swapaxes

→Tensor

swapaxes(x: Tensor, axis0: _int, axis1: _int)

source edit

Implementing kernel

C++swapaxes_opfree fn

Return x with axes axis0 and axis1 exchanged.

NumPy-style spelling of pairwise axis transposition. Verbose alias of swapdims; both names refer to the same composite.

Parameters

xTensor

Input tensor.

axis0int

First axis to swap. Negative values count from the end.

axis1int

Second axis to swap. Negative values count from the end.

Returns

Tensor

Tensor with the same data as x but with axis0 and axis1 exchanged — a view of x's buffer on the CPU, so a write through it reaches x; a copy on metal.

Notes

For a tensor with shape $(s_0, \dots, s_{n-1})$ , the output shape is

s'_i = \begin{cases} s_{a_1}, & i = a_0, \\ s_{a_0}, & i = a_1, \\ s_i, & \text{otherwise.} \end{cases}

As in the reference, the result is a view on the CPU: no data moves, only the two strides trade places.

Examples

>>> import lucid
>>> x = lucid.zeros((2, 3, 4))
>>> lucid.swapaxes(x, 0, 2).shape
(4, 3, 2)