Skip to main content

Pad

Pad

Pads tensor with specified value along boundaries.

Parameters:

  • padding: Padding sizes (left, right) or (left, right, top, bottom) etc.
  • value: Padding value (default: 0)
  • mode: Padding mode - constant, reflect, replicate, circular (default: constant)

Shape Contract:

  • Input: [*shape_in] tensor to pad
  • Output: [*shape_out] padded tensor (larger along padded dimensions)

Notes:

  • No learnable parameters (pure structural operation)
  • padding specified from last dim backward: (left, right, top, bottom, ...)
  • Modes:
  • constant: fill with value
  • reflect: reflect values at boundary
  • replicate: repeat edge values
  • circular: wrap around
  • Essential for maintaining spatial size in convolutions

Signature​

neuron Pad(padding, value=Int(0), mode=Name("constant"))

Ports​

Inputs:

  • default: [*shape_in]

Outputs:

  • default: [*shape_out]

Implementation​

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