Feature request: Protocol for converting something to a pandas DataFrame

[amueller]

This is inspired by the discussion in scikit-learn/enhancement_proposals#25.

NumPy defines an __array__ protocol that allows developers to implement classes that can be converted to an array by calling np.asarray() . That makes it easy to have a common interface between libraries and it's heavily used by pandas and sklearn.

It would be great to have a similar protocol for converting something to a pandas DataFrame. The goal would be to allow users to pass other data structures to libraries that expect a dataframe, say seaborn, as long as the data structures allow conversion to pd.DataFrame.

A workaround is for the developer of the new datastructure to provide an .asframe method, but that creates friction and requires the users to know what data type a particular library or function expects. If instead the developer of the datastructure can declare that conversion to a dataframe is possible, the library author (say seaborn) can request conversion to dataframe in a unified manner.

The implementation of this is probably pretty simple as it requires "only" a special case in pd.DataFrame.__init__. The main work is probably in adding it to developer documentation and publicizing it correctly.

cc @jorisvandenbossche

[TomAugspurger]

Sorry this was ignored @amueller.

I'd be happy to have a hasattr check on the data argument in DataFrame.__init__ (and probably Series.init) that simply forwards through all the relevant arguments to give control back to the library.

Our constructors are already quite complex so I might be missing something, but I'm hopeful that this is doable.

[amueller]

No worries. Not sure @GaelVaroquaux saw it either before opening the discussion.

There's probably subtleties in the implementation but I think the main question is whether this is something the ecosystem wants.

[GaelVaroquaux]

Sorry, I didn't see it. However, I think that it's useful to have the discussion in a wider setting than pandas. Though I am afraid that this discussion is going off track with complexity (out-of-core computation...) rather than on the 80/20 tradeoffs.

[thomasjpfan]

Linking this issue to PR: #46141 which introduces the dataframe exchange protocol to pandas.