Search code examples
python-3.xlightgbmpmml

verbose parameter not working in PMML pipeline

I am trying to put my gbdt+lr pipeline into PMML pipeline. Followed instruction here, this is my test code:

gbm = lgb.LGBMClassifier(n_estimators=100)
clf = GBDTLRClassifier(gbm, LogisticRegression(penalty='l2'))
pipeline = PMMLPipeline([('classifier', clf)])
pipeline.fit(x, y, classifier__gbm_early_stopping_rounds=5, classifier__gbm_eval_set=[(xv, yv)], classifier__gbm_eval_metric='logloss', classifier__gbm_verbose = 2)

During fitting period, the verbose parameter not working. Same setting works well in simple lightgbm classifier. Do anyone have idea about why verbose not working?

Solution

  • Do anyone have idea about why verbose not working?

    Does Scikit-Learn issue any warnings or errors? Such as that "'classifier' does not have 'gbm' attribute"?

    You can take a look at GBDTLRClassifier source code here.

    The attribute that holds GBDT classifier is called GBDTLRClassifier.gbdt (instead of GBDTLRClassifier.gbm). Therefore, if you prefix your fit parameters with classifier__gbdt (instead of classifier__gbm), then everything should work fine:

    pipeline = PMMLPipeline([('classifier', clf)])
pipeline.fit(x, y, classifier__gbdt__verbose = 2)