Skip to content

Commit 921f0f4

Browse files
csukuangfjcsukuangfj
authored
Add UVR models for source separation. (#2266)
1 parent 93e2819 commit 921f0f4

7 files changed

Lines changed: 575 additions & 3 deletions

File tree

‎.github/workflows/as_cmake_sub_project.yaml‎

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -4,9 +4,6 @@ on:
44
push:
55
branches:
66
- master
7-
pull_request:
8-
branches:
9-
- master
107

118
workflow_dispatch:
129

‎.github/workflows/export-uvr-to-onnx.yaml‎

Lines changed: 98 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,98 @@
1+
name: export-uvr-to-onnx
2+
3+
on:
4+
push:
5+
branches:
6+
- uvr
7+
workflow_dispatch:
8+
9+
concurrency:
10+
group: export-uvr-to-onnx-${{ github.ref }}
11+
cancel-in-progress: true
12+
13+
jobs:
14+
export-uvr-to-onnx:
15+
if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
16+
name: export UVR to ONNX
17+
runs-on: ${{ matrix.os }}
18+
strategy:
19+
fail-fast: false
20+
matrix:
21+
os: [macos-latest]
22+
python-version: ["3.10"]
23+
24+
steps:
25+
- uses: actions/checkout@v4
26+
27+
- name: Setup Python ${{ matrix.python-version }}
28+
uses: actions/setup-python@v5
29+
with:
30+
python-version: ${{ matrix.python-version }}
31+
32+
- name: Install dependencies
33+
shell: bash
34+
run: |
35+
pip install "numpy<2" onnx==1.17.0 onnxruntime==1.17.1 onnxmltools kaldi-native-fbank librosa soundfile
36+
37+
- name: Run
38+
shell: bash
39+
run: |
40+
cd scripts/uvr_mdx
41+
curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/source-separation-models/audio_example.wav
42+
ls -lh audio_example.wav
43+
./run.sh
44+
45+
- name: Collect mp3 files
46+
shell: bash
47+
run: |
48+
mv -v scripts/uvr_mdx/*.mp3 ./
49+
ls -lh *.mp3
50+
51+
- uses: actions/upload-artifact@v4
52+
with:
53+
name: generated-mp3
54+
path: ./*.mp3
55+
56+
- name: Collect models
57+
shell: bash
58+
run: |
59+
mv -v scripts/uvr_mdx/models/*.onnx ./
60+
ls -lh *.onnx
61+
62+
- name: Release
63+
uses: svenstaro/upload-release-action@v2
64+
with:
65+
file_glob: true
66+
file: ./*.onnx
67+
overwrite: true
68+
repo_name: k2-fsa/sherpa-onnx
69+
repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
70+
tag: source-separation-models
71+
72+
- name: Publish to huggingface
73+
env:
74+
HF_TOKEN: ${{ secrets.HF_TOKEN }}
75+
uses: nick-fields/retry@v3
76+
with:
77+
max_attempts: 20
78+
timeout_seconds: 200
79+
shell: bash
80+
command: |
81+
git config --global user.email "csukuangfj@gmail.com"
82+
git config --global user.name "Fangjun Kuang"
83+
84+
export GIT_LFS_SKIP_SMUDGE=1
85+
export GIT_CLONE_PROTECTION_ACTIVE=false
86+
87+
rm -rf huggingface
88+
git clone https://huggingface.co/k2-fsa/sherpa-onnx-models huggingface
89+
cd huggingface
90+
mkdir -p source-separation-models
91+
cp -av ../*.onnx ./source-separation-models
92+
git lfs track "*.onnx"
93+
git status
94+
git add .
95+
ls -lh
96+
git status
97+
git commit -m "add source separation models"
98+
git push https://csukuangfj:$HF_TOKEN@huggingface.co/k2-fsa/sherpa-onnx-models main

‎scripts/uvr_mdx/READEME.md‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,5 @@
1+
# Introduction
2+
3+
This folder contains scripts for converting models from
4+
https://github.com/TRvlvr/model_repo/releases/tag/all_public_uvr_models
5+
to sherpa-onnx.

‎scripts/uvr_mdx/add_meta_data_and_quantize.py‎

Lines changed: 118 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,118 @@
1+
#!/usr/bin/env python3
2+
# Copyright 2025 Xiaomi Corp. (authors: Fangjun Kuang)
3+
4+
import argparse
5+
from pathlib import Path
6+
7+
import onnx
8+
import onnxmltools
9+
import onnxruntime
10+
from onnxmltools.utils.float16_converter import convert_float_to_float16
11+
from onnxruntime.quantization import QuantType, quantize_dynamic
12+
13+
14+
def get_args():
15+
parser = argparse.ArgumentParser(
16+
formatter_class=argparse.ArgumentDefaultsHelpFormatter
17+
)
18+
19+
parser.add_argument(
20+
"--filename",
21+
type=str,
22+
required=True,
23+
help="Path to onnx model",
24+
)
25+
26+
return parser.parse_args()
27+
28+
29+
def export_onnx_fp16(onnx_fp32_path, onnx_fp16_path):
30+
onnx_fp32_model = onnxmltools.utils.load_model(onnx_fp32_path)
31+
onnx_fp16_model = convert_float_to_float16(onnx_fp32_model, keep_io_types=True)
32+
onnxmltools.utils.save_model(onnx_fp16_model, onnx_fp16_path)
33+
34+
35+
def validate(model: onnxruntime.InferenceSession):
36+
for i in model.get_inputs():
37+
print(i)
38+
39+
print("-----")
40+
41+
for i in model.get_outputs():
42+
print(i)
43+
44+
assert len(model.get_inputs()) == 1, len(model.get_inputs())
45+
assert len(model.get_outputs()) == 1, len(model.get_outputs())
46+
47+
inp = model.get_inputs()[0]
48+
outp = model.get_outputs()[0]
49+
50+
assert len(inp.shape) == 4, inp.shape
51+
assert len(outp.shape) == 4, outp.shape
52+
53+
assert inp.shape[1:] == outp.shape[1:], (inp.shape, outp.shape)
54+
55+
56+
def add_meta_data(filename, meta_data):
57+
model = onnx.load(filename)
58+
59+
print(model.metadata_props)
60+
61+
while len(model.metadata_props):
62+
model.metadata_props.pop()
63+
64+
for key, value in meta_data.items():
65+
meta = model.metadata_props.add()
66+
meta.key = key
67+
meta.value = str(value)
68+
print("--------------------")
69+
70+
print(model.metadata_props)
71+
72+
onnx.save(model, filename)
73+
74+
75+
def main():
76+
args = get_args()
77+
filename = Path(args.filename)
78+
if not filename.is_file():
79+
raise ValueError(f"{filename} does not exist")
80+
81+
name = filename.stem
82+
print("name", name)
83+
84+
model = onnx.load(str(filename))
85+
86+
session_opts = onnxruntime.SessionOptions()
87+
session_opts.log_severity_level = 3
88+
sess = onnxruntime.InferenceSession(
89+
str(filename), session_opts, providers=["CPUExecutionProvider"]
90+
)
91+
validate(sess)
92+
93+
inp = sess.get_inputs()[0]
94+
outp = sess.get_outputs()[0]
95+
96+
meta_data = {
97+
"model_type": "UVR",
98+
"model_name": name,
99+
"sample_rate": 44100,
100+
"comment": "This model is downloaded from https://github.com/TRvlvr/model_repo/releases",
101+
"n_fft": inp.shape[2] * 2,
102+
"center": 1,
103+
"window_type": "hann",
104+
"win_length": inp.shape[2] * 2,
105+
"hop_length": 1024,
106+
"dim_t": inp.shape[3],
107+
"dim_f": inp.shape[2],
108+
"dim_c": inp.shape[1],
109+
"stems": 2,
110+
}
111+
add_meta_data(str(filename), meta_data)
112+
113+
filename_fp16 = f"./{name}.fp16.onnx"
114+
export_onnx_fp16(filename, filename_fp16)
115+
116+
117+
if __name__ == "__main__":
118+
main()

‎scripts/uvr_mdx/run.sh‎

Lines changed: 54 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,54 @@
1+
#!/usr/bin/env bash
2+
# Copyright 2025 Xiaomi Corp. (authors: Fangjun Kuang)
3+
4+
set -ex
5+
6+
7+
# Please see https://github.com/TRvlvr/model_repo/releases/tag/all_public_uvr_models
8+
models=(
9+
UVR-MDX-NET-Inst_1.onnx
10+
UVR-MDX-NET-Inst_2.onnx
11+
UVR-MDX-NET-Inst_3.onnx
12+
UVR-MDX-NET-Inst_HQ_1.onnx
13+
UVR-MDX-NET-Inst_HQ_2.onnx
14+
UVR-MDX-NET-Inst_HQ_3.onnx
15+
UVR-MDX-NET-Inst_HQ_4.onnx
16+
UVR-MDX-NET-Inst_HQ_5.onnx
17+
UVR-MDX-NET-Inst_Main.onnx
18+
UVR-MDX-NET-Voc_FT.onnx
19+
UVR-MDX-NET_Crowd_HQ_1.onnx
20+
UVR_MDXNET_1_9703.onnx
21+
UVR_MDXNET_2_9682.onnx
22+
UVR_MDXNET_3_9662.onnx
23+
UVR_MDXNET_9482.onnx
24+
UVR_MDXNET_KARA.onnx
25+
UVR_MDXNET_KARA_2.onnx
26+
UVR_MDXNET_Main.onnx
27+
)
28+
29+
mkdir -p models
30+
for m in ${models[@]}; do
31+
if [ ! -f models/$m ]; then
32+
curl -SL --output models/$m https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/$m
33+
fi
34+
done
35+
36+
ls -lh models
37+
38+
for m in ${models[@]}; do
39+
echo "----------$m----------"
40+
python3 ./add_meta_data_and_quantize.py --filename models/$m
41+
42+
ls -lh models/
43+
done
44+
45+
if [ -f ./audio_example.wav ]; then
46+
for m in ${models[@]}; do
47+
./test.py --model-filename ./models/$m --audio-filename ./audio_example.wav
48+
name=$(basename -s .onnx $m)
49+
mv -v vocals.mp3 ${name}_vocals.mp3
50+
mv -v non_vocals.mp3 ${name}_non_vocals.mp3
51+
done
52+
53+
ls -lh *.mp3
54+
fi

‎scripts/uvr_mdx/show.py‎

Lines changed: 39 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,39 @@
1+
#!/usr/bin/env python3
2+
# Copyright 2025 Xiaomi Corp. (authors: Fangjun Kuang)
3+
4+
import onnxruntime
5+
import onnx
6+
7+
"""
8+
[]
9+
NodeArg(name='input', type='tensor(float)', shape=['batch_size', 4, 3072, 256])
10+
-----
11+
NodeArg(name='output', type='tensor(float)', shape=['batch_size', 4, 3072, 256])
12+
"""
13+
14+
15+
def show(filename):
16+
model = onnx.load(filename)
17+
print(model.metadata_props)
18+
19+
session_opts = onnxruntime.SessionOptions()
20+
session_opts.log_severity_level = 3
21+
sess = onnxruntime.InferenceSession(
22+
filename, session_opts, providers=["CPUExecutionProvider"]
23+
)
24+
for i in sess.get_inputs():
25+
print(i)
26+
27+
print("-----")
28+
29+
for i in sess.get_outputs():
30+
print(i)
31+
32+
33+
def main():
34+
# show("./UVR-MDX-NET-Voc_FT.onnx")
35+
show("./UVR_MDXNET_1_9703.onnx")
36+
37+
38+
if __name__ == "__main__":
39+
main()

0 commit comments

Comments
 (0)