Implement NumPy array interface protocol about polars HOT 3 CLOSED

After some research I do not think it is worth implementing the protocol. It is not a good match for describing arrow data, since it does not support bitpacked booleans. Anything with nulls becomes impossible to describe.

We have now optimized our __array__ implementation to provide zero copy conversions where possible. I think that should suffice for integrations.

If someone can make a strong case for implementing __array_interface__, I will consider reopening this.

from polars.

This means one could eventually to import torch; torch.tensor(dataframe), am I right? But it would still be up to Polars dev team to minimise the number of copies made, etc.

from polars.

This means one could eventually to import torch; torch.tensor(dataframe), am I right? But it would still be up to Polars dev team to minimise the number of copies made, etc.

Actually, torch.tensor does not seem to be using the NumPy protocol. We implement __array__ but it still cannot take a DataFrame. You have to explicitly export to NumPy first, e.g. torch.tensor(df.to_numpy()) or use our new functionality df.to_torch().

from polars.