scikit-learn
diff --git a/‎doc/whats_new/v0.24.rst
Copy file name to clipboardExpand all lines: doc/whats_new/v0.24.rst
+11Lines changed: 11 additions & 0 deletions b/‎doc/whats_new/v0.24.rst
Copy file name to clipboardExpand all lines: doc/whats_new/v0.24.rst
+11Lines changed: 11 additions & 0 deletions
diff --git a/‎sklearn/gaussian_process/_gpr.py
Copy file name to clipboardExpand all lines: sklearn/gaussian_process/_gpr.py
+4-1Lines changed: 4 additions & 1 deletion b/‎sklearn/gaussian_process/_gpr.py
Copy file name to clipboardExpand all lines: sklearn/gaussian_process/_gpr.py
+4-1Lines changed: 4 additions & 1 deletion
diff --git a/‎sklearn/gaussian_process/tests/test_gpr.py
Copy file name to clipboardExpand all lines: sklearn/gaussian_process/tests/test_gpr.py
+23Lines changed: 23 additions & 0 deletions b/‎sklearn/gaussian_process/tests/test_gpr.py
Copy file name to clipboardExpand all lines: sklearn/gaussian_process/tests/test_gpr.py
+23Lines changed: 23 additions & 0 deletions
@@ -32,6 +32,17 @@ Changelog
 - |Fix| Fixed a bug in :class:`decomposition.KernelPCA`'s
   ``inverse_transform``.  :pr:`19732` by :user:`Kei Ishikawa <kstoneriv3>`.
 
+:mod:`sklearn.gaussian_process`
+...............................
+
+- |Fix| Avoid division by zero when scaling constant target in
+  :class:`gaussian_process.GaussianProcessRegressor`. It was due to a std. dev.
+  equal to 0. Now, such case is detected and the std. dev. is affected to 1
+  avoiding a division by zero and thus the presence of NaN values in the
+  normalized target.
+  :pr:`19703` by :user:`sobkevich`, :user:`Boris Villazón-Terrazas <boricles>`
+  and :user:`Alexandr Fonari <afonari>`.
+
 :mod:`sklearn.linear_model`
 ...........................
 
 
@@ -14,6 +14,7 @@
 from ..base import BaseEstimator, RegressorMixin, clone
 from ..base import MultiOutputMixin
 from .kernels import RBF, ConstantKernel as C
+from ..preprocessing._data import _handle_zeros_in_scale
 from ..utils import check_random_state
 from ..utils.optimize import _check_optimize_result
 from ..utils.validation import _deprecate_positional_args
@@ -197,7 +198,9 @@ def fit(self, X, y):
         # Normalize target value
         if self.normalize_y:
             self._y_train_mean = np.mean(y, axis=0)
-            self._y_train_std = np.std(y, axis=0)
+            self._y_train_std = _handle_zeros_in_scale(
+                np.std(y, axis=0), copy=False
+            )
 
             # Remove mean and make unit variance
             y = (y - self._y_train_mean) / self._y_train_std
 
@@ -546,3 +546,26 @@ def test_bound_check_fixed_hyperparameter():
                         periodicity_bounds="fixed")  # seasonal component
     kernel = k1 + k2
     GaussianProcessRegressor(kernel=kernel).fit(X, y)
+
+
+# FIXME: we should test for multitargets as well. However, GPR is broken:
+# see: https://github.com/scikit-learn/scikit-learn/pull/19706
+@pytest.mark.parametrize('kernel', kernels)
+def test_constant_target(kernel):
+    """Check that the std. dev. is affected to 1 when normalizing a constant
+    feature.
+    Non-regression test for:
+    https://github.com/scikit-learn/scikit-learn/issues/18318
+    NaN where affected to the target when scaling due to null std. dev. with
+    constant target.
+    """
+    y_constant = np.ones(X.shape[0], dtype=np.float64)
+
+    gpr = GaussianProcessRegressor(kernel=kernel, normalize_y=True)
+    gpr.fit(X, y_constant)
+    assert gpr._y_train_std == pytest.approx(1.0)
+
+    y_pred, y_cov = gpr.predict(X, return_cov=True)
+    assert_allclose(y_pred, y_constant)
+    # set atol because we compare to zero
+    assert_allclose(np.diag(y_cov), 0., atol=1e-9)