Steps toward full compatibility with sklearnsklearn.utils.validation.check_array: missing default __array_namespace__ and dpt.isfinite

[fcharras]

Some work has been started on sklearn to interface estimators that are fully written using only typical array functions with array libraries that implement the Array API, see scikit-learn/scikit-learn#22554

In this effort, the default data validation function sklearn.utils.validation.check_array has been adapted to all array libraries that implement the array API.

For the plugin system we're considering at scikit-learn/scikit-learn#24497 along with our plugin at https://github.com/soda-inria/sklearn-numba-dpex I've found that it could be interesting to re-use sklearn.utils.validation.check_array on usm_ndarray inputs with sklearn validation rules and it might also prevent unnecessary data copies.

I've found that currently a usm_ndarray will fail the check_array for those two reasons:

• requires dpctl.tensor.isfinite to be implemented
• requires the .__array_namespace__ attribute of usm_ndarray array to return dpctl.tensor rather than None

In the meantime it's possible to work around those two missing features with:

import dpctl.tensor as dpt
from sklearn.utils.validation import check_array
from sklearn import config_context

array = dpctl.tensor.asarray(np.arange(12).reshape(4,3))
with config_context(array_api_dispatch=True, assume_finite=True):    # workaround 1: assume_finite=True
    tensor._set_namespace(dpt)    # workaround 2: manually call ._set_namespace
    checked_array = check_array(array, accept_sparse=False, dtype=[np.float32, np.float64], order="C", copy=False, force_all_finite=True)

[oleksandr-pavlyk]

Since gh-1324, dpctl.tensor.usm_ndarray.__array_namespace__ returns dpctl.tensor by default.

dpctl.tensor.isfinite has been implemented for some time as well:

In [1]: import dpctl.tensor as dpt

In [2]: x = dpt.ones(10, dtype=dpt.float64)

In [3]: x
Out[3]: usm_ndarray([1., 1., 1., 1., 1., 1., 1., 1., 1., 1.])

In [4]: dpt.isfinite(_)
Out[4]:
usm_ndarray([ True,  True,  True,  True,  True,  True,  True,  True,
              True,  True])

In [5]: x + 2
Out[5]: usm_ndarray([3., 3., 3., 3., 3., 3., 3., 3., 3., 3.])

In [6]: 'isfinite' in dir(x.__array_namespace__())
Out[6]: True

[oleksandr-pavlyk]

@fcharras Ok to close?

[fcharras]

Yes, thank you @oleksandr-pavlyk