Skip to main content

InstanceNorm

Instance Normalization

Normalizes each sample independently across spatial dimensions. Commonly used in style transfer and image generation.

Parameters:

  • num_features: Number of channels (C dimension)
  • eps: Small constant for numerical stability (default: 1e-5)
  • affine: If true, learnable scale and shift (default: true)

Shape Contract:

  • Input: [batch, num_features, *spatial] feature maps
  • Output: [batch, num_features, *spatial] same shape as input

Notes:

  • Normalizes across spatial dimensions only (not batch or channel)
  • Each instance (sample) normalized independently
  • Equivalent to GroupNorm with num_groups=num_features
  • Removes style information from features
  • Used in neural style transfer, GANs, image-to-image translation

Signature​

neuron InstanceNorm(num_features, eps=Float(1e-5), affine=Bool(true))

Ports​

Inputs:

  • default: [batch, num_features, *spatial]

Outputs:

  • default: [batch, num_features, *spatial]

Implementation​

Source { source: "core", path: "normalization/InstanceNorm" }