Supporting InputTransforms in SparseOutlierLikelihood and `get_po…

…sterior_over_support` (#2659)

Summary:

Adding support to using the `SparseOutlierLikelihood` in conjunction with input transforms. To do this, we need to pass the transformed inputs separately to any of the marginal likelihood computation similar to the [model closures in the model fitting routines](https://github.com/pytorch/botorch/blob/466da73a18731d45b034bfd36011bb3eb150fdd8/botorch/optim/closures/model_closures.py#L185).

Reviewed By: saitcakmak

Differential Revision: D67605578

botorch/models/likelihoods/sparse_outlier_noise.py

@@ -454,7 +454,8 @@ def _optimal_rhos(self, mll: ExactMarginalLogLikelihood) -> Tensor:
         mll.train()  # NOTE: this changes model.train_inputs to be unnormalized.
         X, Y = mll.model.train_inputs[0], mll.model.train_targets
         F = mll.model(X)
-        L = mll.likelihood(F, X)
+        TX = mll.model.transform_inputs(X)
+        L = mll.likelihood(F, TX)  # likelihood expects transformed inputs
         S = L.covariance_matrix  # (Kernel Matrix + Noise Matrix)
 
         # NOTE: The following computation is mathematically equivalent to the formula

botorch/models/relevance_pursuit.py

@@ -720,7 +720,8 @@ def log_prior(
         mll.train()
         X, Y = mll.model.train_inputs[0], mll.model.train_targets
         F = mll.model(X)
-        mll_i = cast(Tensor, mll(F, Y, X))
+        TX = mll.model.transform_inputs(X) if mll.model.training else X
+        mll_i = cast(Tensor, mll(F, Y, TX))
         log_mll_trace.append(mll_i)
         support_size, log_prior_i = log_prior(
             model,

test/models/test_relevance_pursuit.py

@@ -7,8 +7,10 @@
 from __future__ import annotations
 
 import itertools
+import warnings
 
 from functools import partial
+from unittest.mock import patch
 
 import gpytorch
 import torch
@@ -197,6 +199,26 @@ def _test_robust_gp_end_to_end(
         undetected_outliers = set(outlier_indices) - set(sparse_module.support)
         self.assertEqual(len(undetected_outliers), 0)
 
+        with patch.object(
+            SparseOutlierNoise,
+            "forward",
+            wraps=sparse_module.forward,
+        ) as sparse_module_fwd:
+            # testing that posterior inference on training set does not throw warnings,
+            # which means that the passed inputs are the equal to the cached ones.
+            with warnings.catch_warnings(record=True) as warnings_log:
+                map_model.posterior(X)
+                self.assertEqual(warnings_log, [])
+            # Testing that the noise module's forward receives transformed inputs
+            X_in_call = sparse_module_fwd.call_args.kwargs["X"]
+            self.assertIsInstance(X_in_call, list)
+            self.assertEqual(len(X_in_call), 1)
+            X_in_call = X_in_call[0]
+            X_max = X_in_call.amax(dim=0)
+            X_min = X_in_call.amin(dim=0)
+            self.assertAllClose(X_max, torch.ones_like(X_max))
+            self.assertAllClose(X_min, torch.zeros_like(X_min))
+
     def test_robust_relevance_pursuit(self) -> None:
         for optimizer, convex_parameterization, dtype in itertools.product(
             [forward_relevance_pursuit, backward_relevance_pursuit],