Struct MixupLoss 

pub struct MixupLoss {
    pub alpha: f64,
    pub base_loss: Box<dyn Loss>,
}

Mixup data augmentation that mixes training examples and their labels.

Based on “mixup: Beyond Empirical Risk Minimization” (Zhang et al., 2018).

Fields

alpha: f64

Alpha parameter for Beta distribution (typically 1.0).

base_loss: Box<dyn Loss>

Base loss function.

Implementations

impl MixupLoss

pub fn new(alpha: f64, base_loss: Box<dyn Loss>) -> TrainResult<Self>

Create a new Mixup loss.

Arguments

• alpha - Beta distribution parameter (higher = more mixing)
• base_loss - Underlying loss function

pub fn compute_mixup( &self, predictions: &ArrayView<'_, f64, Ix2>, mixed_targets: &ArrayView<'_, f64, Ix2>, ) -> TrainResult<f64>

Compute mixup loss.

Note: In practice, mixup is applied during data loading, not in the loss function. This implementation assumes pre-mixed inputs and targets.

pub fn mix_data( data1: &ArrayView<'_, f64, Ix2>, data2: &ArrayView<'_, f64, Ix2>, lambda: f64, ) -> TrainResult<Array<f64, Ix2>>

Mix two batches of data with random lambda.

Arguments

• data1 - First batch of data
• data2 - Second batch of data
• lambda - Mixing coefficient (0 to 1)

Returns

Mixed data

Trait Implementations

impl Debug for MixupLoss

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

Formats the value using the given formatter.