Struct BYOL 

pub struct BYOL { /* private fields */ }

BYOL (Bootstrap Your Own Latent) framework (Grill et al., 2020).

Self-supervised learning WITHOUT negative samples:

• Online network predicts target network representation
• Target network is momentum-updated from online
• Stops gradient on target to prevent collapse

Reference

Grill, J.B., et al. (2020). Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning. NeurIPS.

Implementations

impl BYOL

pub fn new(momentum: f32) -> Self

Create BYOL framework.

pub fn loss(&self, online_pred: &[Vec<f32>], target_proj: &[Vec<f32>]) -> f32

Compute BYOL loss (MSE between normalized predictions).

Arguments

• online_pred - Predictions from online network
• target_proj - Projections from target network (stop gradient)

pub fn symmetric_loss( &self, pred_1: &[Vec<f32>], proj_2: &[Vec<f32>], pred_2: &[Vec<f32>], proj_1: &[Vec<f32>], ) -> f32

Symmetric BYOL loss (both views predict each other).

pub fn momentum_update(&self, online_params: &[f32], target_params: &mut [f32])

Update target network parameters with momentum.

pub fn momentum(&self) -> f32

Trait Implementations

impl Clone for BYOL

fn clone(&self) -> BYOL

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BYOL

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

Formats the value using the given formatter.