moveaxis

→Tensor

moveaxis(x: Tensor, source: int | Sequence[int], destination: int | Sequence[int])

source

Move one or more axes to new positions.

More flexible than swapaxes — instead of pairwise swapping, moveaxis re-positions a whole set of axes while preserving the relative order of the remaining axes. Thin wrapper around lucid.movedim that accepts either a single int or a sequence on both endpoints.

Parameters

xTensor

Input tensor.

sourceint | Sequence[int]

Original position(s) of the axes to move. Negative values count from the end.

destinationint | Sequence[int]

Final position(s) for each moved axis. Must have the same length as source (after normalisation) and be a permutation-compatible target set.

Returns

Tensor

Tensor whose axes have been re-ordered as requested.

Notes

Equivalent to applying a permutation $\pi$ to the axes such that axis at source[i] ends up at destination[i] while the other axes slide to fill in the gaps preserving their relative order. Useful for vectorised code that needs to ferry a batch dimension across the rank of a tensor.

Examples

>>> import lucid
>>> x = lucid.zeros((2, 3, 4, 5))
>>> lucid.moveaxis(x, 0, -1).shape
(3, 4, 5, 2)
>>> lucid.moveaxis(x, [0, 1], [-1, -2]).shape
(4, 5, 3, 2)