The result of merge_ebm is not a valid classifier, as attributes are missing. The repr raises an AttributeError which makes it horrible to debug.
The following is a minimal reproducible example:
from sklearn.datasets import load_iris
from interpret.glassbox import ExplainableBoostingClassifier, merge_ebms
X, y = load_iris(return_X_y=True)
clf1 = ExplainableBoostingClassifier(interactions=0, outer_bags=2)
clf1.fit(X,y)
clf2 = ExplainableBoostingClassifier(interactions=0, outer_bags=2)
clf2.fit(X,y)
clf = merge_ebms([clf1, clf2])
repr(clf)
Which results in the error AttributeError: 'ExplainableBoostingClassifier' object has no attribute 'cyclic_progress'.
A hotfix is simply copying the attributes of the first classifier in the list:
for attr, val in clf1.get_params(deep=False).items():
if not hasattr(clf, attr):
setattr(clf, attr, val)The question is, what the desired strategy is to fix this error. (I haven't delved into merge_ebms to know what's going on.) Another option would be setting the parameters where all classifiers agree, and setting all other parameters to None. This avoids immediate AttributeErrors but might cause problems later, in case None is not a valid parameter.
The result of
merge_ebmis not a valid classifier, as attributes are missing. Thereprraises anAttributeErrorwhich makes it horrible to debug.The following is a minimal reproducible example:
Which results in the error
AttributeError: 'ExplainableBoostingClassifier' object has no attribute 'cyclic_progress'.A hotfix is simply copying the attributes of the first classifier in the list:
The question is, what the desired strategy is to fix this error. (I haven't delved into
merge_ebmsto know what's going on.) Another option would be setting the parameters where all classifiers agree, and setting all other parameters to None. This avoids immediateAttributeErrors but might cause problems later, in caseNoneis not a valid parameter.