elixir-nx/axon

Builds neural networks as composable Elixir structs using functional `defn` definitions, enabling JIT/AOT compilation to any Nx backend (CPU, GPU) and export to TensorFlow Lite or ONNX formats. Decouples three APIs—functional layers (activations, losses, metrics), model creation, and training—so each can be used independently. Training leverages Polaris optimizers with PyTorch Ignite-style event handlers for metrics and custom callbacks.

1,671 stars. Actively maintained with 5 commits in the last 30 days.