Casting Tensor to LinearOperator in constructor of MultivariateNormal…

… to ensure PSD-safe factorization

gpytorch/distributions/multitask_multivariate_normal.py

@@ -2,7 +2,10 @@
 
 import torch
 from linear_operator import LinearOperator, to_linear_operator
-from linear_operator.operators import BlockDiagLinearOperator, BlockInterleavedLinearOperator, CatLinearOperator
+from linear_operator.operators import (BlockDiagLinearOperator,
+                                       BlockInterleavedLinearOperator,
+                                       CatLinearOperator)
+from torch import Tensor
 
 from .multivariate_normal import MultivariateNormal
 
@@ -24,7 +27,9 @@ class MultitaskMultivariateNormal(MultivariateNormal):
         w.r.t. inter-observation covariance for each task.
     """
 
-    def __init__(self, mean, covariance_matrix, validate_args=False, interleaved=True):
+    def __init__(
+        self, mean: Tensor, covariance_matrix: LinearOperator, validate_args: bool = False, interleaved: bool = True
+    ):
         if not torch.is_tensor(mean) and not isinstance(mean, LinearOperator):
             raise RuntimeError("The mean of a MultitaskMultivariateNormal must be a Tensor or LinearOperator")
 

gpytorch/distributions/multivariate_normal.py

@@ -9,7 +9,8 @@
 
 import torch
 from linear_operator import to_dense, to_linear_operator
-from linear_operator.operators import DiagLinearOperator, LinearOperator, RootLinearOperator
+from linear_operator.operators import (DiagLinearOperator, LinearOperator,
+                                       RootLinearOperator)
 from torch import Tensor
 from torch.distributions import MultivariateNormal as TMultivariateNormal
 from torch.distributions.kl import register_kl
@@ -42,27 +43,36 @@ class MultivariateNormal(TMultivariateNormal, Distribution):
     :ivar torch.Tensor variance: The variance.
     """
 
-    def __init__(self, mean: Tensor, covariance_matrix: Union[Tensor, LinearOperator], validate_args: bool = False):
-        self._islazy = isinstance(mean, LinearOperator) or isinstance(covariance_matrix, LinearOperator)
-        if self._islazy:
-            if validate_args:
-                ms = mean.size(-1)
-                cs1 = covariance_matrix.size(-1)
-                cs2 = covariance_matrix.size(-2)
-                if not (ms == cs1 and ms == cs2):
-                    raise ValueError(f"Wrong shapes in {self._repr_sizes(mean, covariance_matrix)}")
-            self.loc = mean
-            self._covar = covariance_matrix
-            self.__unbroadcasted_scale_tril = None
-            self._validate_args = validate_args
-            batch_shape = torch.broadcast_shapes(self.loc.shape[:-1], covariance_matrix.shape[:-2])
-
-            event_shape = self.loc.shape[-1:]
-
-            # TODO: Integrate argument validation for LinearOperators into torch.distribution validation logic
-            super(TMultivariateNormal, self).__init__(batch_shape, event_shape, validate_args=False)
-        else:
-            super().__init__(loc=mean, covariance_matrix=covariance_matrix, validate_args=validate_args)
+    def __init__(
+        self,
+        mean: Union[Tensor, LinearOperator],
+        covariance_matrix: Union[Tensor, LinearOperator],
+        validate_args: bool = False,
+    ):
+        self._islazy = True
+        # casting Tensor to DenseLinearOperator because the super constructor calls cholesky, which
+        # will fail if the covariance matrix is semi-definite, whereas DenseLinearOperator ends up
+        # calling _psd_safe_cholesky, which factorizes semi-definite matrices by adding to the diagonal.
+        if isinstance(covariance_matrix, Tensor):
+            self._islazy = False  # to allow _unbroadcasted_scale_tril setter
+            covariance_matrix = to_linear_operator(covariance_matrix)
+
+        if validate_args:
+            ms = mean.size(-1)
+            cs1 = covariance_matrix.size(-1)
+            cs2 = covariance_matrix.size(-2)
+            if not (ms == cs1 and ms == cs2):
+                raise ValueError(f"Wrong shapes in {self._repr_sizes(mean, covariance_matrix)}")
+        self.loc = mean
+        self._covar = covariance_matrix
+        self.__unbroadcasted_scale_tril = None
+        self._validate_args = validate_args
+        batch_shape = torch.broadcast_shapes(self.loc.shape[:-1], covariance_matrix.shape[:-2])
+
+        event_shape = self.loc.shape[-1:]
+
+        # TODO: Integrate argument validation for LinearOperators into torch.distribution validation logic
+        super(TMultivariateNormal, self).__init__(batch_shape, event_shape, validate_args=False)
 
     def _extended_shape(self, sample_shape: torch.Size = torch.Size()) -> torch.Size:
         """
@@ -81,16 +91,16 @@ def _extended_shape(self, sample_shape: torch.Size = torch.Size()) -> torch.Size
     def _repr_sizes(mean: Tensor, covariance_matrix: Union[Tensor, LinearOperator]) -> str:
         return f"MultivariateNormal(loc: {mean.size()}, scale: {covariance_matrix.size()})"
 
-    @property
+    @property  # not using lazy_property here, because it does not allow for setter below
     def _unbroadcasted_scale_tril(self) -> Tensor:
-        if self.islazy and self.__unbroadcasted_scale_tril is None:
+        if self.__unbroadcasted_scale_tril is None:
             # cache root decoposition
             ust = to_dense(self.lazy_covariance_matrix.cholesky())
             self.__unbroadcasted_scale_tril = ust
         return self.__unbroadcasted_scale_tril
 
     @_unbroadcasted_scale_tril.setter
-    def _unbroadcasted_scale_tril(self, ust: Tensor):
+    def _unbroadcasted_scale_tril(self, ust: Tensor) -> None:
         if self.islazy:
             raise NotImplementedError("Cannot set _unbroadcasted_scale_tril for lazy MVN distributions")
         else:
@@ -114,10 +124,7 @@ def base_sample_shape(self) -> torch.Size:
 
     @lazy_property
     def covariance_matrix(self) -> Tensor:
-        if self.islazy:
-            return self._covar.to_dense()
-        else:
-            return super().covariance_matrix
+        return self._covar.to_dense()
 
     def confidence_region(self) -> Tuple[Tensor, Tensor]:
         """
@@ -157,10 +164,7 @@ def get_base_samples(self, sample_shape: torch.Size = torch.Size()) -> Tensor:
 
     @lazy_property
     def lazy_covariance_matrix(self) -> LinearOperator:
-        if self.islazy:
-            return self._covar
-        else:
-            return to_linear_operator(super().covariance_matrix)
+        return self._covar
 
     def log_prob(self, value: Tensor) -> Tensor:
         r"""
@@ -304,13 +308,10 @@ def to_data_independent_dist(self) -> torch.distributions.Normal:
 
     @property
     def variance(self) -> Tensor:
-        if self.islazy:
-            # overwrite this since torch MVN uses unbroadcasted_scale_tril for this
-            diag = self.lazy_covariance_matrix.diagonal(dim1=-1, dim2=-2)
-            diag = diag.view(diag.shape[:-1] + self._event_shape)
-            variance = diag.expand(self._batch_shape + self._event_shape)
-        else:
-            variance = super().variance
+        # overwrite this since torch MVN uses unbroadcasted_scale_tril for this
+        diag = self.lazy_covariance_matrix.diagonal(dim1=-1, dim2=-2)
+        diag = diag.view(diag.shape[:-1] + self._event_shape)
+        variance = diag.expand(self._batch_shape + self._event_shape)
 
         # Check to make sure that variance isn't lower than minimum allowed value (default 1e-6).
         # This ensures that all variances are positive

test/distributions/test_multivariate_normal.py

@@ -47,6 +47,20 @@ def test_multivariate_normal_non_lazy(self, cuda=False):
             self.assertTrue(mvn.sample(torch.Size([2])).shape == torch.Size([2, 3]))
             self.assertTrue(mvn.sample(torch.Size([2, 4])).shape == torch.Size([2, 4, 3]))
 
+            # testing with semi-definite input
+            A = torch.randn(len(mean), 1)
+            covmat = A @ A.T
+            handles_psd = False
+            try:
+                # the regular call fails:
+                # mvn = TMultivariateNormal(loc=mean, covariance_matrix=covmat, validate_args=True)
+                mvn = MultivariateNormal(mean=mean, covariance_matrix=covmat, validate_args=True)
+                mvn.sample()
+                handles_psd = True
+            except ValueError:
+                handles_psd = False
+            self.assertTrue(handles_psd)
+
     def test_multivariate_normal_non_lazy_cuda(self):
         if torch.cuda.is_available():
             with least_used_cuda_device():