I think we should continue to have arviz_stats.xyz available for all functions. The functions will continue to be implemented as DataArray level, and accessors should continue to be available, but I think the main thing exposed to the users should be the arviz_stats.xyz functions.
The main reason for this is there are functions that can only work by taking the whole InferenceData because they need data from multiple groups
secondly, we should support more than dataarray/dataset/datatree as inputs whenever possible, we could look into accepting arrays and dispatching to array_stats class instead of dataarray_stats one (though they have quite different APIs so not sure) but the key thing we should support are xarray groupby objects
third, this will allow defining all 3 accessors once, with all the methods using *args, **kwargs and changing only the self._apply method (already implemented in all of them)
fourth, this also clarifies where things should be documented, in the arviz_stats.xyz function. The accessors will not have the same signature (and only the datatree one would be able to either way) but it should be possible to have them pull the docstring from the top level function and remove the arguments that no longer make sense (i.e. group only makes sense for DataTree and var_names+filter_vars only makes sense for DataTree or Dataset)
amaloney
commented
1 week ago
I think we should continue to have
arviz_stats.xyzavailable for all functions. The functions will continue to be implemented asDataArraylevel, and accessors should continue to be available, but I think the main thing exposed to the users should be thearviz_stats.xyzfunctions.array_statsclass instead ofdataarray_statsone (though they have quite different APIs so not sure) but the key thing we should support are xarray groupby objects*args, **kwargsand changing only theself._applymethod (already implemented in all of them)arviz_stats.xyzfunction. The accessors will not have the same signature (and only the datatree one would be able to either way) but it should be possible to have them pull the docstring from the top level function and remove the arguments that no longer make sense (i.e.grouponly makes sense for DataTree andvar_names+filter_varsonly makes sense for DataTree or Dataset)