Module model_pruner 

Model weight pruning with magnitude, structured, and gradual scheduling strategies.

This module implements several classical and modern neural-network pruning approaches:

• Magnitude pruning — zero out individual weights whose absolute value falls below a fixed threshold.
• Percentile-magnitude pruning — zero out the bottom X% of weights ranked by absolute magnitude.
• Structured L1 pruning — remove entire neurons / output channels whose mean L1 norm is below a threshold (structured sparsity that directly speeds up inference on most hardware).
• Random pruning — stochastically mask out X% of weights using a deterministic xorshift64 PRNG seeded from PrunerConfig::seed.
• Gradual pruning — linearly ramp sparsity from an initial value to a final value over a user-specified step window (Zhu & Gupta 2018 style).

An optional binary mask is maintained alongside each LayerWeights tensor so that sparse structure can be preserved across optimiser updates.

Example

use ipfrs_tensorlogic::{
    ModelPruner, PrunerConfig, PruningStrategy, LayerWeights,
};

let cfg = PrunerConfig {
    strategy: PruningStrategy::Magnitude(0.1),
    seed: 42,
    update_mask: true,
};
let mut pruner = ModelPruner::new(cfg);

let mut layer = LayerWeights {
    name: "fc1".to_string(),
    weights: vec![0.05, -0.2, 0.0, 0.3, -0.08],
    mask: None,
};

let result = pruner.prune_layer(&mut layer);
assert!(result.sparsity > 0.0);

Structs

LayerWeights

A named layer’s weight tensor together with an optional sparsity mask.

ModelPruner

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

PrunerConfig

Configuration bundle passed to ModelPruner::new.

PrunerStats

Cumulative statistics tracked by a ModelPruner across all layers and all pruning steps.

PruningResult

Per-layer summary returned by each call to ModelPruner::prune_layer.

Enums

PruningStrategy

Selects the algorithm used to decide which weights to prune.