Struct FeatureDistiller 

pub struct FeatureDistiller {
    pub layers: Vec<(f32, DistilLoss)>,
}

Feature-based knowledge distillation.

Maintains a list of (weight, DistilLoss) pairs, one per distillation layer.

Fields

layers: Vec<(f32, DistilLoss)>

Per-layer distillation configuration: (weight, loss).

Implementations

impl FeatureDistiller

pub fn uniform_mse(n_layers: usize) -> Self

Create a feature distiller with equal-weight MSE loss for each layer.

Parameters

• n_layers — number of intermediate layers to distil.

pub fn with_weights(weights: Vec<f32>, loss: DistilLoss) -> Self

Create a feature distiller with custom per-layer weights and a shared loss.

pub fn compute_layer_loss( &self, layer_index: usize, teacher_feat: &[f32], student_feat: &[f32], ) -> QuantResult<f32>

Compute the distillation loss for a single layer pair.

Errors

• QuantError::DimensionMismatch — layer_index out of range.
• Propagates errors from the underlying DistilLoss.

pub fn compute_total_loss( &self, teacher_feats: &[&[f32]], student_feats: &[&[f32]], ) -> QuantResult<f32>

Compute the total feature distillation loss across all layers.

teacher_feats[l] and student_feats[l] must have matching lengths.

Errors

• QuantError::DimensionMismatch — wrong number of feature arrays.
• Propagates per-layer errors.

pub fn n_layers(&self) -> usize

Number of distillation layers.

pub fn normalise_weights(&mut self)

Normalise layer weights so they sum to 1.

Trait Implementations

impl Clone for FeatureDistiller

fn clone(&self) -> FeatureDistiller

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FeatureDistiller

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

Formats the value using the given formatter.