Struct TensorOptimizationHistory 

pub struct TensorOptimizationHistory {
    pub steps: Vec<OptimizationStep>,
    pub config: OptimizationHistoryConfig,
    pub best_loss: f64,
    pub best_step: u64,
    pub consecutive_no_progress: usize,
}

Records and analyzes a history of optimization steps.

Convergence detection

After each record call the tracker compares the new loss against the previous best. If the improvement (previous_best − new_loss) is below OptimizationHistoryConfig::convergence_threshold the consecutive_no_progress counter is incremented; otherwise it is reset to zero. Once the counter reaches OptimizationHistoryConfig::convergence_patience the status transitions to ConvergenceStatus::Converged; once it reaches patience / 2 (integer division) it transitions to ConvergenceStatus::PossiblyConverged.

Fields

steps: Vec<OptimizationStep>

Retained optimization steps (oldest first).

config: OptimizationHistoryConfig

Configuration.

best_loss: f64

Best (lowest) loss seen so far.

best_step: u64

Step index that achieved best_loss.

consecutive_no_progress: usize

Number of consecutive steps that did not improve loss by at least convergence_threshold.

Implementations

impl TensorOptimizationHistory

pub fn new(config: OptimizationHistoryConfig) -> Self

Creates a new history tracker with the given configuration.

pub fn record(&mut self, step: OptimizationStep)

Records a new optimization step.

If the history buffer is full the oldest entry is evicted. Best loss tracking and consecutive-no-progress counting are updated accordingly.

pub fn convergence_status(&self) -> ConvergenceStatus

Returns the current ConvergenceStatus based on the consecutive no-progress counter.

pub fn recent_improvement(&self, n: usize) -> f64

Computes the total loss improvement over the last n steps.

Returns first_loss_in_window − last_loss_in_window; a positive value means the loss is decreasing (improving). Returns 0.0 when fewer than two steps are available.

pub fn avg_gradient_norm(&self) -> f64

Returns the arithmetic mean of all recorded gradient_norm values.

Returns 0.0 when no steps have been recorded.

pub fn stats(&self) -> HistoryStats

Computes and returns aggregated HistoryStats for the current history buffer.

pub fn last_step(&self) -> Option<&OptimizationStep>

Returns a reference to the most recently recorded step, or None if the history is empty.

pub fn reset(&mut self)

Resets all history and tracking state.