Getting ValueError: axes don't match array while using merge_ebms on multi labels ExplainableBoostingClassifier

[lbalec]

I'm testing merging multi labels ExplainableBoostingClassifier models with merge_ebms method (interpret==0.3.0 from pypi on python 3.7 (my stage env) and 3.8 (clean conda env from scratch only for 0.3.0 version)

models = []
for i in range(10):
    _model_clf = ExplainableBoostingClassifier(random_state=i, n_jobs=-1, interactions=[])
    _model_clf.fit(X_train[features], X_train[target])
    models.append(_model_clf)

and I'm getting following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/tmp/ipykernel_40170/2566246721.py in <module>
----> 1 model_clf = merge_ebms(models)

~/.conda/envs/image-processing/lib/python3.7/site-packages/interpret/glassbox/ebm/utils.py in merge_ebms(models)
    940                         old_mapping[model_idx],
    941                         model.bagged_scores_[term_idx][bag_idx],
--> 942                         model.bin_weights_[term_idx] # we use these to weigh distribution of scores for mulple bins
    943                     )
    944                     new_bagged_scores.append(harmonized_bagged_scores)

~/.conda/envs/image-processing/lib/python3.7/site-packages/interpret/glassbox/ebm/utils.py in _harmonize_tensor(new_feature_idxs, new_bounds, new_bins, old_feature_idxs, old_bounds, old_bins, old_mapping, old_tensor, bin_evidence_weight)
    387     old_tensor = old_tensor.transpose(tuple(axes))
    388     if bin_evidence_weight is not None:
--> 389         bin_evidence_weight = bin_evidence_weight.transpose(tuple(axes))
    390 
    391     mapping = []

ValueError: axes don't match array

I've tried setup different parameters (binning, inner and outer bags, explicitly pointing the features names and types) and the result is always the same - ValueError: axes don't match array. I'm getting this error only with multilabel problem, merging binary classifiers works perfectly fine. Any help here?

[paulbkoch]

Hi @lbalec -- Thanks for reporting this. Lack of multiclass support for merging is an oversight on my part. I'll have a look at this and fix it after the holidays. I'm assuming you meant multiclass here since we do not support multilabel with EBMs.

[lbalec]

@paulbkoch yes, I meant multiclass ofcourse :) Thank you

[paulbkoch]

Hi @lbalec -- multiclass merging is now supported. The changes are in the develop branch for now and should be available in the next release, which will be the release after 0.3.2.