clone, get_params and fitted state

Knowing what clone does explains why search results, CV and nested pipelines behave as they do.

Meta-estimators (pipelines, searches, cross-validation, ensembles) never fit the object you passed. They call sklearn.base.clone(est), which builds a new, unfitted estimator with the same parameters: it reads get_params(deep=False), recursively clones nested estimators, and calls the constructor again.

Consequences worth knowing:

After cross_val_score(model, X, y), model is still unfitted. Use cross_validate(..., return_estimator=True) to get the fitted fold models.
GridSearchCV leaves your original estimator untouched; the trained model is search.best_estimator_.
Fitted attributes (coef_, classes_) are never copied by clone; use copy.deepcopy if you really need a fitted copy.
A random_state given as an integer makes clones reproducible; a RandomState instance is shared and advanced between clones, so results differ.

get_params(deep=True) returns nested parameters with __ paths, exactly the names grid search accepts. set_params(**{"step__param": v}) writes them and returns the estimator.

To check fitted state use check_is_fitted(est), which looks for attributes ending in _. Estimators also carry tags (est.__sklearn_tags__()) describing capabilities such as NaN support or whether they need y, which meta-estimators consult.

Example

from sklearn.base import clone
from sklearn.datasets import load_iris
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score, cross_validate
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

X, y = load_iris(return_X_y=True)
pipe = make_pipeline(StandardScaler(), LogisticRegression(C=0.5))

cross_val_score(pipe, X, y, cv=3)
print(hasattr(pipe[-1], "coef_"))          # False (CV fitted clones, not pipe)

res = cross_validate(pipe, X, y, cv=3, return_estimator=True)
print(res["estimator"][0][-1].coef_.shape)  # (3, 4)

copy = clone(pipe.fit(X, y))
print(hasattr(copy[-1], "coef_"), copy.get_params()["logisticregression__C"])  # False 0.5
print([k for k in pipe.get_params() if k.startswith("standardscaler__")])
# ['standardscaler__copy', 'standardscaler__with_mean', 'standardscaler__with_std']

Remember

Meta-estimators fit clones, never your original object
clone copies parameters, not fitted state
Use integer random_state for reproducible clones

Watch out for

Expecting a model passed to cross_val_score to be fitted afterwards.

Check yourself

After GridSearchCV(est, grid).fit(X, y), where is the trained best model?

Official docs: scikit-learn.org/stable/modules/generated/sklearn.base.clone.html

Example

Remember

Watch out for

Check yourself

Builds on

Unlocks