Struct ConvergenceDetector 

pub struct ConvergenceDetector { /* private fields */ }

Convergence detection for federated learning.

Supports two usage modes:

1. Legacy: new(window_size, threshold) + add_loss() + has_converged()
2. EMA mode: with_config(cfg) + update(loss) returns bool

Implementations

impl ConvergenceDetector

pub fn new(window_size: usize, threshold: f64) -> Self

Create a new convergence detector (legacy mode).

pub fn with_config(config: ConvergenceConfig) -> Self

Create a convergence detector driven by ConvergenceConfig.

pub fn add_loss(&mut self, loss: f64)

Add a loss value (legacy mode).

pub fn has_converged(&self) -> bool

Check if training has converged (legacy mode).

pub fn latest_loss(&self) -> Option<f64>

Get the latest loss (legacy mode).

pub fn history(&self) -> &[f64]

Get loss history (legacy mode).

pub fn update(&mut self, loss: f32) -> bool

Record a new loss value and check for convergence (EMA mode).

Returns true when the smoothed gradient norm drops below config.threshold for config.patience consecutive rounds.

pub fn smoothed_loss(&self) -> f32

Exponentially smoothed loss (EMA mode).

pub fn gradient_norm(&self) -> f32

Gradient norm computed as mean |Δloss| over the window (EMA mode).

pub fn plateau_rounds(&self) -> usize

Number of consecutive rounds the gradient norm has been below threshold (EMA mode).

pub fn reset(&mut self)

Reset all history (both legacy and EMA modes).

pub fn config(&self) -> Option<&ConvergenceConfig>

Return a reference to the ConvergenceConfig (EMA mode).

Returns None when constructed in legacy mode.