If you are interested in writing backends or multimethods for
please look at the documentation for
uarray, which explains how to
unumpy is meant for three groups of individuals:
- Those who write array-like objects, like developers of Dask, Xnd, PyData/Sparse, CuPy, and others.
- Library authors or programmers that hope to target multiple array backends, listed above.
- Users who wish to target their code to other backends.
For example, the following is currently possible:
>>> import uarray as ua >>> import unumpy as np >>> from unumpy.dask_backend import DaskBackend >>> import unumpy.sparse_backend as SparseBackend >>> import sparse, dask.array as da >>> def main(): ... x = np.zeros(5) ... return np.exp(x) >>> with ua.set_backend(DaskBackend()): ... isinstance(main(), da.core.Array) True >>> with ua.set_backend(SparseBackend): ... isinstance(main(), sparse.SparseArray) True
Now imagine some arbitrarily nested code, all for which the implementations can be switched out using a simple context manager.
unumpy is an in-progress mirror of the NumPy API which allows the user
to dynamically switch out the backend that is used. It also allows
auto-selection of the backend based on the arguments passed into a function. It does this by
defining a collection of
uarray multimethods that support dispatch.
Although it currently provides a number of backends, the aspiration is that,
with time, these back-ends will move into the respective libraries and it will be possible
to use the library modules directly as backends.
Note that currently, our coverage is very incomplete. However, we have attempted
to provide at least one of each kind of object in
reference. There are
ufunc s and
ndarray s, which are classes,
ufunc such as
reduce and also functions such as
Where possible, we attempt to provide default implementations so that the whole API does not have to be reimplemented, however, it might be useful to gain speed or to re-implement it in terms of other functions which already exist in your library.
The idea is that once things are more mature, it will be possible to switch out your backend with a simple import statement switch:
import numpy as np # Old method import unumpy as np # Once this project is mature
Currently, the following functions are supported:
- All NumPy universal functions.
For the full range of functions, use
You can use the
uarray.set_backend decorator to set a backend and use the
desired backend. Note that not every backend supports every method. For example,
PyTorch does not have an exact
ufunc equivalent, so we dispatch to actual
methods using a dictionary lookup. The following
backends are supported:
unumpy is based on
uarray, all overrides are done via the
protocols. We strongly recommend you read the
uarray documentation for context.
All functions/methods in
uarray multimethods. This means
you can override them using the
unumpy allows dispatch on
numpy.dtype via the
Dispatching on objects means one can intercept these, convert to an equivalent
native format, or dispatch on their methods, including
We suggest you browse the source for example backends.
Differences between overriding
numpy.ufunc objects and other multimethods¶
Of note here is that there are certain callable objects within NumPy, most
numpy.ufunc objects, that are not typical functions/methods,
and so cannot be directly overridden, the key word here being directly.
In Python, when a method is called, i.e.
x.method(*a, **kw) it is the same
type(x).method(x, *a, **kw) assuming that
method was a regular
method defined on the type. This allows some very interesting things to happen.
For instance, if we make
method a multimethod, it allows us to override
methods, provided we know that the first argument passed in will be
One other thing that is possible (and done in
unumpy) is to override the
__call__ method on a callable object. This is, in fact, exactly how to override
Other interesting things that can be done (but as of now, are not) are to replace
ufunc objects entirely by native equivalents overriding the
This technique can also be applied to
Meta-arrays are arrays that can hold other arrays, such as Dask arrays and XArray datasets.
If meta-arrays and libraries depend on
unumpy instead of NumPy, they can benefit
from containerization and hold arbitrary arrays; not just
__ua_function__ implementation, they might need to do something like the
>>> class Backend: pass >>> meta_backend = Backend() >>> meta_backend.__ua_domain__ = "numpy" >>> def ua_func(f, a, kw): ... # We do this to avoid infinite recursion ... with ua.skip_backend(meta_backend): ... # Actual implementation here ... pass >>> meta_backend.__ua_function__ = ua_func
In this form, one could do something like the following to use the meta-backend:
>>> with ua.set_backend(DaskBackend(inner=SparseBackend)): ... x = np.zeros((2000, 2000)) ... isinstance(x, da.Array) ... isinstance(x.compute(), sparse.SparseArray) True True