peal.training

Training of predictors: trainers, criterions, loggers and their configs.

interfaces defines TrainingConfig and PredictorConfig (the yaml schema of a predictor run); trainers.ModelTrainer runs the epoch loop with optional adversarial, mixup and diffusion augmentation and also hosts the distillation helpers CFKD and DiDAE rely on; criterions maps the task.criterions keys to loss functions; loggers writes per-step and per-epoch metrics to TensorBoard and training_utils collects small shared helpers.

Submodules

criterions	Loss functions ("criterions") used by PEAL's predictor trainers.
interfaces	Pydantic configs of a predictor run: how it is trained and what it is.
loggers	TensorBoard logging for predictor (and generator) training.
trainers	Supervised training loop for PEAL predictors and predictor distillation.
training_utils	Validation helpers shared by PEAL's predictor training loops.