Skip to content

XBArray

XBArray is a Python array library that runs the same code on NumPy, PyTorch and JAX. It is built on the Python Array API Standard.

You write a function once against a compute backend object. You pass in NumpyComputeBackend, PyTorchComputeBackend or JaxComputeBackend, and the function works on that library's arrays: numpy.ndarray, torch.Tensor or jax.Array.

Each backend gives you:

  • The Array API namespace for its library (zeros, matmul, where, linalg, ...), through array-api-compat.
  • The helper functions from array-api-extra, such as at for out-of-place index updates.
  • A few things the standard does not cover: explicit random number generators, conversion to and from NumPy and other backends, device serialization, and mapping a function over nested containers of arrays.

On top of the backends, xbarray.transformations has rotation conversions and point cloud operations written in the same backend-agnostic style.

A short example

from xbarray import ComputeBackend
from xbarray.backends.numpy import NumpyComputeBackend
from xbarray.backends.pytorch import PyTorchComputeBackend

def normalize(backend: ComputeBackend, x):
    """Scale each row of x to unit length."""
    norms = backend.linalg.vector_norm(x, axis=-1, keepdims=True)
    return x / backend.clip(norms, min=1e-12)

for backend in (NumpyComputeBackend, PyTorchComputeBackend):
    x = backend.asarray([[3.0, 4.0], [0.0, 2.0]])
    print(backend, normalize(backend, x))
numpy [[0.6 0.8]
 [0.  1. ]]
pytorch tensor([[0.6000, 0.8000],
        [0.0000, 1.0000]])

The same normalize works with JaxComputeBackend once JAX is installed.

Start here

  • Getting started


    Install XBArray, run a first example on each backend, and get a ComputeBackend.

    Getting started

  • Concepts


    Backends, generic typing, devices and dtypes, random numbers, and moving arrays between backends.

    Concepts

  • Write backend-agnostic code


    The patterns that keep one function working on NumPy, PyTorch and JAX.

    Backend-agnostic code

  • Rotation conversions


    Quaternions, rotation matrices, Euler angles, axis-angle and the 6D representation.

    Rotation conversions

  • Point clouds


    Depth unprojection, farthest point sampling, ball query and jitter.

    Point clouds

  • API reference


    Every module, class and function, generated from the source.

    API reference

Contributing? See Development for running the tests and benchmarks.

Part of UniEnv

XBArray is developed as part of UniEnv, which unifies robot environment and data APIs. The source is at github.com/UniEnvOrg/XBArray under the Apache-2.0 license.

Cite

XBArray is developed as part of UniEnv. If you use it in research, please cite UniEnv; the current author list and BibTeX are on unienvorg.github.io.