NGBClassifier predict error about ngboost HOT 2 CLOSED

Good catch... I looked into it and this is happening because the way predict is implemented right now is a bit silly and only works for distributions that have a loc parameter.

The way to fix this, imo, is to define a predict method for all distributions that returns the conditional mean (i.e. the loc parameter for those we've implemented) for continuous distributions and the conditional mode (i.e. the most likely class) for discrete distributions. We can probably even use the underlying scipy distributions for this. I'll do that sometime soon.

In the meantime, you can easily get the predicted class by doing ngb.pred_dist(X_test).prob > 0.5.

from ngboost.

Ok, I've fixed this issue with c054851. Please do play around with it and let me know if this now works for you and/or if you want other features.

from ngboost.