viva_glyph/codebook

Codebook - Learned vocabulary for vector quantization

A codebook is a set of prototype vectors (centroids). Each vector in latent space is mapped to its nearest centroid.

Theory

Based on Vector Quantization (VQ) from signal processing:

• K centroids partition the space into Voronoi cells
• Encoding: find nearest centroid index
• Decoding: lookup centroid by index

References

• Gray (1984). Vector Quantization. IEEE ASSP Magazine.
• van den Oord et al. (2017). VQ-VAE. NeurIPS.

pub type Codebook {
  Codebook(
    centroids: List(List(Float)),
    dimension: Int,
    size: Int,
  )
}

pub type QuantizeResult {
  QuantizeResult(index: Int, error: Float)
}

pub fn dequantize(codebook: Codebook, index: Int) -> List(Float)

pub fn empty(dimension: Int, size: Int) -> Codebook

pub fn from_vectors(
  vectors: List(List(Float)),
) -> option.Option(Codebook)

pub fn get(
  codebook: Codebook,
  index: Int,
) -> option.Option(List(Float))

pub fn init_deterministic(
  dimension: Int,
  size: Int,
  seed: Int,
) -> Codebook

pub fn quantize(
  codebook: Codebook,
  input: List(Float),
) -> QuantizeResult

pub fn update_centroid(
  codebook: Codebook,
  index: Int,
  new_centroid: List(Float),
) -> Codebook