Struct ModelPruner 

pub struct ModelPruner { /* private fields */ }

Stateful weight pruner. Advance the step counter with ModelPruner::advance_step between training iterations.

Implementations

impl ModelPruner

pub fn new(config: PrunerConfig) -> Self

Create a new pruner from the supplied configuration.

The PRNG seed is initialised to config.seed, falling back to 1 if the seed is zero (xorshift64 must not start from zero).

pub fn prune_layer(&mut self, layer: &mut LayerWeights) -> PruningResult

Prune a single layer in-place according to the configured strategy.

The layer.weights vector is mutated directly; the mask (if enabled) is created or updated.

pub fn prune_all(&mut self, layers: &mut [LayerWeights]) -> Vec<PruningResult>

Prune every layer in layers and return one result per layer.

pub fn current_sparsity_target(&self) -> f64

Compute the sparsity target for the current step.

For PruningStrategy::GradualPruning this linearly interpolates between initial_sparsity and final_sparsity. For all other strategies it returns the equivalent fixed fraction.

pub fn compute_threshold(weights: &[f64], percentile: f64) -> f64

Return the weight value at the given percentile (0–100) of the absolute-magnitude distribution.

Uses a partial sort to avoid allocating a fully sorted copy when only the boundary value is needed. Returns 0.0 for empty slices.

pub fn compute_sparsity(weights: &[f64]) -> f64

Fraction of elements in weights that are exactly zero.

pub fn compute_l1_norm(weights: &[f64]) -> f64

Sum of absolute values of weights.

pub fn advance_step(&mut self)

Advance the step counter by one.

pub fn next_uniform_prng(&mut self) -> f64

Generate the next pseudo-random float in [0, 1) using xorshift64.

The internal state is updated in-place so successive calls yield independent values.

pub fn apply_mask(layer: &mut LayerWeights)

Zero out entries in layer.weights where the corresponding mask entry is false. If no mask is present this is a no-op.

pub fn stats(&self) -> &PrunerStats

Immutable access to the accumulated statistics.