axon

Why we included this project

Elixir doesn't have a huge deep learning ecosystem, but Axon is the library that feels native to it. It builds directly on Nx, so your models are just numerical definitions that run on any Nx backend or compiler, whether that's CPU, GPU, or something else, and you can even export them to ONNX or TensorFlow Lite. The design splits into a low-level functional API covering activations, layers, and losses, a higher-level model creation layer, and a training API inspired by PyTorch Ignite, each of which can be used on its own. If you're already doing tensor work in Nx or want to keep training and inference inside an Elixir service without spinning up a separate Python process, Axon lets you stay in one language. The decoupled structure also means you can plug in your own numerical definitions or a custom training loop and still use the built-in building blocks.