Struct TrainingMonitor 

pub struct TrainingMonitor { /* private fields */ }

Real-time training health monitor.

Attaches to any optimizer to detect pathological training behavior including gradient explosion/vanishing, loss divergence, dead neurons, and convergence. Generates alerts with severity levels and provides actionable suggestions.

Implementations

impl TrainingMonitor

pub fn new() -> Self

Creates a new training monitor with default configuration.

pub fn with_config(config: MonitorConfig) -> Self

Creates a new training monitor with the given configuration.

pub fn record_step(&mut self, loss: f32, grad_norms: &[(&str, f32)], lr: f32)

Records a single training step.

Arguments

• loss - The loss value for this step.
• grad_norms - Slice of (parameter_name, gradient_norm) pairs.
• lr - The current learning rate.

pub fn check_health(&self) -> HealthReport

Returns a full health report for the current training state.

pub fn is_healthy(&self) -> bool

Returns true if training appears healthy (no critical alerts, loss not diverging).

pub fn alerts(&self) -> &[TrainingAlert]

Returns the accumulated alerts.

pub fn clear_alerts(&mut self)

Clears all accumulated alerts.

pub fn loss_trend(&self) -> LossTrend

Analyzes the loss trajectory over the recent window.

Compares the rolling average of the most recent window to the rolling average of the previous window to classify the trend.

pub fn suggest_lr(&self) -> Option<f32>

Suggests a learning rate adjustment based on current training dynamics.

Returns None if no adjustment is needed or training has converged.

pub fn grad_norm_stats(&self) -> (f32, f32, f32)

Returns (mean, std, max) of gradient norms over the recent window.

pub fn convergence_score(&self) -> f32

Returns a convergence score between 0.0 and 1.0.

1.0 indicates full convergence (no loss change over the window). 0.0 indicates the loss is still actively changing.

pub fn summary(&self) -> String

Returns a human-readable summary of the current training state.

Trait Implementations

impl Default for TrainingMonitor

fn default() -> Self

Returns the “default value” for a type.