Closed ExpandingMan closed 1 year ago
This should also now have fixed a much worse issue in which a fix to the predict
output in XGBoost.jl was resulting in mangled output here. This is because I had originally worked around that issue here. How I'm such an idiot that I did all this without realizing it had to be comprehensively fixed I don't know. Definitely in part due to my not taking classification seriously enough.
Not going to waste any time merging and tagging this since the issues were pretty egregious.
Not going to waste any time merging and tagging this since the issues were pretty egregious.
Not sure I understand. Are you saying the issue is too important to wait for a review?
Sounds like we may need some extra tests to catch what we had appeared to overlooked prior to this PR?
Are you saying the issue is too important to wait for a review?
Just that there was basically no way it was going to be worse off than before. :shrug:
Sorry if this was overzealous. Tell you what, I'll wait for you to merge from now on.
Sounds like we may need some extra tests to catch what we had appeared to overlooked prior to this PR?
The real problem was in XGBoost.jl and tests were added.
Sorry if this was overzealous.
I'm not too bothered. I do really appreciate you're taking on some much of the maintenance of this package (and XGBoost.jl). However, I'm still happy and generally available to provide reviews, which I think is helpful and good practice.
Please go ahead and tag a new release, it looks good to me.
Ok, I did so earlier.
Yeah, I agree, it's not a good practice, I was just really eager to get if fixed as I was really bothered by it. I'll refrain from doing it in the future.
There are some shenanigans going on to check the
watchlist
argument to the model, and prior to this PR in some cases it would get passed asnothing
(an invalid value). This should fix #28. I think this ought to be further improved but at the moment it's not clear to me what a "good" design would be here. Might also call for changes in XGBoost.jl since thewatchlist
argument seems to cause a lot of confusion.