Skip to main content

Scale

Scale

Multiplies input by a learnable scale vector. Element-wise multiplicative scaling applied along the last dimension.

Parameters:

  • dim: Size of the scale vector (must match last dimension of input)

Shape Contract:

  • Input: [*, dim] tensor with last dimension matching scale size
  • Output: [*, dim] same shape as input

Notes:

  • Learnable parameter: scale vector of shape [dim], initialized to 1
  • Applied element-wise: out = input * scale
  • Used in normalization layers (gamma parameter)
  • Can be used for learned feature weighting

Signature​

neuron Scale(dim)

Ports​

Inputs:

  • default: [*, dim]

Outputs:

  • default: [*, dim]

Implementation​

Source { source: "core", path: "operations/Scale" }