feat(polars): support to_torch

[ianmcook]

I see Polars recently added to_torch:
https://www.linkedin.com/posts/pola-rs_export-your-polars-dataframe-directly-to-activity-7193568311769542656-kqRz

I see Ibis has a to_torch API: #6518

Is the former supported via the latter?

Code of Conduct

• I agree to follow this project's Code of Conduct

[gforsyth]

It is not, currently.

We go from any backend (including polars) -> pyarrow -> numpy -> torch.from_numpy

Polars has it's own fast (er?) path from a polars frame to a numpy array, then also uses torch.from_numpy.

[ianmcook]

Got it, thanks. Yeah, I don't know if it's worthwhile to directly map Ibis to_torch to Polars to_torch for performance or other reasons.

[alexander-beedie]

Author of the Polars to_torch export here... For the curious, the advantage in a native export path wouldn't be performance (both paths require only a single copy per-array, with no conversion overheads), it would be in avoiding the pyarrow dependency. However, if you've already eaten that import cost then there's no advantage to be had by adding a native path (aside from access to export modes other than "dict").