ad18add8ae7da04213813b5675b903bbf4be1ebd,gpytorch/lazy/kronecker_product_added_diag_lazy_tensor.py,KroneckerProductAddedDiagLazyTensor,_root_decomposition,#KroneckerProductAddedDiagLazyTensor#,192

Before Change

        lt = self.lazy_tensor
        if isinstance(self.diag_tensor, KroneckerProductDiagLazyTensor):
            if all(isinstance(tdiag, ConstantDiagLazyTensor) for tdiag in dlt.lazy_tensors):
                sub_evals, sub_evecs = [], []
                for lt_, dlt_ in zip(lt.lazy_tensors, dlt.lazy_tensors):
                    evals_, evecs_ = lt_.symeig(eigenvectors=True)
                    sub_evals.append(DiagLazyTensor(evals_ / dlt_.diag_values))
                    sub_evecs.append(evecs_)

After Change

        if isinstance(self.diag_tensor, KroneckerProductDiagLazyTensor):
            if all(isinstance(tdiag, ConstantDiagLazyTensor) for tdiag in dlt.lazy_tensors):
                evals_p_i, evecs = _constant_kpadlt_constructor(lt, dlt)
                evals_p_i_root = DiagLazyTensor(evals_p_i.diag().sqrt())
                // here we need to scale the eigenvectors by the constants as
                // A = D^{1/2} Q (\kron a_i^{-1} \Lambda_i + I) Q^\top D^{1/2}
                // so that we compute
                // L = D^{1/2} Q (\kron a_i^{-1} \Lambda_i + I)^{1/2}

Instances