Closed jolars closed 2 years ago
This PR improves update for the sparse case. Previously we were reducing XCsC into a dense vector before computing Lj and XT(y−Xβ), which was not very effective.
Improvements are sometimes considerable.
I'm not really sure actually why news20 doesn't benefit as much. Hm.
I kept the old code for reference for the moment, but we should clean it up before eventually merging.
Amazing @jolars . Thanks
I think we should merge #40 before #39 and I'll keep working on improving the cached version for sparse matrices eventually.
before merging we also have to call the function for the oracle solver when ready.
I've updates the oracle solver too now.
Great, Thank you very much!
This PR improves update for the sparse case. Previously we were reducing $X_C s_C$ into a dense vector before computing $L_j$ and $X^T (y - X\beta)$, which was not very effective.
Improvements are sometimes considerable.
I'm not really sure actually why news20 doesn't benefit as much. Hm.
I kept the old code for reference for the moment, but we should clean it up before eventually merging.