A canonical template for making ONNX inference objects with
Avendish: a single C++ object that runs a .onnx
model with onnxruntime and is compiled, without rewrite, to
the host formats that have a real object model.
It is the ML counterpart of the audio, video, geometry and data templates.
avnd_addon_object(... CATEGORY object) instantiates the object back-ends — not the audio plug-in
formats (VST3 / CLAP):
| Back-end | SDK required |
|---|---|
| Max/MSP, Pure Data | Max SDK, PureData headers |
| TouchDesigner (CHOP message) | TouchDesigner Custom Operator SDK |
Godot (GDExtension node) |
none — fetched by CMake |
| Python (extension module) | pybind11 |
| ossia score | libossia |
onnxruntime is downloaded by CMake (cmake/onnxruntime.cmake, the prebuilt CPU
release) and linked into every back-end. Avendish has no ONNX helper, so the object uses
the onnxruntime C++ API directly; for richer (vision) models see
ossia/score-addon-onnx.
src/Model.hpp implements MyOnnxProcessor: set Model to a .onnx path, feed a
float vector on Input, and it runs the model's single input/output and emits the
result on Output. The session is (re)loaded when the path changes.
A tiny model is bundled: model/scale_by_two.onnx computes y = 2*x
(so [1,2,3] → [2,4,6]). Regenerate / replace it with your own; change the ports in
Model.hpp to match your model's shapes.
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build buildCMake downloads onnxruntime and godot-cpp automatically. Pass -DAVND_MAXSDK_PATH=…,
-DTOUCHDESIGNER_SDK_PATH=…, -DCMAKE_INCLUDE_PATH=…/pure-data/src,
-Dpybind11_DIR=… to enable the other back-ends — see the
Github actions workflow (shared
ossia/actions/avendish-template action).
- onnxruntime's C++ API is compiled with
ORT_NO_EXCEPTIONSto match Avendish's-fno-exceptionsbuild; model/inference errors therefore abort rather than throw, so feed valid models and shapes. - At runtime each plug-in needs the onnxruntime shared library next to it (it is fetched
into the build tree under
_deps/onnxruntime-src/lib).