Nick Erickson
Nick Erickson
Can you provide full logs? This will help identify the issue. Also, consider updating to v0.5.2, it may have fixed your issue already.
Also, try ignoring KNN, that might do the trick (and please tell me if it does)
@donut1856 The foundations of this logic exist with `use_bag_holdout=True` in fit call, but we still don't support a mix of bagged and non-bagged models as of v0.5. One reason is...
Thanks for reporting! This is indeed a bug. You can bypass this bug by not using high_quality / refit_full and then calling fit_extra, then you can call refit_full on the...
> I found that predictor.refit_full() decreases accuracy. I do a fit with time_limit, then fit_extra with time_limit and then a refit_full. > Is it true or am i doing something...
Reopening as I think this requires a deeper dive to understand why this error is occurring.
> @Innixma I think it is because `pip3 install -U pip` may change the pip in Colab. I think this is the root-cause. We should not touch the pip in...
@sxjscience Except we don't use torchtext in AutoGluon, therefore we shouldn't install it. Thus we can't control its version cap, and therefore we can't fix on our end. The only...
Are you referring to the decision threshold for `predict()`? I'm uncertain what you mean when you mention `predict_proba()`. If this is what you mean, then it is always 0.5 currently....
Currently it is always 0.5 for binary or the max for multiclass. We don't adjust thresholds when making predictions.