Fix PairwiseGP comparisons might be implicitly modified (#1811)

Summary:
Pull Request resolved: #1811

Fix the bug where the unconsolidated comparisons might be implicitly modified modified when calling forward.
This is because `self.comparisons` is passed into `set_train_data` in `forward`, which itself might already been consolidated/deduped.

Reviewed By: qingfeng10

Differential Revision: D45472153

fbshipit-source-id: 7f7659e277feed05fa405cd202aea5173351c617

botorch/models/pairwise_gp.py

@@ -914,7 +914,11 @@ def forward(self, datapoints: Tensor) -> MultivariateNormal:
             # We pass in the untransformed datapoints into set_train_data
             # as we will be setting self.datapoints as the untransformed datapoints
             # self.transform_inputs will be called inside before calling _update()
-            self.set_train_data(datapoints, self.comparisons, update_model=True)
+            self.set_train_data(
+                datapoints=datapoints,
+                comparisons=self.unconsolidated_comparisons,
+                update_model=True,
+            )
 
             transformed_dp = self.transform_inputs(self.datapoints)
 
@@ -1099,9 +1103,13 @@ def __init__(self, likelihood, model: GP):
     def forward(self, post: Posterior, comp: Tensor) -> Tensor:
         r"""Calculate approximated log evidence, i.e., log(P(D|theta))
 
+        Note that post will be based on the consolidated/deduped datapoints for
+        numerical stability, but comp will still be the unconsolidated comparisons
+        so that it's still compatible with fit_gpytorch_*.
+
         Args:
-            post: training posterior distribution from self.model
-            comp: Comparisons pairs, see PairwiseGP.__init__ for more details
+            post: training posterior distribution from self.model (after consolidation)
+            comp: Comparisons pairs (before consolidation)
 
         Returns:
             The approximated evidence, i.e., the marginal log likelihood

test/models/test_pairwise_gp.py

@@ -180,11 +180,29 @@ def test_pairwise_gp(self):
             with self.assertRaises(RuntimeError):
                 model.set_train_data(train_X, changed_train_comp, strict=True)
 
+    def test_consolidation(self):
+        for batch_shape, dtype, likelihood_cls in itertools.product(
+            (torch.Size(), torch.Size([2])),
+            (torch.float, torch.double),
+            (PairwiseLogitLikelihood, PairwiseProbitLikelihood),
+        ):
+            tkwargs = {"device": self.device, "dtype": dtype}
+            X_dim = 2
+
+            _, model_kwargs = self._get_model_and_data(
+                batch_shape=batch_shape,
+                X_dim=X_dim,
+                likelihood_cls=likelihood_cls,
+                **tkwargs,
+            )
+            train_X = model_kwargs["datapoints"]
+            train_comp = model_kwargs["comparisons"]
+
             # Test consolidation
             i1, i2 = train_X.shape[-2], train_X.shape[-2] + 1
             dup_comp = torch.cat(
                 [
-                    train_comp,
+                    model_kwargs["comparisons"],
                     torch.tensor(
                         [[i1, i2]], dtype=train_comp.dtype, device=train_comp.device
                     ).expand(*batch_shape, 1, 2),
@@ -206,6 +224,18 @@ def test_pairwise_gp(self):
                     torch.equal(model.utility, model.unconsolidated_utility)
                 )
 
+            # Calling forward with duplicated datapoints should work after consolidation
+            mll = PairwiseLaplaceMarginalLogLikelihood(model.likelihood, model)
+            # make sure model is in training mode
+            self.assertTrue(model.training)
+            pred = model(dup_X)
+            # posterior shape in training should match the consolidated utility
+            self.assertEqual(pred.shape(), model.utility.shape)
+            # which is equal to the non-duplicated train_X
+            self.assertEqual(pred.shape(), train_X.shape[:-1])
+            # Pass the original comparisons through mll should work
+            mll(pred, dup_comp)
+
     def test_condition_on_observations(self):
         for batch_shape, dtype, likelihood_cls in itertools.product(
             (torch.Size(), torch.Size([2])),