Hotfix: fix .diagonal() calls for keops kernel matrices.

[Fixes #2589]

gpytorch/kernels/keops/keops_kernel.py

@@ -12,6 +12,8 @@
     import pykeops  # noqa F401
     from pykeops.torch import LazyTensor
 
+    _Anysor = Union[Tensor, LazyTensor]
+
     def _lazify_and_expand_inputs(
         x1: Tensor, x2: Tensor
     ) -> Tuple[Union[Tensor, LazyTensor], Union[Tensor, LazyTensor]]:
@@ -49,6 +51,8 @@ def __call__(self, *args: Any, **kwargs: Any) -> Union[LinearOperator, Tensor, L
 
 except ImportError:
 
+    _Anysor = Tensor
+
     def _lazify_and_expand_inputs(x1: Tensor, x2: Tensor) -> Tuple[Tensor, Tensor]:
         x1_ = x1[..., :, None, :]
         x2_ = x2[..., None, :, :]

gpytorch/kernels/keops/matern_kernel.py

@@ -2,11 +2,12 @@
 import math
 
 from linear_operator.operators import KernelLinearOperator
+from torch import Tensor
 
-from .keops_kernel import _lazify_and_expand_inputs, KeOpsKernel
+from .keops_kernel import _Anysor, _lazify_and_expand_inputs, KeOpsKernel
 
 
-def _covar_func(x1, x2, nu=2.5, **params):
+def _covar_func(x1: _Anysor, x2: _Anysor, nu: float = 2.5, **params) -> _Anysor:
     x1_, x2_ = _lazify_and_expand_inputs(x1, x2)
 
     sq_distance = ((x1_ - x2_) ** 2).sum(-1)
@@ -57,15 +58,22 @@ class MaternKernel(KeOpsKernel):
 
     has_lengthscale = True
 
-    def __init__(self, nu=2.5, **kwargs):
+    def __init__(self, nu: float = 2.5, **kwargs):
         if nu not in {0.5, 1.5, 2.5}:
             raise RuntimeError("nu expected to be 0.5, 1.5, or 2.5")
         super().__init__(**kwargs)
         self.nu = nu
 
-    def forward(self, x1, x2, **kwargs):
+    def forward(self, x1: Tensor, x2: Tensor, diag: bool = False, **kwargs) -> KernelLinearOperator:
         mean = x1.reshape(-1, x1.size(-1)).mean(0)[(None,) * (x1.dim() - 1)]
         x1_ = (x1 - mean) / self.lengthscale
         x2_ = (x2 - mean) / self.lengthscale
         # return KernelLinearOperator inst only when calculating the whole covariance matrix
-        return KernelLinearOperator(x1_, x2_, covar_func=_covar_func, nu=self.nu, **kwargs)
+        res = KernelLinearOperator(x1_, x2_, covar_func=_covar_func, nu=self.nu, **kwargs)
+
+        # TODO: diag=True mode will be removed with the GpyTorch 2.0 PR to remove LazyEvaluatedKernelTensor
+        # (it will be replaced by a `_symmetric_diag` method for quickly computing the diagonals of symmetric matrices)
+        if diag:
+            return res.diagonal(dim1=-1, dim2=-2)
+        else:
+            return res

gpytorch/kernels/keops/periodic_kernel.py

@@ -3,12 +3,13 @@
 import math
 
 from linear_operator.operators import KernelLinearOperator
+from torch import Tensor
 
 from ..periodic_kernel import PeriodicKernel as GPeriodicKernel
-from .keops_kernel import _lazify_and_expand_inputs, KeOpsKernel
+from .keops_kernel import _Anysor, _lazify_and_expand_inputs, KeOpsKernel
 
 
-def _covar_func(x1, x2, lengthscale, **kwargs):
+def _covar_func(x1: _Anysor, x2: _Anysor, lengthscale: Tensor, **kwargs) -> _Anysor:
     x1_, x2_ = _lazify_and_expand_inputs(x1, x2)
     lengthscale = lengthscale[..., None, None, 0, :]  # 1 x 1 x ndim
     # do not use .power(2.0) as it gives NaN values on cuda
@@ -56,9 +57,16 @@ class PeriodicKernel(KeOpsKernel, GPeriodicKernel):
 
     has_lengthscale = True
 
-    def forward(self, x1, x2, **kwargs):
+    def forward(self, x1: Tensor, x2: Tensor, diag: bool = False, **kwargs) -> KernelLinearOperator:
         x1_ = x1.div(self.period_length / math.pi)
         x2_ = x2.div(self.period_length / math.pi)
         # return KernelLinearOperator inst only when calculating the whole covariance matrix
         # pass any parameters which are used inside _covar_func as *args to get gradients computed for them
-        return KernelLinearOperator(x1_, x2_, lengthscale=self.lengthscale, covar_func=_covar_func, **kwargs)
+        res = KernelLinearOperator(x1_, x2_, lengthscale=self.lengthscale, covar_func=_covar_func, **kwargs)
+
+        # TODO: diag=True mode will be removed with the GpyTorch 2.0 PR to remove LazyEvaluatedKernelTensor
+        # (it will be replaced by a `_symmetric_diag` method for quickly computing the diagonals of symmetric matrices)
+        if diag:
+            return res.diagonal(dim1=-1, dim2=-2)
+        else:
+            return res

gpytorch/kernels/keops/rbf_kernel.py

@@ -2,11 +2,12 @@
 
 # from linear_operator.operators import KeOpsLinearOperator
 from linear_operator.operators import KernelLinearOperator
+from torch import Tensor
 
-from .keops_kernel import _lazify_and_expand_inputs, KeOpsKernel
+from .keops_kernel import _Anysor, _lazify_and_expand_inputs, KeOpsKernel
 
 
-def _covar_func(x1, x2, **kwargs):
+def _covar_func(x1: _Anysor, x2: _Anysor, **kwargs) -> _Anysor:
     x1_, x2_ = _lazify_and_expand_inputs(x1, x2)
     K = (-((x1_ - x2_) ** 2).sum(-1) / 2).exp()
     return K
@@ -40,8 +41,15 @@ class RBFKernel(KeOpsKernel):
 
     has_lengthscale = True
 
-    def forward(self, x1, x2, **kwargs):
+    def forward(self, x1: Tensor, x2: Tensor, diag: bool = False, **kwargs) -> KernelLinearOperator:
         x1_ = x1 / self.lengthscale
         x2_ = x2 / self.lengthscale
         # return KernelLinearOperator inst only when calculating the whole covariance matrix
-        return KernelLinearOperator(x1_, x2_, covar_func=_covar_func, **kwargs)
+        res = KernelLinearOperator(x1_, x2_, covar_func=_covar_func, **kwargs)
+
+        # TODO: diag=True mode will be removed with the GpyTorch 2.0 PR to remove LazyEvaluatedKernelTensor
+        # (it will be replaced by a `_symmetric_diag` method for quickly computing the diagonals of symmetric matrices)
+        if diag:
+            return res.diagonal(dim1=-1, dim2=-2)
+        else:
+            return res

gpytorch/test/base_keops_test_case.py

@@ -42,6 +42,11 @@ def test_forward_x1_eq_x2(self, ard=False, use_keops=True, **kwargs):
                 k2 = kern2(x1, x1).to_dense()
                 self.assertLess(torch.norm(k1 - k2), 1e-4)
 
+                # Test diagonal
+                d1 = kern1(x1, x1).diagonal(dim1=-1, dim2=-2)
+                d2 = kern2(x1, x1).diagonal(dim1=-1, dim2=-2)
+                self.assertLess(torch.norm(d1 - d2), 1e-4)
+
         if use_keops:
             self.assertTrue(keops_mock.called)
 
@@ -68,6 +73,11 @@ def test_forward_x1_neq_x2(self, use_keops=True, ard=False, **kwargs):
                 k2 = kern2(x1, x2).to_dense()
                 self.assertLess(torch.norm(k1 - k2), 1e-3)
 
+                # Test diagonal
+                d1 = kern1(x1, x1).diagonal(dim1=-1, dim2=-2)
+                d2 = kern2(x1, x1).diagonal(dim1=-1, dim2=-2)
+                self.assertLess(torch.norm(d1 - d2), 1e-4)
+
         if use_keops:
             self.assertTrue(keops_mock.called)