autoray 0.2.0

Write backend agnostic numeric code compatible with any numpy-ish array library.

A lightweight python AUTOmatic-arRAY library. Write numeric code that works for:

• numpy

• cupy

• dask

• tensorflow

• autograd

• jax

• mars

• … and indeed any library that provides a numpy-ish api.

For example consider this function that orthogonalizes a matrix using the modified Gram-Schmidt algorithm:

from autoray import do

def modified_gram_schmidt(X):

    Q = []
    for j in range(0, X.shape[0]):

        q = X[j, :]
        for i in range(0, j):
            rij = do('tensordot', do('conj', Q[i]), q, 1)
            q = q - rij * Q[i]

        rjj = do('linalg.norm', q, 2)
        Q.append(q / rjj)

    return do('stack', Q, axis=0, like=X)

Which is now compatible with all of the above mentioned libraries! (N.B. this particular example is also probably slow). If you don’t like the explicit do syntax, then you can import the fake numpy object as a drop-in replacement instead:

>>> from autoray import numpy as np

>>> x = np.random.uniform(size=(2, 3, 4), like='tensorflow')

>>> np.tensordot(x, x, [(2, 1), (2, 1)])
<tf.Tensor 'Tensordot:0' shape=(2, 2) dtype=float32>

>>> np.eye(3, like=x)  # many functions obviously can't dispatch without the `like` keyword
<tf.Tensor 'eye/MatrixDiag:0' shape=(3, 3) dtype=float32>

Of course complete compatibility is not going to be possible for all functions, operations and libraries, but autoray hopefully makes the job much easier.

How does it work?

autoray works using essentially a single dispatch mechanism on the first argument for do, or the like keyword argument if specified, fetching functions from the whichever module defined that supplied array. Additionally, it caches a few custom translations and lookups so as to handle libraries like tensorflow that don’t exactly replicate the numpy api (for example sum gets translated to tensorflow.reduce_sum).

Special Functions

The main function is do, but the following special (i.e. not in numpy) functions are also implemented that may be useful:

• autoray.infer_backend - check what library is being infered for a given array

• autoray.to_backend_dtype - convert a string specified dtype like ‘float32’ to torch.float32 for example

• autoray.get_dtype_name - convert a backend dtype back into the equivalent string specifier like ‘complex64’

• autoray.astype - backend agnostic dtype conversion of arrays

• autoray.to_numpy - convert any array to numpy

Deviations from numpy

autoray doesn’t have an API as such, since it is essentially just a fancy single dispatch mechanism. On the other hand, where translations are in place, they generally use the numpy API. So autoray.do('stack', arrays=pytorch_tensors, axis=0) gets automatically translated into torch.stack(tensors=pytorch_tensors, dims=0) and so forth.

Currently the one place this isn’t true is autoray.do('linalg.svd', x) where instead full_matrices=False is used as the default since this generally makes more sense and many libraries don’t even implement the other case. Autoray also dispatches 'linalg.expm' for numpy arrays to scipy, and may well do with other scipy-only functions at some point.

Installation

You can install autoray as standard with pip. Alternatively, simply copy the monolithic autoray.py into your project internally (if dependencies aren’t your thing).

Alternatives

• The __array_function__ protocol has been suggested and now implemented in numpy. Hopefully this will eventually negate the need for autoray. On the other hand, third party libraries themselves need to implement the interface, which has not been done, for example, in tensorflow yet.

• The uarray project aims to develop a generic array interface but comes with the warning “This is experimental and very early research code. Don’t use this.”.

Contributing

Pull requests such as extra translations are very welcome!