Struct DistributionNormalizer 

pub struct DistributionNormalizer { /* private fields */ }

Distribution Normalizer implementing Stage 1 of NOVAQ

Mathematical formulation from NOVAQ paper: W_hat_{i,:} = (W_{i,:} - μ_i) / s_i

where: μ_i = (1/d) * Σ_j W_{i,j} (per-channel mean) s_i = σ_i / Δ if σ_i in top p%, else 1 (outlier scaling)

Implementations

impl DistributionNormalizer

pub fn new(outlier_threshold: f32, seed: u64) -> Self

pub fn normalize( &self, weights: &mut WeightMatrix, ) -> Result<NormalizationMetadata>

Normalize weight matrix by eliminating per-channel means and rescaling outlier channels

pub fn denormalize( &self, weights: &mut [f32], metadata: &NormalizationMetadata, ) -> Result<()>

Denormalize weights using stored metadata

Trait Implementations

impl Debug for DistributionNormalizer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.