Compass

poseigen_compass is the hyperparameter search and model-evaluation package in the Poseigen ecosystem. It generates candidate configurations, evaluates models across data splits and repeats, and supports ensembling and statistical assessment.

Installation

Requires Python 3.12 or newer. Install from PyPI:

python -m pip install poseigen_compass

Package

Compass's public helpers are re-exported from the poseigen_compass package:

• Candidate generation — RandomCanGen samples continuous, integer, and categorical hyperparameter spaces.
• Model evaluation — StandardCanScorer scores a model with a selected metric, while CanEvaluator and RandomOpt evaluate candidates across splits and repeats, with optional parallel execution.
• Candidate selection — TopCan ranks candidates by their evaluation scores, and TopCanDict returns the selected configurations.
• Ensembling and uncertainty — EnsembleScorer combines model predictions; Bootstrapper, BootstrapConfidenceInterval, and BootstrapStandardError support bootstrap-based assessment.

Compass uses poseigen_seaside for shared utilities and metrics. poseigen_oyster provides Oyster_BaseDict and Dual_Oyster_BaseDict search spaces for Compass.

Usage

RandomCanGen accepts a dictionary that defines the range or choices for each parameter:

import poseigen_compass as co

search_space = {
    "learning_rate": [[1e-4, 1e-2], "float"],
    "batch_size": [[16, 32, 64], "cat"],
    "max_depth": [[2, 12], "int"],
}

candidates = co.RandomCanGen(search_space, num_can=20)

Pass the resulting candidates to RandomOpt to evaluate them with a model, data, and an optional set of splits.