├── .gitignore
├── Dockerfile
├── README.md
├── app.py
├── models
    └── README.md
└── requirements.txt


/.gitignore:
--------------------------------------------------------------------------------
1 | .vscode
2 | *.onnx
3 | 


--------------------------------------------------------------------------------
/Dockerfile:
--------------------------------------------------------------------------------
 1 | FROM nvidia/cuda:11.6.0-cudnn8-runtime-ubuntu20.04
 2 | 
 3 | # Install Python 3.10
 4 | RUN apt-get update && \
 5 |     apt-get install -y software-properties-common && \
 6 |     add-apt-repository ppa:deadsnakes/ppa && \
 7 |     apt-get update && \
 8 |     apt-get install -y python3.10 python3.10-dev python3.10-distutils curl && \
 9 |     apt-get clean && \
10 |     rm -rf /var/lib/apt/lists/*
11 | 
12 | # Install pip
13 | RUN curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py && \
14 |     python3.10 get-pip.py && \
15 |     rm get-pip.py
16 | 
17 | # Install Python packages
18 | WORKDIR /app
19 | COPY requirements.txt /app/
20 | 
21 | RUN pip install --no-cache-dir -r requirements.txt
22 | EXPOSE 6873
23 | COPY app.py models/ /app/
24 | 
25 | CMD python3.10 app.py
26 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # OpenUTAU remote ONNX hosting
 2 | 
 3 | This repository hosts the ONNX models for DiffSinger OpenUTAU. 
 4 | You should put *.onnx files in the `models` directory, after you start the server, you will see their md5 hash in the console.
 5 | 
 6 | ## Usage
 7 | Set your `acoustic` in `dsconfig.yaml` to `http://${ADDRESS}/v1/${md5}/inference`.
 8 | 
 9 | To start the server, run `python3 app.py` in the root directory of this repository.  
10 | There is also a docker image for this server, but it is not tested yet.
11 | 


--------------------------------------------------------------------------------
/app.py:
--------------------------------------------------------------------------------
 1 | from flask import Flask, request
 2 | from pathlib import Path
 3 | from hashlib import md5
 4 | import onnxruntime as ort
 5 | from loguru import logger
 6 | import numpy as np
 7 | 
 8 | app = Flask(__name__)
 9 | all_models = {}
10 | 
11 | for model_path in Path('models').glob('*.onnx'):
12 |     name = md5(model_path.read_bytes()).hexdigest()
13 |     all_models[name] = ort.InferenceSession(str(model_path), providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
14 | 
15 |     logger.info(f'Loaded model {model_path} as {name}')
16 | 
17 | @app.route('/v1/<model_name>/inference', methods=['POST'])
18 | def inference(model_name):
19 |     if model_name not in all_models:
20 |         return dict(error=f'Model {model_name} not found'), 404
21 | 
22 |     model = all_models[model_name]
23 |     payload = request.get_json(silent=True)
24 |     if payload is None:
25 |         return dict(error='Invalid payload'), 400
26 | 
27 |     inputs = {}
28 |     for model_input in model.get_inputs():
29 |         name = model_input.name
30 | 
31 |         if name not in payload:
32 |             return dict(error=f'Input {name} not found'), 400
33 | 
34 |         input = payload[name]
35 |         if 'data' not in input or 'shape' not in input:
36 |             return dict(error=f'Input {name} is invalid'), 400
37 |         
38 |         np_type = None
39 |         if model_input.type == 'tensor(float)':
40 |             np_type = np.float32
41 |         elif model_input.type == 'tensor(int64)':
42 |             np_type = np.int64
43 |         else:
44 |             return dict(error=f'Input {name} has unsupported type {model_input.type}'), 400
45 | 
46 |         data = np.array(
47 |             input['data'], 
48 |             dtype=np_type
49 |         ).reshape(input['shape'])
50 | 
51 |         inputs[name] = data
52 | 
53 |     outputs = model.run(None, inputs)
54 | 
55 |     return dict(
56 |         type="float32",
57 |         shape=outputs[0].shape,
58 |         data=outputs[0].flatten().tolist()
59 |     )
60 | 
61 | 
62 | if __name__ == '__main__':
63 |     app.run(host='0.0.0.0', port=6873, debug=True, threaded=True)
64 | 


--------------------------------------------------------------------------------
/models/README.md:
--------------------------------------------------------------------------------
1 | PUT YOUR MODELS HERE
2 | 


--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | flask 
2 | onnxruntime-gpu
3 | loguru
4 | 


--------------------------------------------------------------------------------