Open spinkney opened 3 years ago
I rewrote the issue as an issue and tagged it as a feature and good first issue, but it's still missing all the boilerplate.
Is it faster or negligibile to just autodiff through the diag version?
A reverse specialization would be faster and more memory efficient. It probably hasn't been optimized because it's mainly used for rescaling a covariance matrix, which we'd rather do in the Cholesky space for efficiency.
Provide a reverse-mode specialization for
quad_form_diag
.