Transpose

Transpose

Permutes the dimensions of a tensor according to the specified ordering. Essential for reshaping data between different layer expectations.

Parameters:

• dims: Tuple/list specifying the new dimension order (e.g., (0, 2, 1))

Shape Contract:

• Input: [*shape_in] tensor with dimensions to permute
• Output: [*shape_out] tensor with reordered dimensions

Notes:

• No learnable parameters (pure structural operation)
• Common use: converting between channel-first and channel-last formats
• For attention: transposing keys for batched matrix multiplication
• Example: [batch, seq, dim] with dims=(0, 2, 1) gives [batch, dim, seq]
• All dimensions must be accounted for in the permutation

Signature

neuron Transpose(dims)

Ports

Inputs:

• default: [*shape_in]

Outputs:

• default: [*shape_out]

Implementation

Source { source: "core", path: "structural/Transpose" }