Struct BayesianLinear 

pub struct BayesianLinear {
    pub in_features: usize,
    pub out_features: usize,
    pub w_mu: Vec<f64>,
    pub w_log_sigma: Vec<f64>,
    pub b_mu: Vec<f64>,
    pub b_log_sigma: Vec<f64>,
    pub prior_std: f64,
}

A single variational Bayesian linear layer.

Weights: w_{ij} ~ N(w_mu_{ij}, exp(w_log_sigma_{ij})^2) Biases: b_j ~ N(b_mu_j, exp(b_log_sigma_j)^2)

Stored as flat row-major vectors of length out_features * in_features.

Fields

in_features: usize

Number of input features

out_features: usize

Number of output features

w_mu: Vec<f64>

Weight posterior means, length out_features * in_features

w_log_sigma: Vec<f64>

Weight posterior log-std, length out_features * in_features

b_mu: Vec<f64>

Bias posterior means, length out_features

b_log_sigma: Vec<f64>

Bias posterior log-std, length out_features

prior_std: f64

Prior standard deviation

Implementations

impl BayesianLinear

pub fn new( in_features: usize, out_features: usize, prior_std: f64, ) -> Result<Self, StatsError>

Create a new BayesianLinear layer.

Weights are initialized from N(0, 0.1) and log-sigma initialized to -3.0 (corresponding to sigma ≈ 0.05, tight initial posterior).

Arguments

• in_features - Input dimensionality
• out_features - Output dimensionality
• prior_std - Standard deviation of the weight prior

Errors

Returns an error if in_features or out_features is zero.

pub fn forward_sample( &self, x: &[f64], rng: &mut impl FnMut() -> f64, ) -> Result<Vec<f64>, StatsError>

Forward pass with sampled weights (reparameterization trick).

Samples w = w_mu + eps * exp(w_log_sigma) for each weight and bias, then computes the matrix-vector product.

Arguments

• x - Input vector of length in_features
• rng - Closure producing standard normal samples N(0,1)

Returns

Output vector of length out_features

Errors

Returns an error if x has incorrect length.

pub fn forward_mean(&self, x: &[f64]) -> Result<Vec<f64>, StatsError>

Deterministic forward pass using posterior means only.

Computes output = W_mu @ x + b_mu. Useful for fast predictive mean.

Arguments

• x - Input vector of length in_features

Errors

Returns an error if x has incorrect length.

pub fn kl_divergence(&self, prior_std: f64) -> f64

Compute the KL divergence KL[q(w) || p(w)] for all weights and biases.

For q = N(mu, sigma^2) and p = N(0, prior_std^2): KL = -0.5 * sum(1 + 2*log_sigma - log(prior_std^2) - mu^2/prior_std^2 - sigma^2/prior_std^2)

Arguments

• prior_std - Prior standard deviation (can differ from initialization value)

pub fn update( &mut self, grad_w_mu: &[f64], grad_w_log_sigma: &[f64], grad_b_mu: &[f64], grad_b_log_sigma: &[f64], lr: f64, ) -> Result<(), StatsError>

Apply a gradient step (SGD) to the variational parameters.

Arguments

• grad_w_mu - Gradient of loss w.r.t. w_mu, length out*in
• grad_w_log_sigma - Gradient of loss w.r.t. w_log_sigma, length out*in
• grad_b_mu - Gradient of loss w.r.t. b_mu, length out
• grad_b_log_sigma - Gradient of loss w.r.t. b_log_sigma, length out
• lr - Learning rate

Errors

Returns an error if gradient dimensions are inconsistent.

pub fn n_params(&self) -> usize

Total number of variational parameters (for KL / ELBO scaling)

Trait Implementations

impl Clone for BayesianLinear

fn clone(&self) -> BayesianLinear

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BayesianLinear

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

Formats the value using the given formatter.