atune

atune is a hyperparameter optimizer in Rust. You declare the hyperparameters, such as a learning rate or a network width; atune proposes values, scores them and spends the remaining budget where the results are best. It runs as a Rust library, a Python package and a command-line tool.

Live demo

The atune live demo tuning CartPole, with trial values and the best policy on a chartThe atune live demo tuning CartPole, with trial values and the best policy on a chart
Open full page

Pick one of nine samplers and a scheduler, then press Start: each trial trains a small CartPole policy in your browser, and the heatmap shows where the search went. You can also tune a test function, such as Rastrigin, or draw your own.

Features

  • Reproducible: with a fixed seed, a stateless sampler over a fixed space gives trial n the same parameters on any number of threads; adaptive samplers need a single thread and a seeded objective.
  • Built for reinforcement learning: multi-fidelity search (DEHB), population-based training (PBT and PB2), pruners that stop weak runs early (median, ASHA, Hyperband, Wilcoxon) and multi-seed evaluation.
  • Open search spaces: a range marked open widens during the search when the best trials cluster at its edge.
  • Interoperable: a Python package via PyO3, Optuna journal import and export, and a command line that tunes any program that reads environment variables and prints a number.

Status

atune is in pre-release development; its source, crate and Python package are not public yet.