Overload NumPy ufuncs and functions

overload_numpy provides easy-to-use tools for working with NumPy's __array_(u)func(tion)__. The library is fully typed and wheels are compiled with mypyc.

Implementing an Overload

First, some imports:

>>> from dataclasses import dataclass, fields
>>> from typing import ClassVar
>>> import numpy as np
>>> from overload_numpy import NumPyOverloader, NPArrayOverloadMixin

Now we can define a NumPyOverloader instance:

>>> W_FUNCS = NumPyOverloader()

The overloads apply to an array wrapping class. Let's define one:

>>> @dataclass
... class Wrap1D(NPArrayOverloadMixin):
...     '''A simple array wrapper.'''
...     x: np.ndarray
...     NP_OVERLOADS: ClassVar[NumPyOverloader] = W_FUNCS

>>> w1d = Wrap1D(np.arange(3))

Now both numpy.ufunc (e.g. numpy.add) and numpy functions (e.g. numpy.concatenate) can be overloaded and registered for Wrap1D.

>>> @W_FUNCS.implements(np.add, Wrap1D)
... def add(w1, w2):
...     return Wrap1D(np.add(w1.x, w2.x))

>>> @W_FUNCS.implements(np.concatenate, Wrap1D)
... def concatenate(w1ds):
...     return Wrap1D(np.concatenate(tuple(w.x for w in w1ds)))

Time to check these work:

>>> np.add(w1d, w1d)
Wrap1D(x=array([0, 2, 4]))

>>> np.concatenate((w1d, w1d))
Wrap1D(x=array([0, 1, 2, 0, 1, 2]))

ufunc also have a number of methods: 'at', 'accumulate', etc. The function dispatch mechanism in NEP13 says that "If one of the input or output arguments implements __array_ufunc__, it is executed instead of the ufunc." Currently the overloaded numpy.add does not work for any of the ufunc methods.

>>> try: np.add.accumulate(w1d)
... except Exception: print("failed")
failed

ufunc method overloads can be registered on the wrapped add implementation:

>>> @add.register('accumulate')
... def add_accumulate(w1):
...     return Wrap1D(np.add.accumulate(w1.x))

>>> np.add.accumulate(w1d)
Wrap1D(x=array([0, 1, 3]))

Dispatching Overloads for Subclasses

What if we defined a subclass of Wrap1D?

>>> @dataclass
... class Wrap2D(Wrap1D):
...     '''A simple 2-array wrapper.'''
...     y: np.ndarray

The overload for numpy.concatenate registered on Wrap1D will not work correctly for Wrap2D. However, NumPyOverloader supports single-dispatch on the calling type for the overload, so overloads can be customized for subclasses.

>>> @W_FUNCS.implements(np.add, Wrap2D)
... def add(w1, w2):
...     print("using Wrap2D implementation...")
...     return Wrap2D(np.add(w1.x, w2.x),
...                   np.add(w1.y, w2.y))

>>> @W_FUNCS.implements(np.concatenate, Wrap2D)
... def concatenate2(w2ds):
...     print("using Wrap2D implementation...")
...     return Wrap2D(np.concatenate(tuple(w.x for w in w2ds)),
...                   np.concatenate(tuple(w.y for w in w2ds)))

Checking these work:

>>> w2d = Wrap2D(np.arange(3), np.arange(3, 6))
>>> np.add(w2d, w2d)
using Wrap2D implementation...
Wrap2D(x=array([0, 2, 4]), y=array([ 6, 8, 10]))

>>> np.concatenate((w2d, w2d))
using Wrap2D implementation...
Wrap2D(x=array([0, 1, 2, 0, 1, 2]), y=array([3, 4, 5, 3, 4, 5]))

Great! But rather than defining a new implementation for each subclass, let's see how we could write a more broadly applicable overload:

>>> @W_FUNCS.implements(np.add, Wrap1D)  # overriding both
... @W_FUNCS.implements(np.add, Wrap2D)  # overriding both
... def add_general(w1, w2):
...     WT = type(w1)
...     return WT(*(np.add(getattr(w1, f.name), getattr(w2, f.name))
...                 for f in fields(WT)))

>>> @W_FUNCS.implements(np.concatenate, Wrap1D)  # overriding both
... @W_FUNCS.implements(np.concatenate, Wrap2D)  # overriding both
... def concatenate_general(ws):
...     WT = type(ws[0])
...     return WT(*(np.concatenate(tuple(getattr(w, f.name) for w in ws))
...                 for f in fields(WT)))

Checking these work:

>>> np.add(w2d, w2d)
Wrap2D(x=array([0, 2, 4]), y=array([ 6, 8, 10]))

>>> np.concatenate((w2d, w2d))
Wrap2D(x=array([0, 1, 2, 0, 1, 2]), y=array([3, 4, 5, 3, 4, 5]))

>>> @dataclass
... class Wrap3D(Wrap2D):
...     '''A simple 3-array wrapper.'''
...     z: np.ndarray

>>> w3d = Wrap3D(np.arange(2), np.arange(3, 5), np.arange(6, 8))
>>> np.add(w3d, w3d)
Wrap3D(x=array([0, 2]), y=array([6, 8]), z=array([12, 14]))
>>> np.concatenate((w3d, w3d))
Wrap3D(x=array([0, 1, 0, 1]), y=array([3, 4, 3, 4]), z=array([6, 7, 6, 7]))

Assisting Groups of Overloads

In the previous examples we wrote implementations for a single NumPy function. Overloading the full set of NumPy functions this way would take a long time.

Wouldn't it be better if we could write many fewer, based on groups of NumPy functions?

>>> add_funcs = {np.add, np.subtract}
>>> @W_FUNCS.assists(add_funcs, types=Wrap1D, dispatch_on=Wrap1D)
... def add_assists(cls, func, w1, w2, *args, **kwargs):
...     return cls(*(func(getattr(w1, f.name), getattr(w2, f.name), *args, **kwargs)
...                     for f in fields(cls)))

>>> stack_funcs = {np.vstack, np.hstack, np.dstack, np.column_stack, np.row_stack}
>>> @W_FUNCS.assists(stack_funcs, types=Wrap1D, dispatch_on=Wrap1D)
... def stack_assists(cls, func, ws, *args, **kwargs):
...     return cls(*(func(tuple(getattr(v, f.name) for v in ws), *args, **kwargs)
...                     for f in fields(cls)))

Checking these work:

>>> np.subtract(w2d, w2d)
Wrap2D(x=array([0, 0, 0]), y=array([0, 0, 0]))

>>> np.vstack((w1d, w1d))
Wrap1D(x=array([[0, 1, 2],
                    [0, 1, 2]]))

>>> np.hstack((w1d, w1d))
Wrap1D(x=array([0, 1, 2, 0, 1, 2]))

We would also like to implement the accumulate method for all the add_funcs overloads:

>>> @add_assists.register("accumulate")
... def add_accumulate_assists(cls, func, w1, *args, **kwargs):
...     return cls(*(func(getattr(w1, f.name), *args, **kwargs)
...                  for f in fields(cls)))

>>> np.subtract.accumulate(w2d)
Wrap2D(x=array([ 0, -1, -3]), y=array([ 3, -1, -6]))

Details

Want to see about type constraints and the API? Check out the docs!