Spaces:
Sleeping
Sleeping
Add application file
Browse files- .gitignore +2 -0
- app.py +95 -0
- requirements.txt +11 -0
.gitignore
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
*.ckpt
|
2 |
+
*__
|
app.py
ADDED
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import librosa
|
3 |
+
import torch
|
4 |
+
import soundfile as sf
|
5 |
+
from speechbrain.inference.separation import SepformerSeparation as separator
|
6 |
+
import torchaudio, torchmetrics, torch
|
7 |
+
|
8 |
+
|
9 |
+
# defineing model class
|
10 |
+
class SepformerFineTune(torch.nn.Module):
|
11 |
+
def __init__(self, model):
|
12 |
+
super(SepformerFineTune, self).__init__()
|
13 |
+
self.model = model
|
14 |
+
# disabling gradient computation
|
15 |
+
for parms in self.model.parameters():
|
16 |
+
parms.requires_grad = False
|
17 |
+
|
18 |
+
# enable gradient computation for the last layer
|
19 |
+
named_layers = dict(model.named_modules())
|
20 |
+
for name, layer in named_layers.items():
|
21 |
+
# print(f"Name: {name}, Layer: {layer}")
|
22 |
+
if name == "mods.masknet.output.0":
|
23 |
+
for param in layer.parameters():
|
24 |
+
param.requires_grad = True
|
25 |
+
if name == "mods.masknet.output_gate":
|
26 |
+
for param in layer.parameters():
|
27 |
+
param.requires_grad = True
|
28 |
+
|
29 |
+
|
30 |
+
# printing all tranble parameters
|
31 |
+
# for model_name, model_params in model.named_parameters():
|
32 |
+
# print(f"Model Layer Name: {model_name}, Model Params: {model_params.requires_grad}")
|
33 |
+
def forward(self, mix):
|
34 |
+
est_sources = self.model.separate_batch(mix)
|
35 |
+
return est_sources[:,:,0], est_sources[:,:,1] # NOTE: Working with 2 sources ONLY
|
36 |
+
|
37 |
+
|
38 |
+
class SourceSeparationApp:
|
39 |
+
def __init__(self, model_path,device="cpu"):
|
40 |
+
self.model = self.load_model(model_path)
|
41 |
+
self.device = device
|
42 |
+
|
43 |
+
def load_model(self, model_path):
|
44 |
+
model = separator.from_hparams(source="speechbrain/sepformer-wsj03mix", savedir='pretrained_models/sepformer-wsj03mix', run_opts={"device": device})
|
45 |
+
checkpoint = torch.load(model_path)
|
46 |
+
fine_tuned_model = SepformerFineTune(model)
|
47 |
+
fine_tuned_model.load_state_dict(checkpoint["model"])
|
48 |
+
return fine_tuned_model
|
49 |
+
|
50 |
+
def separate_sources(self, audio_file):
|
51 |
+
# Load input audio
|
52 |
+
# print(f"[LOG] Audio file: {audio_file}")
|
53 |
+
input_audio_tensor, sr = audio_file[1], audio_file[0]
|
54 |
+
|
55 |
+
if self.model is None:
|
56 |
+
return "Error: Model not loaded."
|
57 |
+
|
58 |
+
# sending input audio to PyTorch tensor
|
59 |
+
input_audio_tensor = torch.tensor(input_audio_tensor,dtype=torch.float).unsqueeze(0)
|
60 |
+
input_audio_tensor = input_audio_tensor.to(self.device)
|
61 |
+
|
62 |
+
# Source separation using the loaded model
|
63 |
+
self.model.to(self.device)
|
64 |
+
self.model.eval()
|
65 |
+
with torch.inference_mode():
|
66 |
+
# print(f"[LOG] mix shape: {mix.shape}, s1 shape: {s1.shape}, s2 shape: {s2.shape}, noise shape: {noise.shape}")
|
67 |
+
source1,source2 = self.model(input_audio_tensor)
|
68 |
+
|
69 |
+
|
70 |
+
# Save separated sources
|
71 |
+
sf.write("source1.wav", source1.squeeze().cpu().numpy(), sr)
|
72 |
+
sf.write("source2.wav", source2.squeeze().cpu().numpy(), sr)
|
73 |
+
|
74 |
+
return "Separation completed", "source1.wav", "source2.wav"
|
75 |
+
|
76 |
+
def run(self):
|
77 |
+
audio_input = gr.Audio(label="Upload or record audio")
|
78 |
+
output_text = gr.Label(label="Status:")
|
79 |
+
audio_output1 = gr.Audio(label="Source 1", type="filepath",)
|
80 |
+
audio_output2 = gr.Audio(label="Source 2", type="filepath",)
|
81 |
+
gr.Interface(
|
82 |
+
fn=self.separate_sources,
|
83 |
+
inputs=audio_input,
|
84 |
+
outputs=[output_text, audio_output1, audio_output2],
|
85 |
+
title="Audio Source Separation",
|
86 |
+
description="Separate sources from a mixed audio signal.",
|
87 |
+
allow_flagging=False
|
88 |
+
).launch()
|
89 |
+
|
90 |
+
|
91 |
+
if __name__ == "__main__":
|
92 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
93 |
+
model_path = "fine_tuned_sepformer-wsj03mix-7sec.ckpt" # Replace with your model path
|
94 |
+
app = SourceSeparationApp(model_path, device=device)
|
95 |
+
app.run()
|
requirements.txt
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
soundfile>=0.10.3.post1
|
2 |
+
tqdm>=4.46.1
|
3 |
+
pysndfx>=0.3.6
|
4 |
+
pandas>=1.0.1
|
5 |
+
numpy>=1.18.1
|
6 |
+
pyloudnorm>=0.1.0
|
7 |
+
scipy>=1.4.1
|
8 |
+
matplotlib>=3.1.3
|
9 |
+
torch==2.2.1
|
10 |
+
torchaudio==2.2.1
|
11 |
+
speechbrain
|