Struct LabelSmoothingLoss 

pub struct LabelSmoothingLoss {
    pub epsilon: f64,
    pub num_classes: usize,
}

Label smoothing cross-entropy loss.

Based on “Rethinking the Inception Architecture for Computer Vision” (Szegedy et al., 2016). Replaces hard 0/1 labels with smoothed distribution:

• True class: 1 - epsilon
• Other classes: epsilon / (num_classes - 1)

Fields

epsilon: f64

Smoothing parameter (typically 0.1).

num_classes: usize

Number of classes.

Implementations

impl LabelSmoothingLoss

pub fn new(epsilon: f64, num_classes: usize) -> TrainResult<Self>

Create a new label smoothing loss.

Arguments

• epsilon - Smoothing parameter (0 = no smoothing, higher = more smoothing)
• num_classes - Number of classes

pub fn smooth_labels( &self, targets: &ArrayView<'_, f64, Ix2>, ) -> Array<f64, Ix2>

Apply label smoothing to targets.

Arguments

• targets - One-hot encoded targets

Returns

Smoothed targets

Trait Implementations

impl Clone for LabelSmoothingLoss

fn clone(&self) -> LabelSmoothingLoss

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LabelSmoothingLoss

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

Formats the value using the given formatter.

impl Loss for LabelSmoothingLoss

fn compute( &self, predictions: &ArrayView<'_, f64, Ix2>, targets: &ArrayView<'_, f64, Ix2>, ) -> TrainResult<f64>

Compute loss value.

fn gradient( &self, predictions: &ArrayView<'_, f64, Ix2>, targets: &ArrayView<'_, f64, Ix2>, ) -> TrainResult<Array<f64, Ix2>>

Compute loss gradient with respect to predictions.

fn name(&self) -> &str

Get the name of the loss function.