def test_inv_quad_many_vectors(self):
# Forward pass
flattened_mats = self.mats_clone.view(-1, *self.mats_clone.shape[-2:])
actual_inv_quad = (
torch.cat([mat.inverse().unsqueeze(0) for mat in flattened_mats])
.view(self.mats_clone.shape)
.matmul(self.vecs_clone)
.mul(self.vecs_clone)
.sum(-2)
.sum(-1)
)
with gpytorch.settings.num_trace_samples(2000):
non_lazy_tsr = NonLazyTensor(self.mats)
res_inv_quad = non_lazy_tsr.inv_quad(self.vecs)
self.assertEqual(res_inv_quad.shape, actual_inv_quad.shape)
self.assertLess(torch.max((res_inv_quad - actual_inv_quad).abs()).item(), 1e-1)
# Backward
inv_quad_grad_output = torch.randn(2, 3, dtype=torch.float)
actual_inv_quad.backward(gradient=inv_quad_grad_output)
res_inv_quad.backward(gradient=inv_quad_grad_output, retain_graph=True)
self.assertLess(torch.max((self.mats_clone.grad - self.mats.grad).abs()).item(), 1e-1)
self.assertLess(torch.max((self.vecs_clone.grad - self.vecs.grad).abs()).item(), 1e-1)
def test_inv_quad_many_vectors(self):
# Forward pass
actual_inv_quad = self.mat_clone.inverse().matmul(self.vecs_clone).mul(self.vecs_clone).sum()
with gpytorch.settings.num_trace_samples(1000):
non_lazy_tsr = NonLazyTensor(self.mat)
res_inv_quad = non_lazy_tsr.inv_quad(self.vecs)
self.assertAlmostEqual(res_inv_quad.item(), actual_inv_quad.item(), places=1)
# Backward
actual_inv_quad.backward()
res_inv_quad.backward(retain_graph=True)
self.assertLess(torch.max((self.mat_clone.grad - self.mat.grad).abs()).item(), 1e-1)
self.assertLess(torch.max((self.vecs_clone.grad - self.vecs.grad).abs()).item(), 1e-1)
