Support gradient boosting classifiers

[iamDecode]

The long awaited gradient boosting model support is finally here.

I have found that PMML generation for GBM models is unfortunately not very consistent. I had to account for many different versions which resulted in a rather massive change.

There are some parts I am yet uncertain about:

• For a yet unknown reason scores in regression trees for xgboost have to be multiplied by 10. PMML's Target.rescaleFactor seems related. I've added a FIXME note, and it should be revisited when supporting regression trees outside of gradient boosting.
• I noticed for some PMML exports, the prediction probabilities to be slightly off compared to the original and pypmml. This seems happen at

  sklearn-pmml-model/sklearn_pmml_model/tree/tree.py

  Line 224 in 27cc103

  value -= 0.00000000001

  , as switching the - for a + fixes the probabilities for those models. However, that inevitably breaks the predictions of other models, which in my testing is way more significant. I am not sure what causes this discrepancy, but was not able to fix it just yet. However, this is such a slight effect I'm willing to accept it for now. Should reinvestigate later.

After this PR, supporting regression should be fairly straightforward.

Closes #20

[codecov]

Codecov Report

Merging #23 (6e698c0) into master (0a42d68) will increase coverage by 0.22%.
The diff coverage is 100.00%.

@@             Coverage Diff             @@
##           master       #23      +/-   ##
===========================================
+ Coverage   99.77%   100.00%   +0.22%     
===========================================
  Files          11        12       +1     
  Lines         440       559     +119     
===========================================
+ Hits          439       559     +120     
+ Misses          1         0       -1     

Impacted Files	Coverage Δ
sklearn_pmml_model/base.py	100.00% <100.00%> (ø)
sklearn_pmml_model/ensemble/__init__.py	100.00% <100.00%> (ø)
sklearn_pmml_model/ensemble/forest.py	100.00% <100.00%> (ø)
sklearn_pmml_model/ensemble/gb.py	100.00% <100.00%> (ø)
sklearn_pmml_model/linear_model/implementations.py	100.00% <100.00%> (ø)
sklearn_pmml_model/naive_bayes/implementations.py	100.00% <100.00%> (ø)
sklearn_pmml_model/tree/__init__.py	100.00% <100.00%> (ø)
sklearn_pmml_model/tree/tree.py	100.00% <100.00%> (+1.02%)	⬆️
... and 2 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0a42d68...6e698c0. Read the comment docs.

[lgtm-com]

This pull request introduces 4 alerts when merging 6df5bcb into 0a42d68 - view on LGTM.com

new alerts:

• 2 for Except block handles 'BaseException'
• 1 for Conflicting attributes in base classes
• 1 for Unused import

[lgtm-com]

This pull request introduces 1 alert when merging 63d5ca3 into 0a42d68 - view on LGTM.com

new alerts:

• 1 for Conflicting attributes in base classes

[lgtm-com]

This pull request fixes 7 alerts when merging cf9da61 into 0a42d68 - view on LGTM.com

fixed alerts:

• 6 for Conflicting attributes in base classes
• 1 for Unused import

[lgtm-com]

This pull request fixes 8 alerts when merging 6e698c0 into 0a42d68 - view on LGTM.com

fixed alerts:

• 7 for Conflicting attributes in base classes
• 1 for Unused import