Converts an ONNX model to ORT format and serializes it to C++ source code, generate custom slimmed ONNX Runtime static libs & xcframework for apple platforms.

The goal here is to create a flexible but tiny inference engine for a specific model e.g. iPlug2 example.

NOTE: due to risk of ODR violations, global/static variable conflicts, and dependency symbol clashes with DAW Hosts that use ORT themselves - think hard before you use this in an audio plugin!

The scripts here are configured to create a minimal ORT binary using only the CPU provider. If you want to experiment with GPU inference, Core ML etc, you will have to modify.

CMake v2.6+

1. Checkout ONNX Runtime submodule $ git submodule update --init

2. Create a virtual environment and activate it

windows

$ py -3 -m venv venv
$ source ./venv/Scripts/activate`

mac/linux

$ python3 -m venv venv
$ source ./venv/bin/activate`

1. Install dependencies $ pip install -r requirements.txt

2. Run $ ./convert-model-to-ort.sh model.onnx This converts the .onnx file to .ort and produces a .config file which slims the onnxruntime library build in the next step. It also serializes the .ort format model to C++ source code, which can be used to bake the model into your app binary. If the model is large this might not be a great solution, and it might be better to locate the .ort file at runtime.

3. Build customized onnx runtime static libraries

$ ./build-mac.sh

$ ./build-ios.sh
$ ./build-ios-simulator.sh

$ ./build-xcframework.sh

Note: windows static lib builds can get very large due to the LTO/LTCG settings in onnxruntime. You can turn that off by applying the change in ltcg_patch_for_windows.patch to the onnxruntime repo. Due to different MSVC runtimes for Debug and Release builds, we need to build two binaries for windows.

$ ./build-win.sh