ENH check_is_fitted calls __is_fitted__ if available (scikit-learn#20657)

Author: adrinjalali

doc/whats_new/v1.0.rst

@@ -789,6 +789,12 @@ Changelog
   unavailable on the basis of state, in a more readable way.
   :pr:`19948` by `Joel Nothman`_.
 
+_ |Enhancement| :func:`utils.validation.check_is_fitted` now uses
+  ``__sklearn_is_fitted__`` if available, instead of checking for attributes ending with
+  an underscore. This also makes :class:`Pipeline` and
+  :class:`preprocessing.FunctionTransformer` pass
+  ``check_is_fitted(estimator)``. :pr:`20657` by `Adrin Jalali`_.
+
 - |Fix| Fixed a bug in :func:`utils.sparsefuncs.mean_variance_axis` where the
   precision of the computed variance was very poor when the real variance is
   exactly zero. :pr:`19766` by :user:`Jérémie du Boisberranger <jeremiedbb>`.

sklearn/pipeline.py

@@ -26,7 +26,9 @@
 from .utils.deprecation import deprecated
 from .utils._tags import _safe_tags
 from .utils.validation import check_memory
+from .utils.validation import check_is_fitted
 from .utils.fixes import delayed
+from .exceptions import NotFittedError
 
 from .utils.metaestimators import _BaseComposition
 
@@ -657,6 +659,18 @@ def n_features_in_(self):
         # delegate to first step (which will call _check_is_fitted)
         return self.steps[0][1].n_features_in_
 
+    def __sklearn_is_fitted__(self):
+        """Indicate whether pipeline has been fit."""
+        try:
+            # check if the last step of the pipeline is fitted
+            # we only check the last step since if the last step is fit, it
+            # means the previous steps should also be fit. This is faster than
+            # checking if every step of the pipeline is fit.
+            check_is_fitted(self.steps[-1][1])
+            return True
+        except NotFittedError:
+            return False
+
     def _sk_visual_block_(self):
         _, estimators = zip(*self.steps)
 

sklearn/preprocessing/_function_transformer.py

@@ -176,5 +176,9 @@ def _transform(self, X, func=None, kw_args=None):
 
         return func(X, **(kw_args if kw_args else {}))
 
+    def __sklearn_is_fitted__(self):
+        """Return True since FunctionTransfomer is stateless."""
+        return True
+
     def _more_tags(self):
         return {"no_validation": not self.validate, "stateless": True}

sklearn/tests/test_pipeline.py

@@ -21,7 +21,8 @@
     MinimalRegressor,
     MinimalTransformer,
 )
-
+from sklearn.exceptions import NotFittedError
+from sklearn.utils.validation import check_is_fitted
 from sklearn.base import clone, is_classifier, BaseEstimator, TransformerMixin
 from sklearn.pipeline import Pipeline, FeatureUnion, make_pipeline, make_union
 from sklearn.svm import SVC
@@ -1361,3 +1362,16 @@ def test_search_cv_using_minimal_compatible_estimator(Predictor):
     else:
         assert_allclose(y_pred, y.mean())
         assert model.score(X, y) == pytest.approx(r2_score(y, y_pred))
+
+
+def test_pipeline_check_if_fitted():
+    class Estimator(BaseEstimator):
+        def fit(self, X, y):
+            self.fitted_ = True
+            return self
+
+    pipeline = Pipeline([("clf", Estimator())])
+    with pytest.raises(NotFittedError):
+        check_is_fitted(pipeline)
+    pipeline.fit(iris.data, iris.target)
+    check_is_fitted(pipeline)

sklearn/utils/estimator_checks.py

@@ -53,6 +53,7 @@
 from ..model_selection import ShuffleSplit
 from ..model_selection._validation import _safe_split
 from ..metrics.pairwise import rbf_kernel, linear_kernel, pairwise_distances
+from ..utils.validation import check_is_fitted
 
 from . import shuffle
 from ._tags import (
@@ -307,6 +308,7 @@ def _yield_all_checks(estimator):
     yield check_dict_unchanged
     yield check_dont_overwrite_parameters
     yield check_fit_idempotent
+    yield check_fit_check_is_fitted
     if not tags["no_validation"]:
         yield check_n_features_in
         yield check_fit1d
@@ -3501,6 +3503,45 @@ def check_fit_idempotent(name, estimator_orig):
             )
 
 
+def check_fit_check_is_fitted(name, estimator_orig):
+    # Make sure that estimator doesn't pass check_is_fitted before calling fit
+    # and that passes check_is_fitted once it's fit.
+
+    rng = np.random.RandomState(42)
+
+    estimator = clone(estimator_orig)
+    set_random_state(estimator)
+    if "warm_start" in estimator.get_params():
+        estimator.set_params(warm_start=False)
+
+    n_samples = 100
+    X = rng.normal(loc=100, size=(n_samples, 2))
+    X = _pairwise_estimator_convert_X(X, estimator)
+    if is_regressor(estimator_orig):
+        y = rng.normal(size=n_samples)
+    else:
+        y = rng.randint(low=0, high=2, size=n_samples)
+    y = _enforce_estimator_tags_y(estimator, y)
+
+    if not _safe_tags(estimator).get("stateless", False):
+        # stateless estimators (such as FunctionTransformer) are always "fit"!
+        try:
+            check_is_fitted(estimator)
+            raise AssertionError(
+                f"{estimator.__class__.__name__} passes check_is_fitted before being"
+                " fit!"
+            )
+        except NotFittedError:
+            pass
+    estimator.fit(X, y)
+    try:
+        check_is_fitted(estimator)
+    except NotFittedError as e:
+        raise NotFittedError(
+            "Estimator fails to pass `check_is_fitted` even though it has been fit."
+        ) from e
+
+
 def check_n_features_in(name, estimator_orig):
     # Make sure that n_features_in_ attribute doesn't exist until fit is
     # called, and that its value is correct.

sklearn/utils/tests/test_estimator_checks.py

@@ -34,6 +34,7 @@
 from sklearn.utils.validation import check_array
 from sklearn.utils import all_estimators
 from sklearn.exceptions import SkipTestWarning
+from sklearn.utils.metaestimators import available_if
 
 from sklearn.utils.estimator_checks import (
     _NotAnArray,
@@ -52,6 +53,7 @@
     check_regressor_data_not_an_array,
     check_outlier_corruption,
     set_random_state,
+    check_fit_check_is_fitted,
 )
 
 
@@ -1006,3 +1008,28 @@ def test_minimal_class_implementation_checks():
     minimal_estimators = [MinimalTransformer(), MinimalRegressor(), MinimalClassifier()]
     for estimator in minimal_estimators:
         check_estimator(estimator)
+
+
+def test_check_fit_check_is_fitted():
+    class Estimator(BaseEstimator):
+        def __init__(self, behavior="attribute"):
+            self.behavior = behavior
+
+        def fit(self, X, y, **kwargs):
+            if self.behavior == "attribute":
+                self.is_fitted_ = True
+            elif self.behavior == "method":
+                self._is_fitted = True
+            return self
+
+        @available_if(lambda self: self.behavior in {"method", "always-true"})
+        def __sklearn_is_fitted__(self):
+            if self.behavior == "always-true":
+                return True
+            return hasattr(self, "_is_fitted")
+
+    with raises(Exception, match="passes check_is_fitted before being fit"):
+        check_fit_check_is_fitted("estimator", Estimator(behavior="always-true"))
+
+    check_fit_check_is_fitted("estimator", Estimator(behavior="method"))
+    check_fit_check_is_fitted("estimator", Estimator(behavior="attribute"))

sklearn/utils/tests/test_validation.py

@@ -52,7 +52,7 @@
     _get_feature_names,
 )
 from sklearn.utils.validation import _check_fit_params
-
+from sklearn.base import BaseEstimator
 import sklearn
 
 from sklearn.exceptions import NotFittedError, PositiveSpectrumWarning
@@ -751,6 +751,20 @@ def test_check_symmetric():
             assert_array_equal(output, arr_sym)
 
 
+def test_check_is_fitted_with_is_fitted():
+    class Estimator(BaseEstimator):
+        def fit(self, **kwargs):
+            self._is_fitted = True
+            return self
+
+        def __sklearn_is_fitted__(self):
+            return hasattr(self, "_is_fitted") and self._is_fitted
+
+    with pytest.raises(NotFittedError):
+        check_is_fitted(Estimator())
+    check_is_fitted(Estimator().fit())
+
+
 def test_check_is_fitted():
     # Check is TypeError raised when non estimator instance passed
     with pytest.raises(TypeError):

sklearn/utils/validation.py

@@ -1142,8 +1142,9 @@ def check_is_fitted(estimator, attributes=None, *, msg=None, all_or_any=all):
     fitted attributes (ending with a trailing underscore) and otherwise
     raises a NotFittedError with the given message.
 
-    This utility is meant to be used internally by estimators themselves,
-    typically in their own predict / transform methods.
+    If an estimator does not set any attributes with a trailing underscore, it
+    can define a ``__sklearn_is_fitted__`` method returning a boolean to specify if the
+    estimator is fitted or not.
 
     Parameters
     ----------
@@ -1194,13 +1195,15 @@ def check_is_fitted(estimator, attributes=None, *, msg=None, all_or_any=all):
     if attributes is not None:
         if not isinstance(attributes, (list, tuple)):
             attributes = [attributes]
-        attrs = all_or_any([hasattr(estimator, attr) for attr in attributes])
+        fitted = all_or_any([hasattr(estimator, attr) for attr in attributes])
+    elif hasattr(estimator, "__sklearn_is_fitted__"):
+        fitted = estimator.__sklearn_is_fitted__()
     else:
-        attrs = [
+        fitted = [
             v for v in vars(estimator) if v.endswith("_") and not v.startswith("__")
         ]
 
-    if not attrs:
+    if not fitted:
         raise NotFittedError(msg % {"name": type(estimator).__name__})