subclassing ndarray

Syntax#

def __array_prepare__(self, out_arr: ndarray, context: Tuple[ufunc, Tuple, int] = None) -> ndarray: # called on the way into a ufunc
def __array_wrap__(self, out_arr: ndarray, context: Tuple[ufunc, Tuple, int] = None) -> ndarray: # called on the way out of a ufunc
__array_priority__: int # used to determine which argument to invoke the above methods on when a ufunc is called
def __array_finalize__(self, obj: ndarray): # called whenever a new instance of this class comes into existence, even if this happens by routes other than __new__

Tracking an extra property on arrays

class MySubClass(np.ndarray):
    def __new__(cls, input_array, info=None):
        obj = np.asarray(input_array).view(cls)
        obj.info = info
        return obj

    def __array_finalize__(self, obj):
        # handles MySubClass(...)
        if obj is None:
            pass

        # handles my_subclass[...] or my_subclass.view(MySubClass) or ufunc output
        elif isinstance(obj, MySubClass):
            self.info = obj.info

        # handles my_arr.view(MySubClass)
        else:
            self.info = None

    def __array_prepare__(self, out_arr, context=None):
        # called before a ufunc runs
        if context is not None:
            func, args, which_return_val = context

        return super().__array_prepare__(out_arr, context)

    def __array_wrap__(self, out_arr, context=None):
        # called after a ufunc runs
        if context is not None:
            func, args, which_return_val = context

        return super().__array_wrap__(out_arr, context)

For the context tuple, func is a ufunc object such as np.add, args is a tuple, and which_return_val is an integer specifying which return value of the ufunc is being processed