Struct ContrastiveLossEngine 

pub struct ContrastiveLossEngine { /* private fields */ }

Contrastive loss function engine.

Implementations

impl ContrastiveLossEngine

pub fn new(config: ContrastiveConfig) -> Self

Create a new contrastive loss engine.

pub fn with_defaults() -> Self

Create with default configuration.

pub fn info_nce_loss( &mut self, anchors: &[Vec<f64>], positives: &[Vec<f64>], negatives: &[Vec<f64>], ) -> ContrastiveLossResult

Compute InfoNCE (Noise Contrastive Estimation) loss.

L = -log( exp(sim(anchor, positive) / τ) / Σ exp(sim(anchor, neg_i) / τ) )

pub fn triplet_loss( &mut self, anchors: &[Vec<f64>], positives: &[Vec<f64>], negatives: &[Vec<f64>], ) -> ContrastiveLossResult

Compute triplet margin loss.

L = max(0, d(anchor, positive) - d(anchor, negative) + margin)

pub fn nt_xent_loss( &mut self, embeddings_a: &[Vec<f64>], embeddings_b: &[Vec<f64>], ) -> ContrastiveLossResult

Compute NT-Xent (Normalised Temperature-Scaled Cross-Entropy) loss.

SimCLR-style loss over a batch of augmented pairs.

pub fn mine_semi_hard( &self, anchor: &[f64], positive: &[f64], negative_pool: &[Vec<f64>], ) -> Vec<usize>

Mine semi-hard negatives from a pool.

Returns indices of negatives that are farther from anchor than positive but within the margin boundary.

pub fn mine_hardest( &self, anchor: &[f64], negative_pool: &[Vec<f64>], ) -> Option<usize>

Mine the hardest negative from a pool (closest to anchor).

pub fn stats(&self) -> &ContrastiveTrainingStats

Get training statistics.

pub fn reset_stats(&mut self)

Reset training statistics.

pub fn config(&self) -> &ContrastiveConfig

Get the configuration.