Remove model.train call from get_X_baseline (meta-pytorch#1289)

Summary:
Pull Request resolved: meta-pytorch#1289

Removes the `model.train()` call from `get_X_baseline` and uses `_has_transformed_inputs` and `_original_train_inputs` attributes instead to get the original train inputs.

Why: `model.train()` destroys caches on the model. For `SaasFullyBayesianSingleTaskGP` it completely destroys the model.

Reviewed By: dme65

Differential Revision: D37692137

fbshipit-source-id: d6360a6325b78e72a0e5e729ef958beb1754952d

Author: saitcakmak

botorch/optim/utils.py

@@ -276,12 +276,12 @@ def get_X_baseline(acq_function: AcquisitionFunction) -> Optional[Tensor]:
                 warnings.warn("Failed to extract X_baseline.", BotorchWarning)
                 return
         try:
-            # make sure input transforms are not applied
-            model.train()
-            if isinstance(model, ModelListGPyTorchModel):
-                X = model.models[0].train_inputs[0]
+            # Make sure we get the original train inputs.
+            m = model.models[0] if isinstance(model, ModelListGPyTorchModel) else model
+            if m._has_transformed_inputs:
+                X = m._original_train_inputs
             else:
-                X = model.train_inputs[0]
+                X = m.train_inputs[0]
         except (BotorchError, AttributeError):
             warnings.warn("Failed to extract X_baseline.", BotorchWarning)
             return