Add support for subclassing cupy.ndarray

Hi!

Many projects (including mine) use subclasses of numpy.ndarray, like this:

class Field(np.ndarray):
    def __new__(cls, data, prefix=1):
        data = np.asarray(data)
        field = data.view(cls)
        field.prefix = prefix
        return field

    def __array_finalize__(self, field):
        if field is None:
            return
        self.prefix = getattr(field, 'prefix', 1)

If one tries to use cupy.ndarray as a numpy replacement, then at creation of the Field instance an error is raised:

AttributeError: 'cupy.core.core.ndarray' object has no attribute 'prefix'

Are there plans to support in cupy subclassing similar to numpy approach? Maybe some workarounds are possible currently?

Regards, Marek

[takagi]

I’m trying to realize cupy.ndarray subclassing by workaround1 in the gist The following TODO is for managing its progress.

• Direct constructor call #6716
• View casting #6720
• New from template
  • ufuncs
  • ndarray methods
  • routines (this part includes many things and is to be expanded)
• documentation

[takagi]

Yes.