Moved posterior_covariance to Posterior class

GPy/core/gp.py

@@ -8,7 +8,6 @@ from .mapping import Mapping
 from .. import likelihoods
 from .. import kern
 from ..inference.latent_function_inference import exact_gaussian_inference, expectation_propagation
-from ..util.linalg import dtrtrs
 from ..util.normalizer import Standardize
 from paramz import ObsAr
 
@@ -680,23 +679,11 @@ class GP(Model):
         mu_star, var_star = self._raw_predict(x_test)
         return self.likelihood.log_predictive_density_sampling(y_test, mu_star, var_star, Y_metadata=Y_metadata, num_samples=num_samples)
 
-    def posterior_covariance(self, X1, X2):
+    def posterior_covariance_between_points(self, X1, X2):
         """
         Computes the posterior covariance between points.
 
         :param X1: some input observations
         :param X2: other input observations
         """
-        # ndim == 3 is a model for missing data
+        return self.posterior.covariance_between_points(self.kern, self.X, X1, X2)
-        if self.posterior.woodbury_chol.ndim != 2:
-            raise RuntimeError("This method does not support posterior for missing data models")
-
-        Kx1 = self.kern.K(self.X, X1)
-        Kx2 = self.kern.K(self.X, X2)
-        K12 = self.kern.K(X1, X2)
-
-        tmp1 = dtrtrs(self.posterior.woodbury_chol, Kx1)[0]
-        tmp2 = dtrtrs(self.posterior.woodbury_chol, Kx2)[0]
-        var = K12 - tmp1.T.dot(tmp2)
-
-        return var

GPy/inference/latent_function_inference/posterior.py

@@ -101,6 +101,29 @@ class Posterior(object):
             #self._covariance = self._K - self._K.dot(self.woodbury_inv).dot(self._K)
         return self._covariance
 
+    def covariance_between_points(self, kern, X, X1, X2):
+        """
+        Computes the posterior covariance between points.
+
+        :param kern: GP kernel
+        :param X: current input observations
+        :param X1: some input observations
+        :param X2: other input observations
+        """
+        # ndim == 3 is a model for missing data
+        if self.woodbury_chol.ndim != 2:
+            raise RuntimeError("This method does not support posterior for missing data models")
+
+        Kx1 = kern.K(X, X1)
+        Kx2 = kern.K(X, X2)
+        K12 = kern.K(X1, X2)
+
+        tmp1 = dtrtrs(self.woodbury_chol, Kx1)[0]
+        tmp2 = dtrtrs(self.woodbury_chol, Kx2)[0]
+        var = K12 - tmp1.T.dot(tmp2)
+
+        return var
+
     @property
     def precision(self):
         """

GPy/testing/model_tests.py

@@ -1084,7 +1084,7 @@ class GradientTests(np.testing.TestCase):
         Y = np.array([[1], [2]])
         m = GPy.models.GPRegression(X1, Y, kernel=k)
 
-        result = m.posterior_covariance(X1, X2)
+        result = m.posterior_covariance_between_points(X1, X2)
         expected = np.array([[0.4, 2.2], [1.0, 1.0]]) / 3.0
 
         self.assertTrue(np.allclose(result, expected))
@@ -1095,7 +1095,7 @@ class GradientTests(np.testing.TestCase):
         m = _create_missing_data_model(k, Q)
 
         with self.assertRaises(RuntimeError):
-            m.posterior_covariance(np.array([[1], [2]]), np.array([[3], [4]]))
+            m.posterior_covariance_between_points(np.array([[1], [2]]), np.array([[3], [4]]))
 
 def _create_missing_data_model(kernel, Q):
     D1, D2, D3, N, num_inducing = 13, 5, 8, 400, 3