Support for sklearn2pmml generated PMML

[asafschers]

Support for sklearn2pmml generated PMML

[annebyrne]

This has fixed issue, and score works fine now. Thanks for your swift help @asafschers !

[asafschers]

Modified model_factory to identify random forest models by random forest text in extension element and released 0.3.2, should solve the issue for now.
@annebyrne @harrysalmon - Are there any other issues? scoring seems to work for me
@vruusmann - seems like miningModel element should have modelName="randomForest_Model" like the R version, what do you think?

[harrysalmon]

Thank you @asafschers

[asafschers]

Hey

1. I can't reproduce the issue - the model is loaded when I run -

Scoruby.load_model('spec/fixtures/sklearn_format.txt')

2. I think the model isn't loaded correctly because of the 10th TreeModel element (in Segment id="10") in your PMML file is missing a space between the attributes algorithmName and missingValueStrategy, when I added the space the .score error stopped and the model scored - I should add an indicative error message.

Please tell me if it works
Thanks!

[vruusmann]

your PMML file is missing a space between the attributes algorithmName and missingValueStrategy

@annebyrne - Can you confirm that you have not edited this PMML file manually? That is, you have no idea how this space character went missing?

[annebyrne]

Apologies - I had uploaded by accident the version I edited to test for the discrepancies in the XML. Will attach the original, trimmed version shortly.

[harrysalmon]

Thank you for looking into this @vruusmann and @asafschers. I've generated example PMML and included the python script to reproduce.

example_sklearn_rf.pmml.txt
example_sklearn_rf.py.txt

[annebyrne]

Brill, thanks for the quick movement on this @asafschers. Will update issue shortly on whether this resolves our issue.

[vruusmann]

@asafschers The MiningModel@modelName attribute is optional free-form text. It is not suitable for identifying the actual scoring algorithm (eg. RF vs. GBT) - please consider using the MiningModel/Segmentation@multipleModelMethod attribute for that (will be (weighted)average for RF-family of algorithms, and (weighted)sum for GBT-family of algorithms).

[asafschers]

Great, I'll use multipleModelMethod.
Thanks!