Describe the workflow you want to enable

I have a subclass of numpy.ndarray which keeps track of axis labels: which axes represent time, spatial directions, independent samples, and response variables. I have a transform() step in my pipeline that runs
x = check_array(x, accept_sparse=("csr", "csc"), ...).
It's nice that this allows me to handle both numpy arrays and sparse inputs . However, this line of scipy.utils.validation.check_array downcasts anything that isn't a sparse array to a numpy array. This is because numpy.asarray does not allow interception via __array_ufunc__ or __array_function__, so subclasses cannot dictate how to handle its behavior.

Describe your proposed solution

isinstance check to see if array is already a numpy array. Not sure exactly how to enforce order and dtype, however.

Describe alternatives you've considered, if relevant

User code instead repeats validation code from check_array. In my case, that means importing the private function _ensure_sparse_format() and skipping or copypasting the other validation checks.

Additional context

No response