Artelnics/opennn

What is OpenNN?

OpenNN is a modern C++ library (written for the C++20 standard) that lets developers build, train, and run neural‑network models directly from C++ code. It supports the usual data types – tabular tables, images, time‑series, and plain text – and can run on a regular CPU or, optionally, on an NVIDIA GPU via CUDA.

Who might use it?

Audience	Why it matters
C++ engineers who need a neural‑network component inside a larger C++ application (e.g., robotics, finance, embedded systems).	No need to switch languages or embed a Python interpreter; the library compiles straight into the binary.
Performance‑focused teams that want fine‑grained control over CPU vs. GPU execution, threading (OpenMP, oneTBB) and low‑level optimisations.	You can pick the backend that matches your hardware and compile‑time constraints.
Researchers or hobbyists who prefer a self‑contained, open‑source stack without heavyweight Python dependencies.	The library ships with examples, benchmarks against PyTorch, and a quick‑start guide.
Companies that need a commercial‑friendly licence for closed‑source products.	Dual‑licensing (AGPL + commercial) lets you stay compliant while keeping proprietary code private.

How does it work?

Installation – Clone the repo, run CMake, and choose either a CPU‑only or a CUDA‑enabled build. The build pulls in Eigen and zlib automatically; other optional runtimes (MKL, oneDNN, etc.) must be supplied by the user.
Programming model – Include the public headers (e.g., opennn/network/network.h) and use the OpenNN:: namespace to create layers, loss functions, optimisers, etc. The library follows a modular layout (core, data, network, backends).
Training & inference – Feed data (CSV, image folders, etc.) into a DataSet object, configure a NeuralNetwork, call train(), then predict() for inference. The same model can be exported as plain C or Python source code for deployment on platforms where the full library isn’t needed.
Benchmarks & validation – The benchmarks/ folder contains scripts that compare OpenNN’s speed and accuracy to PyTorch, following a documented measurement protocol.
Testing & CI – Unit tests live in tests/, and GitHub Actions run the test suite on every push.

When would you choose OpenNN over other tools?

Situation	OpenNN advantage
C++‑only codebase – you don’t want a Python runtime.	Pure C++ API, can be compiled into a static or shared library.
Embedded or real‑time systems – strict binary size or latency constraints.	Exportable C code and fine‑tuned compile options; optional CUDA for GPU acceleration.
Need for dual licensing – you want an AGPL base but a commercial path for proprietary products.	Explicit dual‑license model with a commercial option from Artelnics.
Performance comparison – you want to see how a C++ implementation stacks up against PyTorch.	Built‑in benchmark suite with reproducible protocols.

If you already rely heavily on Python ecosystems (PyTorch, TensorFlow) and don’t need C++ integration, those tools may be simpler. OpenNN shines when C++ integration, licensing flexibility, or low‑level performance control are primary concerns.

Getting started (quick recap)

# Clone & configure a CPU‑only build
git clone https://github.com/Artelnics/opennn.git
cd opennn
cmake -S . -B ../opennn-build -DCMAKE_BUILD_TYPE=Release \
      -DOpenNN_DISABLE_CUDA=ON -DOpenNN_BUILD_EXAMPLES=ON
cmake --build ../opennn-build --target blank --parallel
# Run the minimal example
../opennn-build/bin/blank

From there, follow the Iris example in examples/ to train a classifier, evaluate it, and export the model.

TL;DR

OpenNN is a C++20 neural‑network library that can run on CPU or CUDA GPUs, supports common data types, offers export to plain C/Python, includes benchmarks against PyTorch, and provides a dual AGPL/commercial licence. It’s a solid choice for developers who need a native C++ ML component or want fine‑grained control over deployment and licensing.