In model_builder, _validate_data changes input type about pymc-experimental HOT 1 OPEN

It is intentional, but may not be the best way forward. When I worked on ModelBuilder, I was focused on getting it to be able to integrate into a scikit-learn-based Pipeline. It's been a while, but if I remember right, scikit-learn transformers and estimators accept a variety of input types and then "normalize" them to numpy arrays for the computations. So that was the idea of _validate_data. All the base _validate_data does is call check_array or check_X_y from sklearn.utils.validation. I suppose it would be just as easy to call the validation function directly rather than through _validate_data. In my own fork of this code, I actually don't use _validate_data or check_array or check_X_y at all. In my use, the input is always pandas objects rather than numpy arrays. I have to call set_output(transform='pandas') on the transformers in the pipeline, but it works for me.

From comments in some other issues (#196, #246), it seems there is a desire to refactor ModelBuilder to suit a more general use case, and then specialize it for scikit-learn as an additional class/sub-class. Seems like a good idea. There are other API's besides scikit-learn, for example sktime, which also might be useful to work with ModelBuilder-based PYMC models.

from pymc-experimental.