onnion project

• compile onnx to python
• runtime depends only numpy

See supported operators.

Tutorial

Extract the post-process of Ultraface and run it.

Install tools:

$ pip3 install onnigiri onnion onnion-rt

Download the onnx file:

$ wget -O ultraface.onnx 'https://github.com/onnx/models/raw/main/vision/body_analysis/ultraface/models/version-RFB-640.onnx'

Extract the post-process:

$ onnigiri ultraface.onnx -o ultraface-post.onnx --from 460 --to boxes

Compile onnx to python:

$ onnion ultraface-post.onnx -o ultraface_post.py

Run and check:

$ python check_model.py
check
pass

check_model.py

import onnxruntime
import numpy as np

from ultraface_post import init_graph

if __name__ == "__main__":
    x = np.random.randn(1,17640, 4).astype(np.float32)
    sess = onnxruntime.InferenceSession('ultraface-post.onnx')
    expeced = sess.run(['boxes'], {'460': x})

    graph = init_graph()
    y = graph.run(x)

    for a,b in zip(expeced, y):
        print("check")
        assert np.all(abs(a-b) < 1e-4)

    print("pass")

See also examples.

Development Guide

See each subdirectory:

• onnion
• onnion-rt

Related project

• onnigiri