This RFC proposes adding support to the array API specification for creating a new array by tiling a provided array.

Overview

Based on array comparison data, the API is available in all array libraries.

Prior art

• NumPy: https://numpy.org/doc/stable/reference/generated/numpy.tile.html
• PyTorch: https://pytorch.org/docs/stable/generated/torch.tile.html
• TensorFlow: https://www.tensorflow.org/api_docs/python/tf/tile
• JAX: https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.tile.html
• CuPy: https://docs.cupy.dev/en/stable/reference/generated/cupy.tile.html
• Dask: https://docs.dask.org/en/stable/generated/dask.array.tile.html

Proposal

def tile(x: array, repetitions: Union[int, Tuple[int, ...]], /)

• repetitions: the number of repetitions of x along each axis. Let N = len(repetitions) and let M = len(x.shape).

  • if M < N, then x must be broadcasted to be N dimensional by prepending new axes.
  • if M > N, then repetitions must be broadcasted by prepending new axes.

Questions

• Neither PyTorch nor TensorFlow support providing an int for repetitions. TensorFlow requires repetitions to be a Tensor. All array libraries apart from TensorFlow allow repetitions to be a tuple. Should repetitions be typed more strictly?

Related

• Mention of xarray's need for tile: Common APIs across array libraries (1 year later) #187 (comment)