Merge pull request #148 from scikit-learn-contrib/Remove-the-check_image_is_input

ENH : remove image_input argument add modify tests accordingly

Author: vtaquet

mapie/classification.py

@@ -27,8 +27,7 @@
     check_alpha_and_n_samples,
     check_n_jobs,
     check_verbose,
-    check_input_is_image,
-    fit_estimator,
+    fit_estimator
 )
 from ._compatibility import np_quantile
 
@@ -252,13 +251,8 @@ def _check_estimator(
             If the estimator is not fitted and ``cv`` attribute is "prefit".
         """
         if estimator is None:
-            if not self.image_input:
-                return LogisticRegression(multi_class="multinomial").fit(X, y)
-            else:
-                raise ValueError(
-                    "Default LogisticRegression's input can't be an image."
-                    "Please provide a proper model."
-                )
+            return LogisticRegression(multi_class="multinomial").fit(X, y)
+
         if isinstance(estimator, Pipeline):
             est = estimator[-1]
         else:
@@ -635,7 +629,6 @@ def fit(
         self,
         X: ArrayLike,
         y: ArrayLike,
-        image_input: Optional[bool] = False,
         sample_weight: Optional[ArrayLike] = None,
     ) -> MapieClassifier:
         """
@@ -649,13 +642,6 @@ def fit(
         y : ArrayLike of shape (n_samples,)
             Training labels.
 
-        image_input: Optional[bool] = False
-            Whether or not the X input is an image. If True, you must provide
-            a model that accepts image as input (e.g., a Neural Network). All
-            Scikit-learn classifiers only accept two-dimensional inputs.
-
-            By default False.
-
         sample_weight : Optional[ArrayLike] of shape (n_samples,)
             Sample weights for fitting the out-of-fold models.
             If None, then samples are equally weighted.
@@ -671,12 +657,10 @@ def fit(
             The model itself.
         """
         # Checks
-        self.image_input = image_input
         self._check_parameters()
         cv = check_cv(self.cv)
         estimator = self._check_estimator(X, y, self.estimator)
-        if self.image_input:
-            check_input_is_image(X)
+
         X, y = indexable(X, y)
         y = _check_y(y)
         assert type_of_target(y) == "multiclass"
@@ -849,8 +833,6 @@ def predict(
         include_last_label = self._check_include_last_label(include_last_label)
         alpha = cast(Optional[NDArray], check_alpha(alpha))
         check_is_fitted(self, self.fit_attributes)
-        if self.image_input:
-            check_input_is_image(X)
 
         # Estimate prediction sets
         y_pred = self.single_estimator_.predict(X)

mapie/tests/test_classification.py

@@ -415,10 +415,6 @@
     }
 ]
 
-X_WRONG_IMAGE = [
-    np.zeros((3, 1024, 1024, 3, 1)),
-    np.zeros((3, 512))
-]
 X_good_image = np.zeros((3, 1024, 1024, 3))
 y_toy_image = np.array([0, 0, 1])
 
@@ -827,7 +823,7 @@ def test_image_cumulated_scores(X: Dict[str, ArrayLike]) -> None:
         cv="prefit",
         random_state=42
     )
-    mapie.fit(cumclf.X_calib, cumclf.y_calib, image_input=True)
+    mapie.fit(cumclf.X_calib, cumclf.y_calib)
     np.testing.assert_allclose(mapie.conformity_scores_, cumclf.y_calib_scores)
     # predict
     _, y_ps = mapie.predict(
@@ -894,51 +890,6 @@ def test_classifier_without_classes_attribute(
         mapie.fit(X_toy, y_toy)
 
 
-@pytest.mark.parametrize("X_wrong_image", X_WRONG_IMAGE)
-def test_wrong_image_shape_fit(X_wrong_image: ArrayLike) -> None:
-    """
-    Test that ValueError is raised if image has not 3 or 4 dimensions in fit.
-    """
-    cumclf = ImageClassifier(X_wrong_image, y_toy_image)
-    cumclf.fit(cumclf.X_calib, cumclf.y_calib)
-    mapie = MapieClassifier(
-        cumclf,
-        method="cumulated_score",
-        cv="prefit",
-        random_state=42
-    )
-    with pytest.raises(ValueError, match=r"Invalid X.*"):
-        mapie.fit(cumclf.X_calib, cumclf.y_calib, image_input=True)
-
-
-@pytest.mark.parametrize("X_wrong_image", X_WRONG_IMAGE)
-def test_wrong_image_shape_predict(X_wrong_image: ArrayLike) -> None:
-    """
-    Test that ValueError is raised if image has not
-    3 or 4 dimensions in predict.
-    """
-    cumclf = ImageClassifier(X_good_image, y_toy_image)
-    cumclf.fit(cumclf.X_calib, cumclf.y_calib)
-    mapie = MapieClassifier(
-        cumclf,
-        method="cumulated_score",
-        cv="prefit",
-        random_state=42
-    )
-    mapie.fit(cumclf.X_calib, cumclf.y_calib, image_input=True,)
-    with pytest.raises(ValueError, match=r"Invalid X.*"):
-        mapie.predict(X_wrong_image)
-
-
-def test_undefined_model() -> None:
-    """
-    Test ValueError is raised if no model is specified with image input.
-    """
-    mapie = MapieClassifier()
-    with pytest.raises(ValueError, match=r"LogisticRegression's input.*"):
-        mapie.fit(X_good_image, y_toy_image, image_input=True,)
-
-
 @pytest.mark.parametrize("method", WRONG_METHODS)
 def test_method_error_in_fit(monkeypatch: Any, method: str) -> None:
     """Test else condition for the method in .fit"""

mapie/utils.py

@@ -404,27 +404,3 @@ def check_nan_in_aposteriori_prediction(X: ArrayLike) -> None:
             + "belongs to every resamplings.\n"
             "Increase the number of resamplings"
         )
-
-
-def check_input_is_image(X: ArrayLike) -> None:
-    """
-    Check if the image has 3 or 4 dimensions
-
-    Parameters
-    ----------
-    X: Union[
-        ArrayLike[n_samples, width, height],
-        ArrayLike[n_samples, width, height, n_channels]
-    ]
-        Image input
-
-    Raises
-    ------
-    ValueError
-    """
-    if len(np.array(X).shape) not in [3, 4]:
-        raise ValueError(
-            "Invalid X."
-            "When X is an image, the number of dimensions"
-            "must be equal to 3 or 4."
-        )