Not working jpmml-converter version with GradientBoostingClassifier about sklearn2pmml

[vruusmann]

This is a very-very odd bug report. The SkLearn2PMML project is based on the JPMML-SkLearn library, which includes two integration test cases for GradientBoostingClassifier:
https://github.com/jpmml/jpmml-sklearn/blob/master/src/test/java/org/jpmml/sklearn/ClassifierTest.java#L57-L60
https://github.com/jpmml/jpmml-sklearn/blob/master/src/test/java/org/jpmml/sklearn/ClassifierTest.java#L147-L150

If the JPMML-SkLearn library was generating incorrect PMML markup for the second segment, then these integration test cases should cause the build process to error out.

The working copy of JPMML-SkLearn already includes JPMML-Converter library version 1.2.4:
jpmml/jpmml-sklearn@79a4547

So, you need to wait a couple more days until SkLearn2PMML version 0.20.4 is released, which shall be based on the latest and greatest JPMML-SkLearn library version.

[sabba]

Hi Villu,
yes is very odd!

In order to understand the issue i checked out JPMML-SkLearn to version 1.3.3 which depend of jpmml-converter-1.2.3 and te full test suite run green.

Then, I tried to export a PMML document using the converter-executable-1.3.3.jar produced in "target" with the following command and the tree was surprising right!

java -jar converter-executable-1.3.3.jar --pkl-pipeline-input ..\src\test\resources\pkl\GradientBoostingAudit.pkl --pmml-output GradientBoostingAudit.pmml

Could be the case that the jar included statically as jpmml-converter-1.2.3 in SkLearn2PMML was not generated from the right code base?

I'va have installed SkLearn2PMML using pip pointing to the git repository.

Thank you for the support!

[vruusmann]

Could be the case that the jar included statically as jpmml-converter-1.2.3 in SkLearn2PMML was not generated from the right code base?

I've experienced a situation, where the resources directory of an installed sklearn2pmml package contains several versions of the same library JAR file (eg. JPMML-Converter versions 1.2.0, 1.2.3). The JVM then adds the first out-of-date JAR file (eg. 1.2.0) to classpath, and ignores the second up-to-date JAR file (eg. 1.2.3). Such classpath conflicts may lead to all sorts of unexpected outcomes, starting with JVM class loader/verifier errors and ending with incorrectly generated PMML markup.

Could it be that you had multiple versions of the JPMML-Converter library in the resources directory? This could happen if you install a new version of the sklearn2pmml package, without properly uninstalling the old version first.

Anyway, a good "fix" for library errors is to 1) delete the sklearn2pmml subdirectory from the site-packages directory of your active Python installation, and 2) install a the latest version of the sklearn2pmml package from GitHub.

[sabba]

Could it be that you had multiple versions of the JPMML-Converter library in the resources directory? This could happen if you install a new version of the sklearn2pmml package, without properly uninstalling the old version first.

I think is not my case because i never installed before the package... BUT, In order to have locally this pull request (that you merged today) jpmml/jpmml-sklearn#46 I have put a recompiled version of the jpmml-sklearn jar inside my resources directory.

Could be this the root cause?

[vruusmann]

I have put a recompiled version of the jpmml-sklearn jar inside my resources directory.

For documentation purposes, these are the steps for updating the JPMML-SkLearn dependency of the SkLearn2PMML package:

1. Update JPMML-SkLearn version in pom.xml. When working with locally tweaked JPMML-SkLearn library, then the version identifier is typically 1.3-SNAPSHOT.
2. Manually delete all JAR files from the /sklearn2pmml/resources directory (only __init__.py should remain).
3. Execute Apache Maven build mvn clean package. This will download the JPMML-SkLearn library, and all its transitive dependencies into the /sklearn2pmml/resources directory.
4. Optionally, update the version identifier in sklearn2pmml/metadata.py. For example, introduce some sort of suffix to signal that this is a "tweaked" version, not an official version.

[sabba]

Ok! Thank You.