ENH: implement checks for CrossConformalRegression, make utils.py non-public

Author: Valentin-Laurent

mapie_v1/_utils.py

@@ -0,0 +1,49 @@
+import warnings
+from typing import Union, List
+
+from numpy import array
+from numpy.typing import ArrayLike
+from sklearn.model_selection import BaseCrossValidator
+
+
+def transform_confidence_level_to_alpha_list(
+    confidence_level: Union[float, List[float]]
+) -> List[float]:
+    if isinstance(confidence_level, float):
+        confidence_levels = [confidence_level]
+    else:
+        confidence_levels = confidence_level
+    return [1 - level for level in confidence_levels]
+
+
+def check_method_not_naive(method: str) -> None:
+    if method == "naive":
+        raise ValueError(
+            '"naive" method not available in MAPIE >= v1'
+        )
+
+
+def check_cv_not_string(cv: Union[int, str, BaseCrossValidator]):
+    if isinstance(cv, str):
+        raise ValueError(
+            "'cv' string options not available in MAPIE >= v1"
+        )
+
+
+def hash_X_y(X: ArrayLike, y: ArrayLike) -> int:
+    # Known issues:
+    # - the hash calculated with `hash` changes between Python processes
+    # - two arrays with  the same content but different shapes will all have
+    #   the same hash because .tobytes() ignores shape
+    return hash(array(X).tobytes() + array(y).tobytes())
+
+
+def check_if_X_y_different_from_fit(
+    X: ArrayLike,
+    y: ArrayLike,
+    previous_X_y_hash: int
+) -> None:
+    if hash_X_y(X, y) != previous_X_y_hash:
+        warnings.warn(
+            "You have to use the same X and y in .fit and .conformalize"
+        )

mapie_v1/conformity_scores/utils.py mapie_v1/conformity_scores/_utils.pymapie_v1/conformity_scores/utils.py renamed to mapie_v1/conformity_scores/_utils.py

@@ -3,7 +3,7 @@
 from . import REGRESSION_CONFORMITY_SCORES_STRING_MAP
 
 
-def check_and_select_split_conformity_score(
+def check_and_select_regression_conformity_score(
     conformity_score: Union[str, BaseRegressionScore]
 ) -> BaseRegressionScore:
     if isinstance(conformity_score, BaseRegressionScore):

mapie_v1/integration_tests/tests/test_regression.py

@@ -15,8 +15,8 @@
 
 from mapiev0.regression import MapieRegressor as MapieRegressorV0  # noqa
 
-from mapie_v1.conformity_scores.utils import \
-    check_and_select_split_conformity_score
+from mapie_v1.conformity_scores._utils import \
+    check_and_select_regression_conformity_score
 from mapie_v1.integration_tests.utils import (filter_params,
                                               train_test_split_shuffle)
 from sklearn.model_selection import LeaveOneOut, GroupKFold
@@ -66,7 +66,7 @@ def test_exact_interval_equality_split(
     v0_params = {
         "estimator": estimator,
         "method": method,
-        "conformity_score": check_and_select_split_conformity_score(
+        "conformity_score": check_and_select_regression_conformity_score(
             conformity_score
         ),
         "alpha": 1 - confidence_level,

mapie_v1/regression.py

@@ -10,10 +10,12 @@
 from mapie.conformity_scores import BaseRegressionScore
 from mapie.regression import MapieRegressor
 from mapie.utils import check_estimator_fit_predict
-from mapie_v1.conformity_scores.utils import (
-    check_and_select_split_conformity_score,
+from mapie_v1.conformity_scores._utils import (
+    check_and_select_regression_conformity_score,
 )
-from mapie_v1.utils import transform_confidence_level_to_alpha_list
+from mapie_v1._utils import transform_confidence_level_to_alpha_list, \
+    check_method_not_naive, check_cv_not_string, hash_X_y, \
+    check_if_X_y_different_from_fit
 
 
 class SplitConformalRegressor:
@@ -87,7 +89,7 @@ def __init__(
         check_estimator_fit_predict(estimator)
         self._estimator = estimator
         self._prefit = prefit
-        self._conformity_score = check_and_select_split_conformity_score(
+        self._conformity_score = check_and_select_regression_conformity_score(
             conformity_score)
 
         # Note to developers: to implement this v1 class without touching the
@@ -323,20 +325,27 @@ def __init__(
         verbose: int = 0,
         random_state: Optional[Union[int, np.random.RandomState]] = None
     ) -> None:
-        self.mapie_regressor = MapieRegressor(
-            estimator=self.estimator,
+        check_method_not_naive(method)
+        check_cv_not_string(cv)
+
+        self._mapie_regressor = MapieRegressor(
+            estimator=estimator,
             method=method,
             cv=cv,
             n_jobs=n_jobs,
             verbose=verbose,
-            conformity_score=self.conformity_score,
+            conformity_score=check_and_select_regression_conformity_score(
+                conformity_score
+            ),
             random_state=random_state,
         )
 
         self._alphas = transform_confidence_level_to_alpha_list(
             confidence_level
         )
 
+        self.hashed_X_y: int = 0
+
     def fit(
         self,
         X: ArrayLike,
@@ -363,10 +372,14 @@ def fit(
         Self
             The fitted CrossConformalRegressor instance.
         """
-        X, y, sample_weight, groups = self.init_fit(
+        self.hashed_X_y = hash_X_y(X, y)
+
+        X, y, sample_weight, groups = self._mapie_regressor.init_fit(
             X, y, fit_params=fit_params
         )
-        self.mapie_regressor.fit_estimator(X, y, sample_weight, groups)
+        self._mapie_regressor.fit_estimator(X, y, sample_weight, groups)
+
+        return self
 
     def conformalize(
         self,
@@ -403,7 +416,9 @@ def conformalize(
         Self
             The conformalized SplitConformalRegressor instance.
         """
-        self.mapie_regressor.conformalize(
+        check_if_X_y_different_from_fit(X, y, self.hashed_X_y)
+
+        self._mapie_regressor.conformalize(
             X,
             y,
             groups=groups,