Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,82 +1,82 @@
|
|
1 |
-
|
2 |
-
|
3 |
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
|
10 |
-
|
11 |
-
|
12 |
|
13 |
-
CHECKPOINTS_DIR =
|
14 |
|
15 |
-
|
16 |
|
17 |
-
#
|
18 |
MODEL_NAME = CHECKPOINTS_DIR / "audiosep_base_4M_steps.ckpt"
|
19 |
-
|
20 |
config_yaml="config/audiosep_base.yaml",
|
21 |
-
checkpoint_path=
|
22 |
-
|
23 |
)
|
24 |
|
25 |
|
26 |
-
|
27 |
-
# AudioSep:
|
28 |
-
[[
|
29 |
|
30 |
-
AudioSep
|
31 |
-
AudioSep
|
32 |
-
|
33 |
"""
|
34 |
|
35 |
|
36 |
-
def inference(
|
37 |
-
print(f"
|
38 |
-
|
39 |
|
40 |
-
|
41 |
-
|
42 |
|
43 |
-
|
44 |
-
|
45 |
)
|
46 |
|
47 |
input_dict = {
|
48 |
-
"
|
49 |
-
"
|
50 |
}
|
51 |
|
52 |
-
sep_segment =
|
53 |
|
54 |
sep_segment = sep_segment.squeeze(0).squeeze(0).data.cpu().numpy()
|
55 |
|
56 |
-
|
57 |
|
58 |
|
59 |
-
|
60 |
-
gr.Markdown(
|
61 |
-
|
62 |
-
|
63 |
input_audio = gr.Audio(label="Mixture", type="filepath")
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
output_audio = gr.Audio(label="
|
68 |
-
|
69 |
-
"
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
)
|
75 |
-
|
76 |
-
fn=
|
77 |
)
|
78 |
|
79 |
-
gr.Markdown("##
|
80 |
-
gr.Examples(
|
81 |
|
82 |
-
demo.queue().launch(
|
|
|
1 |
+
from pathlib import Path
|
2 |
+
from threading import Thread
|
3 |
|
4 |
+
import gdown
|
5 |
+
import gradio as gr
|
6 |
+
import librosa
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
|
10 |
+
from gradio_examples import EXAMPLES
|
11 |
+
from pipeline import build_audiosep
|
12 |
|
13 |
+
CHECKPOINTS_DIR = Path("checkpoint")
|
14 |
|
15 |
+
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
16 |
|
17 |
+
# The model will be loaded in the future
|
18 |
MODEL_NAME = CHECKPOINTS_DIR / "audiosep_base_4M_steps.ckpt"
|
19 |
+
MODEL = build_audiosep(
|
20 |
config_yaml="config/audiosep_base.yaml",
|
21 |
+
checkpoint_path=MODEL_NAME,
|
22 |
+
device=DEVICE,
|
23 |
)
|
24 |
|
25 |
|
26 |
+
description = """
|
27 |
+
# AudioSep: Separate Anything You Describe
|
28 |
+
[[Project Page]](https://audio-agi.github.io/Separate-Anything-You-Describe) [[Paper]](https://audio-agi.github.io/Separate-Anything-You-Describe/AudioSep_arXiv.pdf) [[Code]](https://github.com/Audio-AGI/AudioSep)
|
29 |
|
30 |
+
AudioSep is a foundation model for open-domain sound separation with natural language queries.
|
31 |
+
AudioSep demonstrates strong separation performance and impressivezero-shot generalization ability on
|
32 |
+
numerous tasks such as audio event separation, musical instrument separation, and speech enhancement.
|
33 |
"""
|
34 |
|
35 |
|
36 |
+
def inference(audio_file_path: str, text: str):
|
37 |
+
print(f"Separate audio from [{audio_file_path}] with textual query [{text}]")
|
38 |
+
mixture, _ = librosa.load(audio_file_path, sr=32000, mono=True)
|
39 |
|
40 |
+
with torch.no_grad():
|
41 |
+
text = [text]
|
42 |
|
43 |
+
conditions = MODEL.query_encoder.get_query_embed(
|
44 |
+
modality="text", text=text, device=DEVICE
|
45 |
)
|
46 |
|
47 |
input_dict = {
|
48 |
+
"mixture": torch.Tensor(mixture)[None, None, :].to(DEVICE),
|
49 |
+
"condition": conditions,
|
50 |
}
|
51 |
|
52 |
+
sep_segment = MODEL.ss_model(input_dict)["waveform"]
|
53 |
|
54 |
sep_segment = sep_segment.squeeze(0).squeeze(0).data.cpu().numpy()
|
55 |
|
56 |
+
return 32000, np.round(sep_segment * 32767).astype(np.int16)
|
57 |
|
58 |
|
59 |
+
with gr.Blocks(title="AudioSep") as demo:
|
60 |
+
gr.Markdown(description)
|
61 |
+
with gr.Row():
|
62 |
+
with gr.Column():
|
63 |
input_audio = gr.Audio(label="Mixture", type="filepath")
|
64 |
+
text = gr.Textbox(label="Text Query")
|
65 |
+
with gr.Column():
|
66 |
+
with gr.Column():
|
67 |
+
output_audio = gr.Audio(label="Separation Result", scale=10)
|
68 |
+
button = gr.Button(
|
69 |
+
"Separate",
|
70 |
+
variant="primary",
|
71 |
+
scale=2,
|
72 |
+
size="lg",
|
73 |
+
interactive=True,
|
74 |
)
|
75 |
+
button.click(
|
76 |
+
fn=inference, inputs=[input_audio, text], outputs=[output_audio]
|
77 |
)
|
78 |
|
79 |
+
gr.Markdown("## Examples")
|
80 |
+
gr.Examples(examples=EXAMPLES, inputs=[input_audio, text])
|
81 |
|
82 |
+
demo.queue().launch(share=True)
|