Fix box decomposition behavior with empty or None Y (#1489)

Summary:
Pull Request resolved: #1489

See T126108893

Reviewed By: SebastianAment

Differential Revision: D41170229

fbshipit-source-id: 0325791c1f190bb3a3f09c427097c5c0582c18dc

Author: esantorella

botorch/utils/multi_objective/box_decompositions/dominated.py

@@ -56,6 +56,9 @@ def compute_hypervolume(self) -> Tensor:
             A `(batch_shape)`-dim tensor containing the hypervolume dominated by
                 each Pareto frontier.
         """
+        if not hasattr(self, "_neg_pareto_Y"):
+            return torch.tensor(0.0).to(self._neg_ref_point)
+
         if self._neg_pareto_Y.shape[-2] == 0:
             return torch.zeros(
                 self._neg_pareto_Y.shape[:-2],

botorch/utils/multi_objective/box_decompositions/non_dominated.py

@@ -347,6 +347,9 @@ def compute_hypervolume(self) -> Tensor:
         Returns:
             `(batch_shape)`-dim tensor containing the dominated hypervolume.
         """
+        if not hasattr(self, "_neg_pareto_Y"):
+            return torch.tensor(0.0).to(self._neg_ref_point)
+
         if self._neg_pareto_Y.shape[-2] == 0:
             return torch.zeros(
                 self._neg_pareto_Y.shape[:-2],
@@ -460,13 +463,15 @@ def _partition_space_2d(self) -> None:
         )
         self.register_buffer("hypercell_bounds", cell_bounds)
 
-    def compute_hypervolume(self):
+    def compute_hypervolume(self) -> Tensor:
         r"""Compute hypervolume that is dominated by the Pareto Froniter.
 
         Returns:
             A `(batch_shape)`-dim tensor containing the hypervolume dominated by
                 each Pareto frontier.
         """
+        if not hasattr(self, "_neg_pareto_Y"):
+            return torch.tensor(0.0).to(self._neg_ref_point)
         if self._neg_pareto_Y.shape[-2] == 0:
             return torch.zeros(
                 self._neg_pareto_Y.shape[:-2],

test/utils/multi_objective/box_decompositions/test_box_decomposition.py

@@ -15,6 +15,13 @@
     BoxDecomposition,
     FastPartitioning,
 )
+from botorch.utils.multi_objective.box_decompositions.dominated import (
+    DominatedPartitioning,
+)
+from botorch.utils.multi_objective.box_decompositions.non_dominated import (
+    FastNondominatedPartitioning,
+    NondominatedPartitioning,
+)
 from botorch.utils.multi_objective.box_decompositions.utils import (
     update_local_upper_bounds_incremental,
 )
@@ -262,3 +269,50 @@ def test_fast_partitioning(self):
             if m == 2:
                 with self.assertRaises(NotImplementedError):
                     DummyFastPartitioning(ref_point=ref_point, Y=Y.unsqueeze(0))
+
+
+class TestBoxDecomposition_Hypervolume(BotorchTestCase):
+    def helper_hypervolume(self, Box_Decomp_cls: type) -> None:
+        """
+        This test should be run for each non-abstract subclass of `BoxDecomposition`.
+        """
+        # batching
+        n_outcomes, batch_dim, n = 2, 3, 4
+
+        ref_point = torch.zeros(n_outcomes)
+        Y = torch.ones(batch_dim, n, n_outcomes)
+
+        box_decomp = Box_Decomp_cls(ref_point=ref_point, Y=Y)
+        hv = box_decomp.compute_hypervolume()
+        self.assertEqual(hv.shape, (batch_dim,))
+        self.assertTrue(torch.allclose(hv, torch.ones(batch_dim)))
+
+        # no batching
+        Y = torch.ones(n, n_outcomes)
+
+        box_decomp = Box_Decomp_cls(ref_point=ref_point, Y=Y)
+        hv = box_decomp.compute_hypervolume()
+
+        self.assertEqual(hv.shape, ())
+        self.assertTrue(torch.allclose(hv, torch.tensor(1.0)))
+
+        # cases where there is nothing in Y, either because n=0 or Y is None
+        n = 0
+        Y_and_expected_shape = [
+            (torch.ones(batch_dim, n, n_outcomes), (batch_dim,)),
+            (torch.ones(n, n_outcomes), ()),
+            (None, ()),
+        ]
+        for Y, expected_shape in Y_and_expected_shape:
+            box_decomp = Box_Decomp_cls(ref_point=ref_point, Y=Y)
+            hv = box_decomp.compute_hypervolume()
+            self.assertEqual(hv.shape, expected_shape)
+            self.assertTrue(torch.allclose(hv, torch.tensor(0.0)))
+
+    def test_hypervolume(self) -> None:
+        for cl in [
+            NondominatedPartitioning,
+            DominatedPartitioning,
+            FastNondominatedPartitioning,
+        ]:
+            self.helper_hypervolume(cl)