YOLO11-to-ONNX

This repository will be used as a guide to convert YOLO11 to ONNX using ultralitics to train a model and then use C++ with ONNX for inference, some details that I notice will be recorded here for future guidance.

This repo is meant for myself, but I will try to make it so anyone can follow the process easily, I will add the resources I used to guide me.

NIX

I am using NIXOS, this distro has the nix package manager, this will be used to make the compilation of C++ easier, since the packages required can be added to a nix.shell file and will be found at compile time.

The program itself can also be made available as a package, this means that it will be compiled and available as a binary in the path when you enter the shell created by nix.

The use of nix-shell can make specific CUDA versions available directly, this is important for some libraries that will only work with older CUDA versions.

EXTRA

The rest of the info there is a README file for training and another for the C++ implementation

KEY LEARNINGS AND SKILLS SHOWN IN THIS REPO

• The use of nix-shell with a specific cuda version.
• Training with ultralytics and exporting the model to onnx.
• The use of onnxruntime alongside C++.
• The use of cmake and make to compile C++ code.
• The use of nix to create a container, although simple this shows the potential to use it for deployment on web environments.
• How to create a package using nix and how to use it in a shell or containerized environment.
• The use of uv to manage python packages.
• Some specific details about YOLO11, mainly the format of the image used by it when used with ONNX.

Sources

• https://docs.ultralytics.com/datasets/detect/sku-110k/
• https://github.com/clementpoiret/nix-python-devenv/blob/cuda/devenv.nix
• https://github.com/clementpoiret/nix-python-devenv/blob/cuda/devenv.yaml
• https://devenv.sh/reference/yaml-options/
• https://discourse.nixos.org/t/where-can-i-get-libgthread-2-0-so-0/16937
• https://github.com/K4HVH/YOLOv11-ONNXRuntime-CPP

Important for cuda to have a fixed version, the version numbers work despite them not being in the nix package search page

• https://nixos.org/manual/nixpkgs/unstable/#cuda