API: trainers

The training loop. Module: autoStructN2V.trainers.

AutoStructN2VTrainer (trainers/auto_struct_n2v.py)

AutoStructN2VTrainer(
    model, optimizer, scheduler, device, hparams, stage,
    experiment_name='experiment',
    clean_stack=None,            # optional GT: enables the aux PSNR diagnostic
    val_indices=None, aux_psnr_indices=None,
    norm_stats=None,
)

• stage accepts 'n2v' / 'structn2v' (legacy 'stage1' / 'stage2' still work)
• Masked validation loss drives the scheduler, early stopping, and best-checkpoint selection; the auxiliary clean-PSNR (when clean_stack is given) is logged only
• Writes TensorBoard logs under the experiment's logs/<timestamp>/
• Saves the best-validation checkpoint as the run's model.pth

BaseTrainer (trainers/base.py)

The epoch/validation scaffolding AutoStructN2VTrainer builds on — subclass it for custom losses or logging.

EarlyStopping (trainers/callbacks.py)

Patience-based stopper on the masked validation loss (early_stopping_patience, default 10). The publication runs disable early stopping and train the fixed 100-epoch budget; keeping it enabled is convenient for exploration since the epoch count then acts as a ceiling.

Related

• Training user guide — the loop in context
• Pipeline API — how dataloaders and models are assembled around the trainer