NEP 30 — Duck typing for NumPy arrays - implementation#

Author:

Peter Andreas Entschev <pentschev@nvidia.com>

Author:

Stephan Hoyer <shoyer@google.com>

Status:

Superseded

Replaced-By:

NEP 56 — Array API standard support in NumPy’s main namespace

Type:

Standards Track

Created:

2019-07-31

Updated:

2019-07-31

Resolution:

https://mail.python.org/archives/list/numpy-discussion@python.org/message/Z6AA5CL47NHBNEPTFWYOTSUVSRDGHYPN/

Abstract#

We propose the __duckarray__ protocol, following the high-level overview described in NEP 22, allowing downstream libraries to return arrays of their defined types, in contrast to np.asarray, that coerces those array_like objects to NumPy arrays.

Detailed description#

NumPy’s API, including array definitions, is implemented and mimicked in countless other projects. By definition, many of those arrays are fairly similar in how they operate to the NumPy standard. The introduction of __array_function__ allowed dispatching of functions implemented by several of these projects directly via NumPy’s API. This introduces a new requirement, returning the NumPy-like array itself, rather than forcing a coercion into a pure NumPy array.

For the purpose above, NEP 22 introduced the concept of duck typing to NumPy arrays. The suggested solution described in the NEP allows libraries to avoid coercion of a NumPy-like array to a pure NumPy array where necessary, while still allowing that NumPy-like array libraries that do not wish to implement the protocol to coerce arrays to a pure NumPy array via np.asarray.

Usage Guidance#

Code that uses np.duckarray is meant for supporting other ndarray-like objects that “follow the NumPy API”. That is an ill-defined concept at the moment – every known library implements the NumPy API only partly, and many deviate intentionally in at least some minor ways. This cannot be easily remedied, so for users of np.duckarray we recommend the following strategy: check if the NumPy functionality used by the code that follows your use of np.duckarray is present in Dask, CuPy and Sparse. If so, it’s reasonable to expect any duck array to work here. If not, we suggest you indicate in your docstring what kinds of duck arrays are accepted, or what properties they need to have.

To exemplify the usage of duck arrays, suppose one wants to take the mean() of an array-like object arr. Using NumPy to achieve that, one could write np.asarray(arr).mean() to achieve the intended result. If arr is not a NumPy array, this would create an actual NumPy array in order to call .mean(). However, if the array is an object that is compliant with the NumPy API (either in full or partially) such as a CuPy, Sparse or a Dask array, then that copy would have been unnecessary. On the other hand, if one were to use the new __duckarray__ protocol: np.duckarray(arr).mean(), and arr is an object compliant with the NumPy API, it would simply be returned rather than coerced into a pure NumPy array, avoiding unnecessary copies and potential loss of performance.

Implementation#

The implementation idea is fairly straightforward, requiring a new function duckarray to be introduced in NumPy, and a new method __duckarray__ in NumPy-like array classes. The new __duckarray__ method shall return the downstream array-like object itself, such as the self object, while the __array__ method raises TypeError. Alternatively, the __array__ method could create an actual NumPy array and return that.

The new NumPy duckarray function can be implemented as follows:

def duckarray(array_like):
    if hasattr(array_like, '__duckarray__'):
        return array_like.__duckarray__()
    return np.asarray(array_like)

Example for a project implementing NumPy-like arrays#

Now consider a library that implements a NumPy-compatible array class called NumPyLikeArray, this class shall implement the methods described above, and a complete implementation would look like the following:

class NumPyLikeArray:
    def __duckarray__(self):
        return self

    def __array__(self):
        raise TypeError("NumPyLikeArray can not be converted to a NumPy "
                         "array. You may want to use np.duckarray() instead.")

The implementation above exemplifies the simplest case, but the overall idea is that libraries will implement a __duckarray__ method that returns the original object, and an __array__ method that either creates and returns an appropriate NumPy array, or raises a``TypeError`` to prevent unintentional use as an object in a NumPy array (if np.asarray is called on an arbitrary object that does not implement __array__, it will create a NumPy array scalar).

In case of existing libraries that don’t already implement __array__ but would like to use duck array typing, it is advised that they introduce both __array__ and``__duckarray__`` methods.

Usage#

An example of how the __duckarray__ protocol could be used to write a stack function based on concatenate, and its produced outcome, can be seen below. The example here was chosen not only to demonstrate the usage of the duckarray function, but also to demonstrate its dependency on the NumPy API, demonstrated by checks on the array’s shape attribute. Note that the example is merely a simplified version of NumPy’s actual implementation of stack working on the first axis, and it is assumed that Dask has implemented the __duckarray__ method.

def duckarray_stack(arrays):
    arrays = [np.duckarray(arr) for arr in arrays]

    shapes = {arr.shape for arr in arrays}
    if len(shapes) != 1:
        raise ValueError('all input arrays must have the same shape')

    expanded_arrays = [arr[np.newaxis, ...] for arr in arrays]
    return np.concatenate(expanded_arrays, axis=0)

dask_arr = dask.array.arange(10)
np_arr = np.arange(10)
np_like = list(range(10))

duckarray_stack((dask_arr, dask_arr))   # Returns dask.array
duckarray_stack((dask_arr, np_arr))     # Returns dask.array
duckarray_stack((dask_arr, np_like))    # Returns dask.array

In contrast, using only np.asarray (at the time of writing of this NEP, this is the usual method employed by library developers to ensure arrays are NumPy-like) has a different outcome:

def asarray_stack(arrays):
    arrays = [np.asanyarray(arr) for arr in arrays]

    # The remaining implementation is the same as that of
    # ``duckarray_stack`` above

asarray_stack((dask_arr, dask_arr))     # Returns np.ndarray
asarray_stack((dask_arr, np_arr))       # Returns np.ndarray
asarray_stack((dask_arr, np_like))      # Returns np.ndarray

Backward compatibility#

This proposal does not raise any backward compatibility issues within NumPy, given that it only introduces a new function. However, downstream libraries that opt to introduce the __duckarray__ protocol may choose to remove the ability of coercing arrays back to a NumPy array via np.array or np.asarray functions, preventing unintended effects of coercion of such arrays back to a pure NumPy array (as some libraries already do, such as CuPy and Sparse), but still leaving libraries not implementing the protocol with the choice of utilizing np.duckarray to promote array_like objects to pure NumPy arrays.

Previous proposals and discussion#

The duck typing protocol proposed here was described in a high level in NEP 22.

Additionally, longer discussions about the protocol and related proposals took place in numpy/numpy #13831