Struct GradientNoiseInjector 

pub struct GradientNoiseInjector { /* private fields */ }

Injects configurable noise into gradient arrays for training regularization.

Implementations

impl GradientNoiseInjector

pub fn new(config: GradientNoiseConfig) -> Self

Create a new injector from the given configuration.

The internal PRNG is seeded with config.seed (or a fallback value if the seed is zero, since xorshift64 requires a non-zero state).

pub fn next_gaussian(&mut self) -> f64

Sample from a standard Gaussian (mean 0, std 1) using the Box-Muller transform.

pub fn next_uniform(&mut self, low: f64, high: f64) -> f64

Sample from a uniform distribution on [low, high).

pub fn next_laplacian(&mut self, scale: f64) -> f64

Sample from a Laplacian distribution with location 0 and the given scale, using the inverse-CDF method.

pub fn clip_noise(&self, value: f64) -> f64

Clip value to [-clip, +clip] if a clip bound is configured, otherwise return value unchanged.

pub fn current_scale(&self) -> f64

Compute the current effective noise scale, accounting for decay when using ScheduledGaussian.

pub fn inject(&mut self, gradients: &mut [f64])

Inject noise into gradients in-place.

Each element of the slice receives an independent noise sample drawn from the configured distribution. Statistics are updated accordingly.

pub fn sample_noise(&mut self, count: usize) -> NoiseSample

Generate count noise samples without applying them to any gradient array. Returns a NoiseSample with summary statistics.

pub fn step(&mut self)

Advance the training step counter by one. This affects the noise scale when using ScheduledGaussian.

pub fn reset(&mut self)

Reset the step counter, statistics, and re-seed the PRNG.

pub fn stats(&self) -> &NoiseStats

Read-only access to the accumulated statistics.

pub fn set_scale(&mut self, scale: f64)

Override the current noise scale (applies to all non-Scheduled types; for ScheduledGaussian this sets the initial scale used in decay).