oaax

Python bindings for runtime libraries that implement the OAAX runtime interface (OAAX 2.0). The package loads any compliant runtime library with ctypes and lets you drive it with NumPy arrays: initialise, load models, enqueue inputs, retrieve outputs, and clean up. It is not tied to any particular vendor's runtime or model format.

Installation

pip install oaax

You need an OAAX runtime shared library (.dll on Windows, .so on Linux) and a model file that runtime can load. The package doesn't ship either.

A generic CPU runtime, built on ONNX Runtime, is available as an archive for each platform:

• Windows x86_64: https://oaax.nbg1.your-objectstorage.com/runtimes/2.0.0/CPU/x86_64/Windows/library.tar.gz
• Linux x86_64: https://oaax.nbg1.your-objectstorage.com/runtimes/2.0.0/CPU/x86_64/Ubuntu/library.tar.gz
• Linux aarch64: https://oaax.nbg1.your-objectstorage.com/runtimes/2.0.0/CPU/aarch64/Ubuntu/library.tar.gz

Each archive contains the runtime library and the libraries it depends on. Unpack it and pass the library path to Runtime(...).

To try it with a real model, the examples use UltraFace, a face detection model (MIT licensed) available from https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB.

Usage

import numpy as np
from oaax import Runtime

with Runtime("path/to/RuntimeLibrary.dll") as rt:
    print(rt.name, rt.version)
    rt.init({})  # runtime-specific options, if any

    model_id = rt.load_model("path/to/model.onnx")
    io = rt.model_io(model_id)  # None if the runtime doesn't report its inputs and outputs
    if io is not None:
        inputs, outputs = io
        print(inputs)

    image = np.random.rand(1, 3, 240, 320).astype(np.float32)
    result = rt.run(model_id, {"input": image}, timeout_ms=10_000)  # the model's real input name
    for tensor in result:
        print(tensor.name, tensor.data.shape, tensor.data.dtype)

Use the input names and shapes your model expects. If you pass an input name the model doesn't have, run() raises ValueError when the runtime reports its inputs; otherwise the request may time out.

For pipelined use, rt.enqueue(model_id, inputs) submits a request without waiting, and rt.retrieve(timeout_ms) returns whichever request finishes first.

Configuration keys passed to rt.init(...) and per-model options are defined by each runtime vendor, not by the OAAX interface. A compliant runtime ignores keys it doesn't recognise.

Documentation and source

• Full API notes, memory ownership rules and the test suite: README
• Examples with sample models: examples/
• Contributing: CONTRIBUTING.md
• Changes between versions: CHANGELOG.md
• Issues: github.com/OAAX-standard/PythonBindings/issues

License

Apache License 2.0. See LICENSE.