Added LOO interface for core.gp

GPy/core/gp.py

@@ -604,3 +604,10 @@ 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 LOO(self):
+        """
+        Evaluate the approximate leave one out error using the current state of the inference method
+        """
+        return self.inference_method.LOO(self.kern, self.X, self.Y, self.likelihood, self.posterior, Y_metadata=self.Y_metadata)
+
+

GPy/inference/latent_function_inference/__init__.py

@@ -41,6 +41,13 @@ class LatentFunctionInference(object):
         """
         pass
 
+    def LOO(self, kern, X, Y, likelihood, posterior, *args, **kwargs):
+        """
+        Leave one out error for the inference algorithm
+        """
+        raise NotImplementedError
+
+
 class InferenceMethodList(LatentFunctionInference, list):
 
     def on_optimization_start(self):

GPy/testing/model_tests.py

@@ -406,7 +406,16 @@ class MiscTests(unittest.TestCase):
         warp_f.plot(X.min()-10, X.max()+10)
         plt.show()
 
-
+    def test_LOO_laplace(self):
+        import GPy
+        gauss = GPy.likelihoods.Gaussian()
+        kern = GPy.kern.RBF(self.X.shape[1]) + GPy.kern.White(self.X.shape[1])
+        exact = GPy.inference.latent_function_inference.ExactGaussianInference()
+        m_exact = GPy.core.GP(X=self.X, Y=self.Y, kernel=kern.copy(), likelihood=gauss.copy(), inference_method=exact)
+        laplace = GPy.inference.latent_function_inference.Laplace()
+        m_laplace = GPy.core.GP(X=self.X, Y=self.Y, kernel=kern.copy(), likelihood=gauss.copy(), inference_method=laplace)
+        m_laplace[:] = m_exact[:]
+        np.testing.assert_almost_equal(m_exact.LOO(), m_laplace.LOO())
 
 class GradientTests(np.testing.TestCase):
     def setUp(self):