Spaces:
Running
on
L40S
Running
on
L40S
hainazhu
commited on
Commit
·
c4aaa82
1
Parent(s):
208580f
add separator.py
Browse files- separator.py +50 -0
separator.py
ADDED
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torchaudio
|
2 |
+
import os
|
3 |
+
import torch
|
4 |
+
from third_party.demucs.models.pretrained import get_model_from_yaml
|
5 |
+
|
6 |
+
|
7 |
+
class Separator(torch.nn.Module):
|
8 |
+
def __init__(self, dm_model_path='third_party/demucs/ckpt/htdemucs.pth', dm_config_path='third_party/demucs/ckpt/htdemucs.yaml', gpu_id=0) -> None:
|
9 |
+
super().__init__()
|
10 |
+
if torch.cuda.is_available() and gpu_id < torch.cuda.device_count():
|
11 |
+
self.device = torch.device(f"cuda:{gpu_id}")
|
12 |
+
else:
|
13 |
+
self.device = torch.device("cpu")
|
14 |
+
self.demucs_model = self.init_demucs_model(dm_model_path, dm_config_path)
|
15 |
+
|
16 |
+
def init_demucs_model(self, model_path, config_path):
|
17 |
+
model = get_model_from_yaml(config_path, model_path)
|
18 |
+
model.to(self.device)
|
19 |
+
model.eval()
|
20 |
+
return model
|
21 |
+
|
22 |
+
def load_audio(self, f):
|
23 |
+
a, fs = torchaudio.load(f)
|
24 |
+
if (fs != 48000):
|
25 |
+
a = torchaudio.functional.resample(a, fs, 48000)
|
26 |
+
if a.shape[-1] >= 48000*10:
|
27 |
+
a = a[..., :48000*10]
|
28 |
+
else:
|
29 |
+
a = torch.cat([a, a], -1)
|
30 |
+
return a[:, 0:48000*10]
|
31 |
+
|
32 |
+
def run(self, audio_path, output_dir='tmp', ext=".flac"):
|
33 |
+
os.makedirs(output_dir, exist_ok=True)
|
34 |
+
name, _ = os.path.splitext(os.path.split(audio_path)[-1])
|
35 |
+
output_paths = []
|
36 |
+
|
37 |
+
for stem in self.demucs_model.sources:
|
38 |
+
output_path = os.path.join(output_dir, f"{name}_{stem}{ext}")
|
39 |
+
if os.path.exists(output_path):
|
40 |
+
output_paths.append(output_path)
|
41 |
+
if len(output_paths) == 1: # 4
|
42 |
+
vocal_path = output_paths[0]
|
43 |
+
else:
|
44 |
+
drums_path, bass_path, other_path, vocal_path = self.demucs_model.separate(audio_path, output_dir, device=self.device)
|
45 |
+
for path in [drums_path, bass_path, other_path]:
|
46 |
+
os.remove(path)
|
47 |
+
full_audio = self.load_audio(audio_path)
|
48 |
+
vocal_audio = self.load_audio(vocal_path)
|
49 |
+
bgm_audio = full_audio - vocal_audio
|
50 |
+
return full_audio, vocal_audio, bgm_audio
|