`FunctionTransformer` serialization in `DataFrameMapper` failing on `numpy.log` about jpmml-sklearn

[vruusmann]

An ufunc is identified by a combination of module and name attributes. The list of "supported" modules includes numpy, numpy.core.numeric and numpy.lib.function_base:
https://github.com/jpmml/jpmml-sklearn/blob/master/src/main/java/sklearn/preprocessing/FunctionTransformer.java#L84

How does the statsmodels module fit in here? How is it related to NumPy?

And the typical questions:

1. What is your Python version?
2. What is your NumPy version?

[jrwalk]

numpy version 1.11.1, python version 3.5.2. The FunctionTransformer object is encoding the numpy.log function, which has previously worked (though I suspect I've update numpy since then). I'm basing the statsmodels surmise on the error trace -- I agree, there shouldn't need to be any usage of that package to simply return a logarithm.

[vruusmann]

Can you avoid the statsmodels dependency if you explicitly import the log function from one of the above "supported" modules?

Something like from numpy.core(.numeric) import log

[jrwalk]

tested, still throws the same error with an explicit log import (previously was simply import numpy -> numpy.log)