6bca4eed74e22ee369c166649f15db56f8565e35,test/lazy/_lazy_tensor_test_case.py,LazyTensorTestCase,test_inv_quad_logdet_no_reduce,#LazyTensorTestCase#,586

Before Change

                [torch.logdet(flattened_evaluated[i]).unsqueeze(0) for i in range(lazy_tensor.batch_shape.numel())]
            ).view(lazy_tensor.batch_shape)

            diff_invq = (res_inv_quad.sum(-1) - actual_inv_quad).abs() / actual_inv_quad.abs().clamp(1, math.inf)
            diff_logdet = (res_logdet - actual_logdet).abs() / res_logdet.abs().clamp(1, math.inf)
            self.assertLess(diff_invq.max().item(), 0.01)
            self.assertLess(diff_logdet.max().item(), 0.3)

After Change

            vecs_copy = vecs.clone().detach_().requires_grad_(True)

            with gpytorch.settings.num_trace_samples(128):
                with gpytorch.settings.max_cholesky_numel(0), gpytorch.settings.cg_tolerance(1e-5):
                    res_inv_quad, res_logdet = lazy_tensor.inv_quad_logdet(
                        inv_quad_rhs=vecs, logdet=True, reduce_inv_quad=False
                    )

            actual_inv_quad = evaluated.inverse().matmul(vecs_copy).mul(vecs_copy).sum(-2).sum(-1)
            actual_logdet = torch.cat(
                [torch.logdet(flattened_evaluated[i]).unsqueeze(0) for i in range(lazy_tensor.batch_shape.numel())]
            ).view(lazy_tensor.batch_shape)

Instances