movedim_op

→TensorImplPtr

int movedim_op(const int & a, const int & source, const int & destination)

source edit

Python wrappers

pylucid.ops.movedimfunction

Move dimensions from source to destination positions.

Composite over permute_op. Builds an explicit permutation in which every axis listed in source is placed at the matching index in destination; axes not listed preserve their original relative order and fill the remaining slots left to right. No new backward node — gradient flows through PermuteBackward.

Math

If $\pi$ is the constructed permutation of $[0, n)$ , then

y_{i_0, \ldots, i_{n-1}} = a_{i_{\pi(0)}, \ldots, i_{\pi(n-1)}}

See Also

permute_op — underlying primitive.

Parameters

aTensorImplPtr

Source tensor.

sourcevector<int>

Axes of a to move. Negative values wrap modulo a.ndim. Must be unique.

destinationvector<int>

Target positions for the moved axes, parallel to source. Negative values wrap. Must be unique and have the same length as source.

Returns

TensorImplPtr

A view (or contiguous copy, per the primitive's policy) of a with the requested axis order.

Raises

Failure

If a is null, the lists differ in length, or either list contains duplicates.

IndexError

If any axis is out of range.

Examples

movedim_op(a, {0}, {-1}) moves the leading axis to the trailing
position — equivalent to a cyclic left rotation of axes.