`transpose`¶

Swap two non-batch axes of the tensor.

Category: shape · Identity: transpose@1

Shape¶

x -> out

Relation: out[dim0] == x[dim1]; out[dim1] == x[dim0]; every other axis is unchanged

Port	Direction	Pattern	dtype
`x`	input	`x`	compute
`out`	output	`out`	compute

Arguments¶

Name	Type	Default	Constraints	Description
`dim0` (positional)	int	required	>= 1; <= 3	First axis of the swap; 1 is the first non-batch axis and 3 is the last axis of a rank-4 tensor. The batch axis 0 cannot be swapped.
`dim1` (positional)	int	required	>= 1; <= 3	Second axis of the swap. Equal to dim0 means the identity.

Description¶

Returns x.transpose(dim0, dim1): the extents at dim0 and dim1 trade places and every other axis keeps its extent. The batch axis 0 is never part of a swap, so both arguments are at least 1 and at most 3 (the last axis of the deepest supported rank). Passing the same axis twice is the identity.

Axis conventions. A rank-3 tensor is [B, T, D] for the sequence operations — linear and the activations act on the last axis, D, across T positions — while 1-D convolution and pooling read rank 3 as [B, C, L], channels before length. transpose(1, 2) is the bridge between the two readings, and a second transpose(1, 2) afterwards returns to the sequence layout. At rank 4 the layout is [B, C, H, W], so transpose(2, 3) swaps height and width and transpose(1, 3) exchanges channels with width.

The result is a view: it shares storage and autograd history with the input and is generally not contiguous. Downstream operations handle non-contiguous inputs; reshape and flatten copy when they must. There are no parameters, no buffers, and no difference between train and eval mode. The computation is a stride permutation, so it is exact in every compute dtype.

Examples¶

Example 1¶

Bridges a [B, T, D] sequence to the [B, C, L] layout 1-D convolutions expect.

transpose(1, 2)

Input ['B', 16, 32] → output ['B', 32, 16].

Network: [B, 16, 32] -> [B, 32, 16]  dtype=float32
index  name  operation  input shapes   output shapes
0      n0    transpose  x=[B, 16, 32]  out=[B, 32, 16]

Parameters: 0

Example 2¶

Swaps height and width of an image tensor, leaving the channel axis alone.

transpose(2, 3)

Input ['B', 3, 8, 16] → output ['B', 3, 16, 8].

Network: [B, 3, 8, 16] -> [B, 3, 16, 8]  dtype=float32
index  name  operation  input shapes     output shapes
0      n0    transpose  x=[B, 3, 8, 16]  out=[B, 3, 16, 8]

Parameters: 0

Example 3¶

The swap is bidirectional: the projection width 10 is read back through it.

linear()
transpose(1, 2)

Input ['B', 4, 6] → output ['B', 10, 4].

Network: [B, 4, 6] -> [B, 10, 4]  dtype=float32
index  name  operation  input shapes  output shapes
0      n0    linear     x=[B, 4, 6]   out=[B, 4, 10]
1      n1    transpose  x=[B, 4, 10]  out=[B, 10, 4]

Parameters: 70

transpose¶