peal.training.training_utils

Validation helpers shared by PEAL’s predictor training loops.

The one public function, calculate_validation_statistics(), walks validation dataloaders, measures the accuracy and error structure of a classifier and runs an explainer on every batch so that counterfactual statistics (flip rates, hints, indices, …) are collected alongside the plain validation metrics. Adaptors such as CFKD call it once per fine-tuning round.

Functions

calculate_validation_statistics

Compute accuracy / error statistics and explainer outputs on validation data.