ENH improve HGBT predict classes #27844
Conversation
thomasjpfan
left a comment
There was a problem hiding this comment.
Codewise this looks more efficient. Did you find a measurable performance benefit?
In any case, LGTM.
betatim
left a comment
There was a problem hiding this comment.
Looks good to me. I assume it is faster because it does less work (we save the detour via self._loss).
One thing that would have helped me understand why this is an equivalent thing to do is more documentation on what _raw_predict returns. Mostly because my editor couldn't help me find the implementation of self._loss.predict_proba. There is a _loss.c in the directory, but no _loss.pyx, so I guess the actual code for this comes from sklearn/_loss/loss.py?? (you see confusion reigns supreme :D)
|
@betatim Think GLMs: You compute a raw prediction in "link space". This is called linear predictor for GLM, just |
Co-authored-by: Guillaume Lemaitre <[email protected]>
…scikit-learn into hgbt_predict_class
|
@jeremiedbb it seems this has not been merged for the 1.4. Would you be inclined in including it in the 1.4 or should we postpone for 1.5? |
|
The whatsnew entry needs to be moved to 1.5, right? |
|
Indeed. We need to. I'll make the fix. |
Co-authored-by: Guillaume Lemaitre <[email protected]>
Reference Issues/PRs
This PR avoids the call to
predict_probawhen executingpredictin HGBT.What does this implement/fix? Explain your changes.
Any other comments?