TuiML's own AutoML layer: pick the model, tune it, hand back a spec.
Everything here is built from parts the library already owns -- the registry, each algorithm's published parameter schema, TuiML's splitters and metrics -- so an AutoML run introduces no new dependency and no second modelling stack.
The pieces are usable separately:
search_space Turns any algorithm's get_parameter_schema() into a ParameterDistribution. Useful on its own with RandomSearchCV. portfolio Queries the registry for the algorithms that fit a task and ranks them cheap-and-strong first. automl AutoMLClassifier and AutoMLRegressor: the search itself. ensembling Greedy ensemble selection (Caruana et al., 2004) over the trial pool.
The distinguishing output is best_spec_: the winning configuration written as a train spec, so the result of a search is portable data rather than a pickled object.
>>> from tuiml.automl import AutoMLClassifier
>>> from tuiml.datasets import load_iris
>>> data = load_iris()
>>> automl = AutoMLClassifier(time_budget=5, cv=3, random_state=0)
>>> _ = automl.fit(data.X, data.y)
>>> sorted(automl.best_spec_)
['evaluation', 'model', 'pipeline']Automatic model selection and tuning under a wall-clock budget....
Greedy ensemble selection over an AutoML trial pool....
Pick and rank the algorithms an AutoML run should try....
Turn a component's JSON Schema into a searchable parameter space....
Model-guided sampling over one candidate's search space....