Expand description
Diagnostics for online models.
This module provides bounded-memory diagnostic primitives that help upper-layer applications answer:
- How many samples has the model seen?
- What is the recent error?
- Is the model still warming up?
- Is the model beating its baseline?
- Are the model parameters healthy (no NaN / Infinity)?
- What is a reasonable prediction interval?
Diagnostics are intentionally decoupled from the core model traits. A model implementation remains free to return a plain prediction; the diagnostic wrappers here layer on top without polluting the base API.
Structs§
- Baseline
Comparator - Compares multiple models by their rolling error and tracks the current best.
- Comparator
Entry - A single tracked model entry inside a
BaselineComparator. - Model
Health Report - Snapshot of model parameter health.
- Online
Model Selector - Online model selector with cooling period and minimum sample requirements.
- Prediction
Interval - An immutable prediction interval
[lower, upper]. - Prediction
Report - An immutable snapshot of diagnostics for a single prediction.
- Prediction
Reporter - Diagnostic wrapper that integrates interval estimation, warmup tracking, and training summary statistics.
- Residual
Interval - Residual-based prediction interval estimator.
- Residual
Interval Config - Configuration for
ResidualInterval. - Selector
Config - Configuration for
OnlineModelSelector. - Training
Summary - Bounded-memory summary of a training process.
- Training
Summary Config - Configuration for
TrainingSummary. - Warmup
Config - Configuration for
WarmupTracker. - Warmup
Tracker - Bounded-memory warmup state tracker.
Enums§
- Confidence
- Coarse confidence level derived from the warmup state and baseline comparison.
- Switch
Reason - Why the active best entry changed (or could not be determined).
- Warmup
State - Lifecycle state of a model during warmup.