Minerva-OPT

Hyperparameter optimisation extensions for Minerva, powered by Ray Tune.

minerva-opt wraps Ray Tune and PyTorch Lightning into Minerva-compatible Pipeline objects so you can run distributed hyperparameter searches and ablation studies on top of any LightningModule without boilerplate.

Features

  • Drop-in Minerva pipeline — inherits from minerva.pipelines.base.Pipeline, picking up logging, reproducibility, and run-status tracking for free.

  • Flexible search algorithms — random, grid, or Bayesian optimisation (via HyperOptSearch); pass any ray.tune.search.Searcher.

  • ASHA early stopping — underperforming trials are pruned automatically; grace_period and max_t are derived from max_epochs.

  • Ablation studies — AblationStudyPipeline runs every named condition across multiple seeds and aggregates results for per-component analysis.

  • Distributed training — RayDDPStrategy + RayLightningEnvironment make each trial distributed-ready out of the box.

  • Configurable checkpointing — interval-based or keep-only-last strategies, scored on your target metric.

Quick install

pip install minerva-opt

# Bayesian optimisation support
pip install "minerva-opt[hyperopt]"