softmax

layernorm<S: Into<Shape>>(
    &self,
    normalized_shape: S,
    gamma: Option<&Tensor<T>>,
    beta: Option<&Tensor<T>>,
    eps: T
) -> Result<Tensor<T>, TensorError>

Applies Layer Normalization over a specified axes.

Parameters:

normalized_shape: shape that must match the dimension size from input tensor shape (from right to left)

gamma: Optional scale tensor of shape [normalized_shape]

beta: Optional bias tensor of shape [normalized_shape]

eps: A value added to the denominator for numerical stability.

Returns:

A new tensor with the same shape as input with type C

Examples:

use hpt::{
    common::TensorInfo,
    error::TensorError,
    ops::{NormalizationOps, Random, TensorCreator},
    Tensor,
};

fn main() -> Result<(), TensorError> {
    // Create a 2x3x4 tensor
    let x = Tensor::<f32>::randn(&[2, 3, 4])?;

    // Create optional gamma and beta parameters
    let gamma = Tensor::<f32>::ones(&[4])?;
    let beta = Tensor::<f32>::zeros(&[4])?;

    // Apply layer normalization over the last dimension
    let result = x.layernorm(&[4], Some(&gamma), Some(&beta), 1e-5)?;
    println!("LayerNorm result shape: {:?}", result.shape());

    Ok(())
}

Backend Support

Backend	Supported
CPU	✅
Cuda	❌