✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 238869d. Link to the linter CI: here}

Another thought that I am 50/50 on - currently when an array that is not able to be converted via numpy.asarray we get an error like:

RuntimeError: Can not convert array on the 'array_api_strict.Device('device1')' device to a Numpy array.

What do we think about adding something along the lines of - "if you wish to enable array API support and get {namespace} arrays as output, set the configuration array_api_dispatch=True"? If they have input e.g., a cupy array, it is possible that they want to enable array API support but they just forgot/didn't know they need to set array_api_dispatch=True...

At the moment this change would be a bit complicated because we perform the asarray in different places for classification metrics vs regression metrics vs clustering metrics etc, but once we start adding support for mixed array inputs (#31829), I think the error could be raised in move_to, which would be called at the start of functions.

I think it would be beneficial to mention that array API dispatch should be activated. I think we try to handle non NumPy arrays with array_api_dispatch=False by trying to convert to NumPy for metrics but I don't think that is something that should be recommended because we can run into unexpected errors.

Yeah I agree, this seems like a niche/odd thing that we do, but we wouldn't really recommend anyone to use it. I've amended the wording.

I think it is important to document, so people know this happens, because it can be quite unexpected to get a different namespaced array than what you've input e.g., you've forgotten to set array_api_dispatch=True.

I'm not sure if the array API section is the right place for this information. I think we should have this in a more general section of the scikit-learn docs, possibly we already state it there :-/

It has happened that as part of the array API we've changed what happens when you pass torch arrays to scikit-learn. In which case we get reports about it breaking. This makes me think that it is something people rely and is possible on purpose. But I don't fully know the "rule of thumb" for what works and how. I think it is something like "if it looks enough like a Numpy array" and/or if it supports __array_function__

But I don't fully know the "rule of thumb" for what works and how. I think it is something like "if it looks enough like a Numpy array" and/or if it supports array_function

My naive guess would be arrays on CPU device can be converted. I have tried torch 'cpu' and jax (import jax.numpy as jnp; arr = jnp.array([1, 2, 3])) and both can be converted to numpy with np.asarray.

In which case we get reports about it breaking.

Interesting, can you point me to any references?

I think we should have this in a more general section of the scikit-learn docs, possibly we already state it there :-/

Tricky to look up. In the SVM user guide we state:

The support vector machines in scikit-learn support both dense (numpy.ndarray and convertible to that by numpy.asarray) and sparse (any scipy.sparse) sample vectors as input.

In "Developing scikit-learn estimators" -> "Input validation" we state:

The module sklearn.utils contains various functions for doing input validation and conversion. Sometimes, np.asarray suffices for validation;

I looked for "array " and "asarray" in our .rst files and couldn't find anything else relevant.

I am happy to add a note about this behaviour in the metrics section of the user guide and link it to here?

#29107 is an example of a bug like this.

@betatim @ogrisel any other comments on this?

I'll merge it. Olivier can comment or make a follow up PR if there is something we missed

I don't think this is worth doing something about it but

we generally recommend setting array_api_dispatch=True when using array API inputs.

I think the recommendation is quite mild so that's fine but passing a PyTorch cpu tensor has been working for ever and I guess people kind of expect the implicit conversion to numpy array ¹.

Something I did not think about until now, is that when array API becomes the default the output will change to a PyTorch tensor and there may be one workflow somewhere that breaks but it's probably quite unlikely to happen?

I think this is something to consider as part of making array API the default. In some way the pytorch on CPU -> Numpy workflow is part of the scikit-learn API, in some way you could also argue that it is undocumented behaviour :-/

in some way you could also argue that it is undocumented behaviour

It isn't explicitly stated anywhere, just implied by the glossary entry: "array-like is any type object for which :func:numpy.asarray will produce an array of appropriate shape".

Maybe I should amend this glossary entry to make it more explicit that inputs will be converted to numpy arrays, and array outputs will be numpy arrays...?

I'd not do anything for now