peal.training.trainers¶

Supervised training loop for PEAL predictors and predictor distillation.

ModelTrainer trains a classifier (or a generator exposing track_generator_performance) from a PredictorConfig: it builds the model, dataloaders, optimizer, criterions and TensorBoard logger, runs epochs with optional mixup, label smoothing and PGD adversarial training, applies an early-stopping/regularisation scheme and writes model.cpl, checkpoints/*.cpl and config.yaml into model_path. The distill_* helpers relabel datasets with a teacher predictor’s outputs and train a student on them, which is how CFKD/DiDAE turn a repaired predictor into a stand-alone model. calculate_test_accuracy() is the shared (worst-)group accuracy evaluator.

Functions

`calculate_test_accuracy`	Compute the accuracy of `model` on a dataloader, optionally per group.
`distill_1ofn_dataset`	Relabel `Image2ClassDataset` (folder-per-class) data with a predictor.
`distill_binary_dataset`	Relabel `Image2MixedDataset`-style datasets with a predictor's outputs.
`distill_dataloader_mixer`	Recursively relabel every loader inside a `DataloaderMixer`.
`distill_predictor`	Train a student predictor on data relabelled by `predictor`.
`get_predictor`	Instantiate the predictor described by a `PredictorConfig`.
`load_first_loadable`	First checkpoint in paths that actually loads, or None if none does.
`mixup`	Mix each sample with a random partner (Zhang et al. mixup).

Classes

ModelTrainer

Train a predictor from a PredictorConfig with early stopping.