scikit-learn · jeremiedbb · Feb 11, 2021 · Feb 10, 2021 · Feb 10, 2021 · Feb 10, 2021
diff --git a/sklearn/calibration.py b/sklearn/calibration.py
@@ -656,10 +656,13 @@ def predict_proba(self, X):
         if n_classes == 2:
             proba[:, 0] = 1. - proba[:, 1]
         else:
-            proba /= np.sum(proba, axis=1)[:, np.newaxis]
-
-        # XXX : for some reason all probas can be 0
-        proba[np.isnan(proba)] = 1. / n_classes
+            denominator = np.sum(proba, axis=1)[:, np.newaxis]
+            # In the edge case where for each class calibrator returns a null
+            # probability for a given sample, use the uniform distribution
+            # instead.
+            uniform_proba = np.full_like(proba, 1 / n_classes)
+            proba = np.divide(proba, denominator, out=uniform_proba,
+                              where=denominator != 0)
 
         # Deal with cases where the predicted probability minimally exceeds 1.0
         proba[(1.0 < proba) & (proba <= 1.0 + 1e-5)] = 1.0

diff --git a/sklearn/tests/test_calibration.py b/sklearn/tests/test_calibration.py
@@ -7,6 +7,7 @@
 from scipy import sparse
 
 from sklearn.base import BaseEstimator
+from sklearn.dummy import DummyClassifier
 from sklearn.model_selection import LeaveOneOut, train_test_split
 
 from sklearn.utils._testing import (assert_array_almost_equal,
@@ -26,7 +27,7 @@
 from sklearn.pipeline import Pipeline
 from sklearn.impute import SimpleImputer
 from sklearn.metrics import brier_score_loss
-from sklearn.calibration import CalibratedClassifierCV
+from sklearn.calibration import CalibratedClassifierCV, _CalibratedClassifier
 from sklearn.calibration import _sigmoid_calibration, _SigmoidCalibration
 from sklearn.calibration import calibration_curve
 
@@ -275,6 +276,29 @@ def multiclass_brier(y_true, proba_pred, n_classes):
     assert calibrated_brier < 1.1 * uncalibrated_brier
 
 
+def test_calibration_zero_probability():
+    # Test an edge case where _CalibratedClassifier avoids numerical errors
+    # in the multiclass normalization step if all the calibrators output
+    # are zero all at once for a given sample and instead fallback to uniform
+    # probabilities.
+    class ZeroCalibrator():
+        # This function is called from _CalibratedClassifier.predict_proba.
+        def predict(self, X):
+            return np.zeros(X.shape[0])
+
+    X, y = make_blobs(n_samples=50, n_features=10, random_state=7,
+                      centers=10, cluster_std=15.0)
+    clf = DummyClassifier().fit(X, y)
+    calibrator = ZeroCalibrator()
+    cal_clf = _CalibratedClassifier(
+        base_estimator=clf, calibrators=[calibrator], classes=clf.classes_)
+
+    probas = cal_clf.predict_proba(X)
+
+    # Check that all probabilities are uniformly 1. / clf.n_classes_
+    assert_allclose(probas, 1. / clf.n_classes_)
+
+
 def test_calibration_prefit():
     """Test calibration for prefitted classifiers"""
     n_samples = 50