libonnx

Libonnx

A lightweight, portable pure C99 onnx inference engine for embedded devices with hardware acceleration support.

Getting Started

The library's .c and .h files can be dropped into a project and compiled along with it. Before use, should be allocated struct onnx_context_t * and you can pass an array of struct resolver_t * for hardware acceleration.

The filename is path to the format of onnx model.

struct onnx_context_t * ctx = onnx_context_alloc_from_file(filename, NULL, 0);

Then, you can get input and output tensor using onnx_tensor_search function.

struct onnx_tensor_t * input = onnx_tensor_search(ctx, "input-tensor-name");
struct onnx_tensor_t * output = onnx_tensor_search(ctx, "output-tensor-name");

When the input tensor has been setting, you can run inference engine using onnx_run function and the result will putting into the output tensor.

onnx_run(ctx);

Finally, you must free struct onnx_context_t * using onnx_context_free function.

onnx_context_free(ctx);

Compilation Instructions

Just type make at the root directory, you will see a static library and some binary of examples and tests for usage.

cd libonnx
make

To compile the mnist example, you will have to install SDL2 and SDL2 GFX. On systems like Ubuntu run

    apt-get install libsdl2-dev libsdl2-gfx-dev

to install the required Simple DirectMedia Layer libraries to run the GUI.

Cross compilation example (for arm64)

Run make CROSS_COMPILE=path/to/toolchains/aarch64-linux-gnu- at the root directory to compile all libraries, tests and examples for the platform.

Change CROSS_COMPILE to point the toolchains that you plan to use.

How to run examples

After compiling all the files, you can run an example by using:

cd libonnx/examples/hello/output
./hello

Screenshots

• Mnist handwritten digit prediction

Running tests

To run tests, for example on those in the tests/model folder use:

cd libonnx/tests/output
./tests ../model

Here is the output:

…

Note that running the test on the other folders may not succeed. Some operators have not been implemented, look bat the notes section for more info.

Notes

• This library based on the onnx version v1.17.0 with the newest opset 24 support. The supported operator table in the documents directory.
• Checkout the tools folder for help with ONNX model files.
• You can use xxd -i (on Linux) to convert your onnx model into a unsigned char array and then use the function onnx_context_alloc to use it. This is how the models are loaded in the examples - hello and mnist.

Links

• The chinese discussion posts
• The onnx documentation
• The onnx operators documentation
• The tutorials for creating ONNX models
• The pre-trained onnx models

License

This library is free software; you can redistribute it and or modify it under the terms of the MIT license. See MIT License for details.