55import numpy as np
66
77from ..base import is_classifier
8+ from .multiclass import type_of_target
89from .validation import _check_response_method , check_is_fitted
910
1011
12+ def _process_predict_proba (* , y_pred , target_type , classes , pos_label ):
13+ """Get the response values when the response method is `predict_proba`.
14+
15+ This function process the `y_pred` array in the binary and multi-label cases.
16+ In the binary case, it selects the column corresponding to the positive
17+ class. In the multi-label case, it stacks the predictions if they are not
18+ in the "compressed" format `(n_samples, n_outputs)`.
19+
20+ Parameters
21+ ----------
22+ y_pred : ndarray
23+ Output of `estimator.predict_proba`. The shape depends on the target type:
24+
25+ - for binary classification, it is a 2d array of shape `(n_samples, 2)`;
26+ - for multiclass classification, it is a 2d array of shape
27+ `(n_samples, n_classes)`;
28+ - for multilabel classification, it is either a list of 2d arrays of shape
29+ `(n_samples, 2)` (e.g. `RandomForestClassifier` or `KNeighborsClassifier`) or
30+ an array of shape `(n_samples, n_outputs)` (e.g. `MLPClassifier` or
31+ `RidgeClassifier`).
32+
33+ target_type : {"binary", "multiclass", "multilabel-indicator"}
34+ Type of the target.
35+
36+ classes : ndarray of shape (n_classes,) or list of such arrays
37+ Class labels as reported by `estimator.classes_`.
38+
39+ pos_label : int, float, bool or str
40+ Only used with binary and multiclass targets.
41+
42+ Returns
43+ -------
44+ y_pred : ndarray of shape (n_samples,), (n_samples, n_classes) or \
45+
46+ Compressed predictions format as requested by the metrics.
47+ """
48+ if target_type == "binary" and y_pred .shape [1 ] < 2 :
49+ # We don't handle classifiers trained on a single class.
50+ raise ValueError (
51+ f"Got predict_proba of shape { y_pred .shape }
52+ "classifier with two classes."
53+ )
54+
55+ if target_type == "binary" :
56+ col_idx = np .flatnonzero (classes == pos_label )[0 ]
57+ return y_pred [:, col_idx ]
58+ elif target_type == "multilabel-indicator" :
59+ # Use a compress format of shape `(n_samples, n_output)`.
60+ # Only `MLPClassifier` and `RidgeClassifier` return an array of shape
61+ # `(n_samples, n_outputs)`.
62+ if isinstance (y_pred , list ):
63+ # list of arrays of shape `(n_samples, 2)`
64+ return np .vstack ([p [:, - 1 ] for p in y_pred ]).T
65+ else :
66+ # array of shape `(n_samples, n_outputs)`
67+ return y_pred
68+
69+ return y_pred
70+
71+
72+ def _process_decision_function (* , y_pred , target_type , classes , pos_label ):
73+ """Get the response values when the response method is `decision_function`.
74+
75+ This function process the `y_pred` array in the binary and multi-label cases.
76+ In the binary case, it inverts the sign of the score if the positive label
77+ is not `classes[1]`. In the multi-label case, it stacks the predictions if
78+ they are not in the "compressed" format `(n_samples, n_outputs)`.
79+
80+ Parameters
81+ ----------
82+ y_pred : ndarray
83+ Output of `estimator.predict_proba`. The shape depends on the target type:
84+
85+ - for binary classification, it is a 1d array of shape `(n_samples,)` where the
86+ sign is assuming that `classes[1]` is the positive class;
87+ - for multiclass classification, it is a 2d array of shape
88+ `(n_samples, n_classes)`;
89+ - for multilabel classification, it is a 2d array of shape `(n_samples,
90+ n_outputs)`.
91+
92+ target_type : {"binary", "multiclass", "multilabel-indicator"}
93+ Type of the target.
94+
95+ classes : ndarray of shape (n_classes,) or list of such arrays
96+ Class labels as reported by `estimator.classes_`.
97+
98+ pos_label : int, float, bool or str
99+ Only used with binary and multiclass targets.
100+
101+ Returns
102+ -------
103+ y_pred : ndarray of shape (n_samples,), (n_samples, n_classes) or \
104+
105+ Compressed predictions format as requested by the metrics.
106+ """
107+ if target_type == "binary" and pos_label == classes [0 ]:
108+ return - 1 * y_pred
109+ return y_pred
110+
111+
11112def _get_response_values (
12113 estimator ,
13114 X ,
@@ -16,12 +117,18 @@ def _get_response_values(
16117):
17118 """Compute the response values of a classifier or a regressor.
18119
19- The response values are predictions, one scalar value for each sample in X
20- that depends on the specific choice of `response_method`.
120+ The response values are predictions such that it follows the following shape:
121+
122+ - for binary classification, it is a 1d array of shape `(n_samples,)`;
123+ - for multiclass classification, it is a 2d array of shape `(n_samples, n_classes)`;
124+ - for multilabel classification, it is a 2d array of shape `(n_samples, n_outputs)`;
125+ - for regression, it is a 1d array of shape `(n_samples,)`.
21126
22127 If `estimator` is a binary classifier, also return the label for the
23128 effective positive class.
24129
130+ This utility is used primarily in the displays and the scikit-learn scorers.
131+
25132 .. versionadded:: 1.3
26133
27134 Parameters
@@ -51,8 +158,9 @@ def _get_response_values(
51158
52159 Returns
53160 -------
54- y_pred : ndarray of shape (n_samples,)
55- Target scores calculated from the provided response_method
161+ y_pred : ndarray of shape (n_samples,), (n_samples, n_classes) or \
162+
163+ Target scores calculated from the provided `response_method`
56164 and `pos_label`.
57165
58166 pos_label : int, float, bool, str or None
@@ -72,32 +180,33 @@ def _get_response_values(
72180 if is_classifier (estimator ):
73181 prediction_method = _check_response_method (estimator , response_method )
74182 classes = estimator .classes_
75- target_type = "binary" if len (classes ) <= 2 else "multiclass"
183+ target_type = type_of_target (classes )
76184
77- if pos_label is not None and pos_label not in classes .tolist ():
78- raise ValueError (
79- f"pos_label={ pos_label }
80- f"one of { classes }
81- )
82- elif pos_label is None and target_type == "binary" :
83- pos_label = pos_label if pos_label is not None else classes [- 1 ]
185+ if target_type in ("binary" , "multiclass" ):
186+ if pos_label is not None and pos_label not in classes .tolist ():
187+ raise ValueError (
188+ f"pos_label={ pos_label }
189+ f"one of { classes }
190+ )
191+ elif pos_label is None and target_type == "binary" :
192+ pos_label = classes [- 1 ]
84193
85194 y_pred = prediction_method (X )
195+
86196 if prediction_method .__name__ == "predict_proba" :
87- if target_type == "binary" and y_pred .shape [1 ] <= 2 :
88- if y_pred .shape [1 ] == 2 :
89- col_idx = np .flatnonzero (classes == pos_label )[0 ]
90- y_pred = y_pred [:, col_idx ]
91- else :
92- err_msg = (
93- f"Got predict_proba of shape { y_pred .shape }
94- "classifier with two classes."
95- )
96- raise ValueError (err_msg )
197+ y_pred = _process_predict_proba (
198+ y_pred = y_pred ,
199+ target_type = target_type ,
200+ classes = classes ,
201+ pos_label = pos_label ,
202+ )
97203 elif prediction_method .__name__ == "decision_function" :
98- if target_type == "binary" :
99- if pos_label == classes [0 ]:
100- y_pred *= - 1
204+ y_pred = _process_decision_function (
205+ y_pred = y_pred ,
206+ target_type = target_type ,
207+ classes = classes ,
208+ pos_label = pos_label ,
209+ )
101210 else : # estimator is a regressor
102211 if response_method != "predict" :
103212 raise ValueError (
0 commit comments