = torchcrepe.filter.mean(f0, 3)
|
130 |
+
f0[pd < 0.1] = 0
|
131 |
+
f0 = f0[0].cpu().numpy()
|
132 |
+
f0 *= pow(2, f0_up_key / 12)
|
133 |
+
# with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
|
134 |
+
tf0 = self.sr // self.window # 每秒f0点数
|
135 |
+
if inp_f0 is not None:
|
136 |
+
delta_t = np.round(
|
137 |
+
(inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
|
138 |
+
).astype("int16")
|
139 |
+
replace_f0 = np.interp(
|
140 |
+
list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
|
141 |
+
)
|
142 |
+
shape = f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]
|
143 |
+
f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
|
144 |
+
:shape
|
145 |
+
]
|
146 |
+
# with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
|
147 |
+
f0bak = f0.copy()
|
148 |
+
f0_mel = 1127 * np.log(1 + f0 / 700)
|
149 |
+
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
|
150 |
+
f0_mel_max - f0_mel_min
|
151 |
+
) + 1
|
152 |
+
f0_mel[f0_mel <= 1] = 1
|
153 |
+
f0_mel[f0_mel > 255] = 255
|
154 |
+
f0_coarse = np.rint(f0_mel).astype(np.int)
|
155 |
+
return f0_coarse, f0bak # 1-0
|
156 |
+
|
157 |
+
def vc(
|
158 |
+
self,
|
159 |
+
model,
|
160 |
+
net_g,
|
161 |
+
sid,
|
162 |
+
audio0,
|
163 |
+
pitch,
|
164 |
+
pitchf,
|
165 |
+
times,
|
166 |
+
index,
|
167 |
+
big_npy,
|
168 |
+
index_rate,
|
169 |
+
version,
|
170 |
+
protect,
|
171 |
+
): # ,file_index,file_big_npy
|
172 |
+
feats = torch.from_numpy(audio0)
|
173 |
+
if self.is_half:
|
174 |
+
feats = feats.half()
|
175 |
+
else:
|
176 |
+
feats = feats.float()
|
177 |
+
if feats.dim() == 2: # double channels
|
178 |
+
feats = feats.mean(-1)
|
179 |
+
assert feats.dim() == 1, feats.dim()
|
180 |
+
feats = feats.view(1, -1)
|
181 |
+
padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
|
182 |
+
|
183 |
+
inputs = {
|
184 |
+
"source": feats.to(self.device),
|
185 |
+
"padding_mask": padding_mask,
|
186 |
+
"output_layer": 9 if version == "v1" else 12,
|
187 |
+
}
|
188 |
+
t0 = ttime()
|
189 |
+
with torch.no_grad():
|
190 |
+
logits = model.extract_features(**inputs)
|
191 |
+
feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
|
192 |
+
if protect < 0.5 and pitch != None and pitchf != None:
|
193 |
+
feats0 = feats.clone()
|
194 |
+
if (
|
195 |
+
isinstance(index, type(None)) == False
|
196 |
+
and isinstance(big_npy, type(None)) == False
|
197 |
+
and index_rate != 0
|
198 |
+
):
|
199 |
+
npy = feats[0].cpu().numpy()
|
200 |
+
if self.is_half:
|
201 |
+
npy = npy.astype("float32")
|
202 |
+
|
203 |
+
# _, I = index.search(npy, 1)
|
204 |
+
# npy = big_npy[I.squeeze()]
|
205 |
+
|
206 |
+
score, ix = index.search(npy, k=8)
|
207 |
+
weight = np.square(1 / score)
|
208 |
+
weight /= weight.sum(axis=1, keepdims=True)
|
209 |
+
npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
|
210 |
+
|
211 |
+
if self.is_half:
|
212 |
+
npy = npy.astype("float16")
|
213 |
+
feats = (
|
214 |
+
torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
|
215 |
+
+ (1 - index_rate) * feats
|
216 |
+
)
|
217 |
+
|
218 |
+
feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
|
219 |
+
if protect < 0.5 and pitch != None and pitchf != None:
|
220 |
+
feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(
|
221 |
+
0, 2, 1
|
222 |
+
)
|
223 |
+
t1 = ttime()
|
224 |
+
p_len = audio0.shape[0] // self.window
|
225 |
+
if feats.shape[1] < p_len:
|
226 |
+
p_len = feats.shape[1]
|
227 |
+
if pitch != None and pitchf != None:
|
228 |
+
pitch = pitch[:, :p_len]
|
229 |
+
pitchf = pitchf[:, :p_len]
|
230 |
+
|
231 |
+
if protect < 0.5 and pitch != None and pitchf != None:
|
232 |
+
pitchff = pitchf.clone()
|
233 |
+
pitchff[pitchf > 0] = 1
|
234 |
+
pitchff[pitchf < 1] = protect
|
235 |
+
pitchff = pitchff.unsqueeze(-1)
|
236 |
+
feats = feats * pitchff + feats0 * (1 - pitchff)
|
237 |
+
feats = feats.to(feats0.dtype)
|
238 |
+
p_len = torch.tensor([p_len], device=self.device).long()
|
239 |
+
with torch.no_grad():
|
240 |
+
if pitch != None and pitchf != None:
|
241 |
+
audio1 = (
|
242 |
+
(net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0])
|
243 |
+
.data.cpu()
|
244 |
+
.float()
|
245 |
+
.numpy()
|
246 |
+
)
|
247 |
+
else:
|
248 |
+
audio1 = (
|
249 |
+
(net_g.infer(feats, p_len, sid)[0][0, 0]).data.cpu().float().numpy()
|
250 |
+
)
|
251 |
+
del feats, p_len, padding_mask
|
252 |
+
if torch.cuda.is_available():
|
253 |
+
torch.cuda.empty_cache()
|
254 |
+
t2 = ttime()
|
255 |
+
times[0] += t1 - t0
|
256 |
+
times[2] += t2 - t1
|
257 |
+
return audio1
|
258 |
+
|
259 |
+
def pipeline(
|
260 |
+
self,
|
261 |
+
model,
|
262 |
+
net_g,
|
263 |
+
sid,
|
264 |
+
audio,
|
265 |
+
input_audio_path,
|
266 |
+
times,
|
267 |
+
f0_up_key,
|
268 |
+
f0_method,
|
269 |
+
file_index,
|
270 |
+
# file_big_npy,
|
271 |
+
index_rate,
|
272 |
+
if_f0,
|
273 |
+
filter_radius,
|
274 |
+
tgt_sr,
|
275 |
+
resample_sr,
|
276 |
+
rms_mix_rate,
|
277 |
+
version,
|
278 |
+
protect,
|
279 |
+
f0_file=None,
|
280 |
+
):
|
281 |
+
logging.info("pipeline starts")
|
282 |
+
if (
|
283 |
+
file_index != ""
|
284 |
+
# and file_big_npy != ""
|
285 |
+
# and os.path.exists(file_big_npy) == True
|
286 |
+
and os.path.exists(file_index) == True
|
287 |
+
and index_rate != 0
|
288 |
+
):
|
289 |
+
try:
|
290 |
+
index = faiss.read_index(file_index)
|
291 |
+
# big_npy = np.load(file_big_npy)
|
292 |
+
big_npy = index.reconstruct_n(0, index.ntotal)
|
293 |
+
except:
|
294 |
+
traceback.print_exc()
|
295 |
+
index = big_npy = None
|
296 |
+
else:
|
297 |
+
index = big_npy = None
|
298 |
+
logging.info("index: ")
|
299 |
+
audio = signal.filtfilt(bh, ah, audio)
|
300 |
+
audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
|
301 |
+
opt_ts = []
|
302 |
+
if audio_pad.shape[0] > self.t_max:
|
303 |
+
audio_sum = np.zeros_like(audio)
|
304 |
+
for i in range(self.window):
|
305 |
+
audio_sum += audio_pad[i : i - self.window]
|
306 |
+
for t in range(self.t_center, audio.shape[0], self.t_center):
|
307 |
+
opt_ts.append(
|
308 |
+
t
|
309 |
+
- self.t_query
|
310 |
+
+ np.where(
|
311 |
+
np.abs(audio_sum[t - self.t_query : t + self.t_query])
|
312 |
+
== np.abs(audio_sum[t - self.t_query : t + self.t_query]).min()
|
313 |
+
)[0][0]
|
314 |
+
)
|
315 |
+
logging.info("opt_ts: ")
|
316 |
+
s = 0
|
317 |
+
audio_opt = []
|
318 |
+
t = None
|
319 |
+
t1 = ttime()
|
320 |
+
audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
|
321 |
+
p_len = audio_pad.shape[0] // self.window
|
322 |
+
inp_f0 = None
|
323 |
+
if hasattr(f0_file, "name") == True:
|
324 |
+
try:
|
325 |
+
with open(f0_file.name, "r") as f:
|
326 |
+
lines = f.read().strip("\n").split("\n")
|
327 |
+
inp_f0 = []
|
328 |
+
for line in lines:
|
329 |
+
inp_f0.append([float(i) for i in line.split(",")])
|
330 |
+
inp_f0 = np.array(inp_f0, dtype="float32")
|
331 |
+
except:
|
332 |
+
traceback.print_exc()
|
333 |
+
sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
|
334 |
+
logging.info("sid: ")
|
335 |
+
pitch, pitchf = None, None
|
336 |
+
logging.info("if_f0: ")
|
337 |
+
if if_f0 == 1:
|
338 |
+
logging.info("inside if_f0: ")
|
339 |
+
pitch, pitchf = self.get_f0(
|
340 |
+
input_audio_path,
|
341 |
+
audio_pad,
|
342 |
+
p_len,
|
343 |
+
f0_up_key,
|
344 |
+
f0_method,
|
345 |
+
filter_radius,
|
346 |
+
inp_f0,
|
347 |
+
)
|
348 |
+
logging.info("after get_f0")
|
349 |
+
pitch = pitch[:p_len]
|
350 |
+
pitchf = pitchf[:p_len]
|
351 |
+
if self.device == "mps":
|
352 |
+
pitchf = pitchf.astype(np.float32)
|
353 |
+
logging.info("12345")
|
354 |
+
pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
|
355 |
+
pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
|
356 |
+
logging.info("6789")
|
357 |
+
logging.info("after if_f0: ")
|
358 |
+
t2 = ttime()
|
359 |
+
times[1] += t2 - t1
|
360 |
+
for t in opt_ts:
|
361 |
+
t = t // self.window * self.window
|
362 |
+
if if_f0 == 1:
|
363 |
+
audio_opt.append(
|
364 |
+
self.vc(
|
365 |
+
model,
|
366 |
+
net_g,
|
367 |
+
sid,
|
368 |
+
audio_pad[s : t + self.t_pad2 + self.window],
|
369 |
+
pitch[:, s // self.window : (t + self.t_pad2) // self.window],
|
370 |
+
pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
|
371 |
+
times,
|
372 |
+
index,
|
373 |
+
big_npy,
|
374 |
+
index_rate,
|
375 |
+
version,
|
376 |
+
protect,
|
377 |
+
)[self.t_pad_tgt : -self.t_pad_tgt]
|
378 |
+
)
|
379 |
+
else:
|
380 |
+
audio_opt.append(
|
381 |
+
self.vc(
|
382 |
+
model,
|
383 |
+
net_g,
|
384 |
+
sid,
|
385 |
+
audio_pad[s : t + self.t_pad2 + self.window],
|
386 |
+
None,
|
387 |
+
None,
|
388 |
+
times,
|
389 |
+
index,
|
390 |
+
big_npy,
|
391 |
+
index_rate,
|
392 |
+
version,
|
393 |
+
protect,
|
394 |
+
)[self.t_pad_tgt : -self.t_pad_tgt]
|
395 |
+
)
|
396 |
+
s = t
|
397 |
+
|
398 |
+
if if_f0 == 1:
|
399 |
+
audio_opt.append(
|
400 |
+
self.vc(
|
401 |
+
model,
|
402 |
+
net_g,
|
403 |
+
sid,
|
404 |
+
audio_pad[t:],
|
405 |
+
pitch[:, t // self.window :] if t is not None else pitch,
|
406 |
+
pitchf[:, t // self.window :] if t is not None else pitchf,
|
407 |
+
times,
|
408 |
+
index,
|
409 |
+
big_npy,
|
410 |
+
index_rate,
|
411 |
+
version,
|
412 |
+
protect,
|
413 |
+
)[self.t_pad_tgt : -self.t_pad_tgt]
|
414 |
+
)
|
415 |
+
else:
|
416 |
+
audio_opt.append(
|
417 |
+
self.vc(
|
418 |
+
model,
|
419 |
+
net_g,
|
420 |
+
sid,
|
421 |
+
audio_pad[t:],
|
422 |
+
None,
|
423 |
+
None,
|
424 |
+
times,
|
425 |
+
index,
|
426 |
+
big_npy,
|
427 |
+
index_rate,
|
428 |
+
version,
|
429 |
+
protect,
|
430 |
+
)[self.t_pad_tgt : -self.t_pad_tgt]
|
431 |
+
)
|
432 |
+
audio_opt = np.concatenate(audio_opt)
|
433 |
+
|
434 |
+
|
435 |
+
if rms_mix_rate != 1:
|
436 |
+
audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
|
437 |
+
if resample_sr >= 16000 and tgt_sr != resample_sr:
|
438 |
+
audio_opt = librosa.resample(
|
439 |
+
audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
|
440 |
+
)
|
441 |
+
audio_max = np.abs(audio_opt).max() / 0.99
|
442 |
+
max_int16 = 32768
|
443 |
+
if audio_max > 1:
|
444 |
+
max_int16 /= audio_max
|
445 |
+
audio_opt = (audio_opt * max_int16).astype(np.int16)
|
446 |
+
del pitch, pitchf, sid
|
447 |
+
if torch.cuda.is_available():
|
448 |
+
torch.cuda.empty_cache()
|
449 |
+
return audio_opt
|