MLPClassifier mis-interprets label (multiclass vs multioutput) about sklearn2pmml

[vruusmann]

Appreciate your reproducible example - helps me to see this mysterious ClassCastException myself:

WARNING: Attribute 'sklearn2pmml.pipeline.PMMLPipeline.target_fields' is not set. Assuming [y1, y2, y3, y4, y5] as the name of target fields
Exception in thread "main" java.lang.ClassCastException: java.lang.Integer cannot be cast to java.util.List
        at sklearn2pmml.pipeline.PMMLPipeline.initLabel(PMMLPipeline.java:581)
        at sklearn2pmml.pipeline.PMMLPipeline.encodePMML(PMMLPipeline.java:130)
        at com.sklearn2pmml.Main.run(Main.java:91)
        at com.sklearn2pmml.Main.main(Main.java:66)

[vruusmann]

For an unknown reason, the SkLearn2PMML has trouble figuring out the intent of the classification task. It currently thinks that it's dealing with a multioutput task (ie. five independent labels). The correct thinking would be a multiclass task (a single label with five independent classes).

The problem can be fixed by wrapping the y(_train) from Numpy array to Pandas series, and calling PMMLPipeline.fit(X, y) with it (instead of MLPClassifier.fit(X, y)):

from pandas import Series

# THIS!
y_train = Series(y_train, name = "y")

model = MLPClassifier(hidden_layer_sizes=(10,5,2), max_iter=1000)
# Don't do MLPClassifier.fit(X, y)
#model.fit(X_train, y_train)

pmml_pipeline = PMMLPipeline([	
 	("classifier", model)
])
# THIS!
pmml_pipeline.fit(X_train, y_train)

sklearn2pmml(pmml_pipeline, 'model01.pmml')

These two changes together help SkLearn2PMML to understand that it's dealing with a multiclass classification problem (one label, five category levels).

[algarnims]

Thank you very much for your quick response. We hope to see this resolved in the future which will improve the package overall.

[vruusmann]

The SkLearn2PMML package is fine. It is your pipeline that needs upgrading.

Most importantly, you should be calling PMMLPipeline.fit(X, y) always (as opposed to calling Model.fit(X, y) and then wrapping it into a dummy PMMLPipeline object), because that will set the PMMLPipeline.active_fields and PMMLPipeline.target_fields attributes to meaningful values. This way, the SkLearn2PMML package does not need to make guesses about the nature of the classification problem.

Look at your own SkLearn2PMML console log that precedes the error. It clearly states that SkLearn2PMML makes an assumption that there must be five labels present (ie. multioutput).

[vruusmann]

The fix will probably be in the form of making the "has the PMMLPipeline object been fitted or not" check more strict.

So, instead of making a guess, it would raise a clear "the PMMLPipeline object is not fitted" error.