def test_nongrid_nongrid(self):
base_kernel = RBFKernel()
data = torch.randn(5, 2)
kernel = GridKernel(base_kernel, grid)
grid_eval = kernel(data, data).evaluate()
actual_eval = base_kernel(data, data).evaluate()
self.assertLess(torch.norm(grid_eval - actual_eval), 1e-5)
def test_grid_grid(self):
base_kernel = RBFKernel()
kernel = GridKernel(base_kernel, grid)
grid_covar = kernel(grid_data, grid_data).evaluate_kernel()
self.assertIsInstance(grid_covar, KroneckerProductLazyTensor)
grid_eval = kernel(grid_data, grid_data).evaluate()
actual_eval = base_kernel(grid_data, grid_data).evaluate()
self.assertLess(torch.norm(grid_eval - actual_eval), 1.2e-5)
def test_nongrid_nongrid(self, toeplitz=True):
with gpytorch.settings.use_toeplitz(toeplitz):
base_kernel = RBFKernel(ard_num_dims=2)
data = torch.randn(5, d)
kernel = GridKernel(base_kernel, grid)
grid_eval = kernel(data, data).evaluate()
actual_eval = base_kernel(data, data).evaluate()
self.assertLess(torch.norm(grid_eval - actual_eval), 1e-5)
def test_grid_grid(self, toeplitz=True):
with gpytorch.settings.use_toeplitz(toeplitz):
base_kernel = RBFKernel(ard_num_dims=2)
kernel = GridKernel(base_kernel, grid)
grid_covar = kernel(grid_data, grid_data).evaluate_kernel()
self.assertIsInstance(grid_covar, KroneckerProductLazyTensor)
grid_eval = kernel(grid_data, grid_data).evaluate()
actual_eval = base_kernel(grid_data, grid_data).evaluate()
self.assertLess(torch.norm(grid_eval - actual_eval), 2e-5)
def test_non_stationary_base(self):
base_kernel = LinearKernel()
with self.assertRaisesRegex(
RuntimeError,
"The base_kernel for GridKernel must be stationary."):
GridKernel(base_kernel, grid)
def test_grid_grid(self):
base_kernel = RBFKernel()
kernel = GridKernel(base_kernel, grid)
grid_eval = kernel(grid, grid).evaluate()
actual_eval = base_kernel(grid_data, grid_data).evaluate()
self.assertLess(torch.norm(grid_eval - actual_eval), 1e-5)