Struct TensorBatchNorm 

pub struct TensorBatchNorm {
    pub config: BatchNormConfig,
    pub gamma: Vec<f64>,
    pub beta: Vec<f64>,
    pub running_mean: Vec<f64>,
    pub running_var: Vec<f64>,
    pub mode: NormMode,
    pub stats: BatchNormStats,
}

Batch normalisation layer.

Layout

Input is expected to have shape [batch_size, num_features].

Example

use ipfrs_tensorlogic::{TensorBatchNorm, BatchNormConfig, NormMode};

let config = BatchNormConfig::default_for(4);
let mut bn = TensorBatchNorm::new(config);

let batch = vec![
    vec![1.0_f64, 2.0, 3.0, 4.0],
    vec![5.0_f64, 6.0, 7.0, 8.0],
];
let output = bn.forward(&batch).expect("forward failed");
assert_eq!(output.len(), 2);
assert_eq!(output[0].len(), 4);

Fields

config: BatchNormConfig

Layer configuration.

gamma: Vec<f64>

Learnable scale parameters, one per feature (γ, initialised to 1.0).

beta: Vec<f64>

Learnable shift parameters, one per feature (β, initialised to 0.0).

running_mean: Vec<f64>

Exponential moving average of per-feature batch means.

running_var: Vec<f64>

Exponential moving average of per-feature batch variances.

mode: NormMode

Current execution mode.

stats: BatchNormStats

Collected statistics.

Implementations

impl TensorBatchNorm

pub fn new(config: BatchNormConfig) -> Self

Construct a new TensorBatchNorm from a BatchNormConfig.

pub fn set_mode(&mut self, mode: NormMode)

Switch the layer between training and inference mode.

pub fn forward(&mut self, input: &[Vec<f64>]) -> Option<Vec<Vec<f64>>>

Run a forward pass through the batch normalisation layer.

• input — a slice of rows, each of length num_features.

Returns None when:

• input is empty, or
• any row does not have exactly num_features elements.

pub fn set_gamma(&mut self, gamma: Vec<f64>) -> bool

Replace the gamma (scale) parameters.

Returns false and leaves gamma unchanged when gamma.len() != num_features.

pub fn set_beta(&mut self, beta: Vec<f64>) -> bool

Replace the beta (shift) parameters.

Returns false and leaves beta unchanged when beta.len() != num_features.

pub fn stats(&self) -> &BatchNormStats

Borrow the accumulated statistics.