Struct WeightInitializer 

pub struct WeightInitializer { /* private fields */ }

Weight initializer that generates tensors according to a chosen strategy.

Implementations

impl WeightInitializer

pub fn new(config: WeightInitConfig) -> Self

Create a new initializer from the given configuration.

pub fn initialize(&mut self, shape: &TensorShape) -> Vec<f64>

Generate weights for the given tensor shape.

pub fn compute_fan(shape: &TensorShape) -> (usize, usize)

Compute (fan_in, fan_out) from a tensor shape.

• 1D: (dims[0], dims[0])
• 2D: (dims[1], dims[0])
• 3D+: (dims[1] * receptive, dims[0] * receptive) where receptive = product of dims[2..]

pub fn xavier_bound(fan_in: usize, fan_out: usize, gain: f64) -> f64

Compute the Xavier/Glorot bound: gain * sqrt(6 / (fan_in + fan_out)).

pub fn he_bound(fan: usize, gain: f64) -> f64

Compute the He/Kaiming bound: gain * sqrt(3 / fan).

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

Generate a uniform random value in [low, high) using xorshift64.

pub fn next_normal(&mut self, mean: f64, std: f64) -> f64

Generate a normally distributed value via the Box-Muller transform.

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

Compute basic statistics for a slice of weights.

pub fn reset_seed(&mut self, seed: u64)

Reset the PRNG seed.

pub fn initialized_count(&self) -> u64

Return the total number of weights initialised so far.