WIP: generalize categorical to N-dimensions

[shoyer]

No description provided.

[shoyer]

closing until I make more progress

[jankatins]

is is more used like a _constructor(...).

[jankatins]

Maybe put in the _finalize... thingies which are used in pandas/numpy? Each method would need a change to call finalize then on new categoricals...

[jreback]

@shoyer so you want an n-dim codes array (rather than 1-d). How would this be used?
conceptually this is the 'same' as a DataFrame of categorical series (that maybe share levels, for 2-d). What is your usecase for this (aside from xray)?

[shoyer]

@jreback I expect that allowing for n-dimensional categoricals could lead to much higher performance for a DataFrame with multiple columns of the same type, in the same way that multi-dimensional arrays are positive for performance for other dtypes.

e.g., you could call unstack on a Categorical series essentially for free (since it's just a numpy reshape under the covers). (though I'm not sure enough about the pandas Block system to be sure about this)

[jreback]

@shoyer ok, that is reasonable. And in fact right now Categoricals are kept in completely separate blocks regardless of their internal structure. It is possible to combine them, so your soln would make sense.