training_cfgs

Standalone config manager migrated from svaichu/robotdataset: a fluent, self-learning Config object for training runs, with typed CLI overrides, W&B sweep export, and two-way Optuna integration — all driven by the same schema.

📖 Full documentation on Read the Docs

Install

pyyaml and optuna are core dependencies — a plain install always gets YAML config loading and Optuna support, no extras to remember:

pip install -e .          # pulls in pyyaml + optuna
pip install -e ".[dev]"   # + pytest, for running the test suite

Quick start

from training_cfgs import CategoricalDistribution, Config, FloatDistribution, create_study

cfg = Config.from_file("config.yaml")
cfg.set_distribution("training", "learning_rate", FloatDistribution(1e-5, 1e-2, log=True))
cfg.set_distribution("training", "optimizer", CategoricalDistribution(["adam", "sgd"]))

# W&B sweep export
sweep = cfg.to_sweep()

# CLI overrides: python train.py --training.learning_rate 1e-3
cfg = Config.from_cli(default_config="config.yaml")

# Optuna search, same schema as the sweep export above
study = create_study(direction="minimize")
cfg.single_objective_optimization(study, train, n_trials=50)
best_cfg = cfg.best_from_optuna(study)

Docs

The full guide lives on Read the Docs, built from the same sources in this repo: doc/basics.md (loading, CLI overrides, exporting), doc/wandb_sweeps.md (HPO bounds, W&B sweep export), doc/optuna.md (Optuna integration) — and examples/ for runnable scripts.