Struct StatRegressionDetector 

pub struct StatRegressionDetector {
    pub metric_name: String,
    pub baseline: StatBaselineStats,
    pub config: StatRegressionConfig,
    pub detection_history: Vec<StatRegressionEvent>,
}

Statistical regression detector for a single scalar training metric.

Uses a fixed StatBaselineStats as the reference distribution and emits StatRegressionEvents whenever a new observation deviates significantly from that baseline.

Example

use trustformers_debug::regression::statistical::{
    StatBaselineStats, StatRegressionConfig, StatRegressionDetector, StatRegressionDirection,
};

let samples = [1.0f64, 1.1, 0.9, 1.05, 0.95];
let baseline = StatRegressionDetector::build_baseline(&samples);

let config = StatRegressionConfig {
    z_score_threshold: 2.0,
    relative_threshold: 0.15,
    min_samples_for_detection: 2,
    direction: StatRegressionDirection::Either,
};

let mut detector = StatRegressionDetector::new("loss", baseline, config);
// Far-out value should trigger a regression
let event = detector.check_point(10, 5.0);
assert!(event.is_some());

Fields

metric_name: String

baseline: StatBaselineStats

config: StatRegressionConfig

detection_history: Vec<StatRegressionEvent>

Implementations

impl StatRegressionDetector

pub fn new( metric_name: &str, baseline: StatBaselineStats, config: StatRegressionConfig, ) -> Self

Creates a new detector for the named metric.

pub fn z_score(&self, value: f64) -> f64

Computes the z-score of a new observation relative to the baseline.

pub fn relative_change(&self, value: f64) -> f64

Computes the relative change (value − mean) / mean.

pub fn check_point( &mut self, step: u64, value: f64, ) -> Option<StatRegressionEvent>

Evaluates a new data point and returns a StatRegressionEvent if a regression is detected, or None otherwise.

A regression is detected when:

1. |z_score| > config.z_score_threshold and
2. |relative_change| > config.relative_threshold and
3. The direction constraint is satisfied.

Detection only fires once baseline.sample_count >= min_samples_for_detection.

pub fn recent_events(&self, n: usize) -> &[StatRegressionEvent]

Returns the last n events in detection history (oldest first).

pub fn build_baseline(samples: &[f64]) -> StatBaselineStats

Builds a StatBaselineStats from a slice of samples using Welford’s one-pass algorithm for numerically stable variance computation.

Returns a baseline with sample_count = 0 and all zeros when samples is empty.