Struct Normalize 

pub struct Normalize { /* private fields */ }

Normalizes a tensor with mean and standard deviation.

Supports both scalar normalization (applied uniformly) and per-channel normalization (PyTorch-style transforms.Normalize(mean=[...], std=[...])). For per-channel mode, the tensor is expected to have shape [C, H, W] or [N, C, H, W].

Implementations

impl Normalize

pub fn new(mean: f32, std: f32) -> Self

Creates a new scalar Normalize transform (applied uniformly to all elements).

pub fn per_channel(mean: Vec<f32>, std: Vec<f32>) -> Self

Creates a per-channel Normalize transform (PyTorch-compatible).

For a [C, H, W] tensor, mean and std must have length C. Each channel is normalized independently: output[c] = (input[c] - mean[c]) / std[c].

pub fn standard() -> Self

Creates a Normalize for standard normal distribution (mean=0, std=1).

pub fn zero_centered() -> Self

Creates a Normalize for [0,1] to [-1,1] conversion (mean=0.5, std=0.5).

pub fn imagenet() -> Self

Creates a Normalize for ImageNet (3-channel RGB).

Trait Implementations

impl Transform for Normalize

fn apply(&self, input: &Tensor<f32>) -> Tensor<f32>

Applies the transform to a tensor.