LogisticRegression.predict_proba is incorrect when multiclass='multinomial'.

[jagapiou]

Looking at the code for LogisticRegression.predict_proba it seems to assume that multiclass='ovr': it does sigmoid(wTx+b) and then normalizes over the classes. I think that when multiclass='multinomial' is used it should use softmax instead.

[agramfort]

indeed... that's ugly. Thanks for the bug report. any volunteer to fix this? @MechCoder @TomDLT ?

[MechCoder]

OK, I can fix this up tonight.

[jagapiou]

I think this is an issue other places too: e.g. sklearn.lda.LDA.predict_proba

[amueller]

Yes, looks to be the case.... argh :-/

[MechCoder]

This was fixed