XGBClassifier.predict_proba outputs incorrect array dimension!

[jyu-theartofml]

Hi, I'm using the following code to fit the xgb classifier:
clf = XGBClassifier(objective='multi:softprob', n_estimators=300, subsample=0.8, gamma=1.3)
clf.fit(training_d[features], training_d['outcome'], early_stopping_rounds=10, eval_set=[(training_d[features], training_d['outcome']),(validation_d[features], validation_d['outcome'])],
eval_metric='mlogloss')
but when I use the predict_proba function;
predict1=np.array(clf.predict_proba(validation_d[features]))
it spits out the wrong dimension, the validation_d[features] is (6238, 37), and the prediction should be (6238,5), but predict1 gives me (5,12476). I can't figure out where the 12476 is coming from.

Does anyone have similar issue? Thanks!

--Jenny

[KhaoticMind]

12476 = 6238 * 2

I'm getting this same behavior when using XGBClassifier from python. using the xgb native interface (objective = 'multi:softprob') and booster.predict I get correct results.

I'm running on commit 9c26566.

I can try to update to latest commit and see if the problem persists.

[KhaoticMind]

Ok! Found the issue!

The problem is here:

if self.objective == "multi:softprob":
   return class_probs
else:
   classone_probs = class_probs
   classzero_probs = 1.0 - classone_probs
   return np.vstack((classzero_probs, classone_probs)).transpose()

self.object is not being updated on the fit method (where it detects the multi:softprob):

if self.n_classes_ > 2:
   # Switch to using a multiclass objective in the underlying XGB instance
   xgb_options["objective"] = "multi:softprob"
   xgb_options['num_class'] = self.n_classes_

The simple fix should be update self.objective inside this if.
I will inplement it and do a PR latter/tomorrow.

[jyu-theartofml]

Thank you!