Fix pathwise sampler bug (#2337)

Summary:
Pull Request resolved: #2337

`_draw_matheron_paths_ExactGP` calls `_gaussian_update_ExactGP` (as the update_strategy argument). It was passing an incorrect argument `train_targets` in place of `target_values`. This led the `Y` values to being silently ignored.

Reviewed By: saitcakmak

Differential Revision: D57175013

fbshipit-source-id: f3a1c304ee221d1460284b51f423af894c7c6a9f

Author: esantorella

botorch/sampling/pathwise/posterior_samplers.py

@@ -139,7 +139,7 @@ def _draw_matheron_paths_ExactGP(
         update_paths = update_strategy(
             model=model,
             sample_values=sample_values,
-            train_targets=train_Y,
+            target_values=train_Y,
         )
 
     return MatheronPath(

botorch/sampling/pathwise/update_strategies.py

@@ -112,7 +112,6 @@ def _gaussian_update_ExactGP(
     points: Optional[Tensor] = None,
     noise_covariance: Optional[Union[Tensor, LinearOperator]] = None,
     scale_tril: Optional[Union[Tensor, LinearOperator]] = None,
-    **ignore: Any,
 ) -> GeneralizedLinearPath:
     if points is None:
         (points,) = get_train_inputs(model, transformed=True)

test/sampling/pathwise/test_update_strategies.py

@@ -26,6 +26,7 @@
 from botorch.utils.testing import BotorchTestCase
 from gpytorch.kernels import MaternKernel, RBFKernel, ScaleKernel
 from gpytorch.likelihoods import BernoulliLikelihood
+from gpytorch.models import ExactGP
 from linear_operator.operators import ZeroLinearOperator
 from linear_operator.utils.cholesky import psd_safe_cholesky
 from torch import Size
@@ -204,3 +205,20 @@ def _test_gaussian_updates(self, model):
             with patch.object(model, "likelihood", new=BernoulliLikelihood()):
                 with self.assertRaises(NotImplementedError):
                     gaussian_update(model=model, sample_values=sample_values)
+
+            with self.subTest("Exact models with `None` target_values"):
+                assert isinstance(model, ExactGP)
+                torch.manual_seed(0)
+                path_none_target_values = gaussian_update(
+                    model=model,
+                    sample_values=sample_values,
+                )
+                torch.manual_seed(0)
+                path_with_target_values = gaussian_update(
+                    model=model,
+                    sample_values=sample_values,
+                    target_values=get_train_targets(model, transformed=True),
+                )
+                self.assertAllClose(
+                    path_none_target_values.weight, path_with_target_values.weight
+                )