mml.core.scripts.schedulers.transfer_scheduler

class TransferScheduler

Inherited from TrainingScheduler this scheduler supports the same subroutines: - model training - model prediction - model testing

But it adds the option to finetune an existing model by choosing a mode.pretrain_task in the config.

__init__(cfg: DictConfig)

after_training_hook(datamodule: LightningDataModule, model: LightningModule, trainer: Trainer, fold: int, task_name: str) → None

If necessary unfreezes the model backbone.

before_training_hook(datamodule: LightningDataModule, model: LightningModule, trainer: Trainer, fold: int, task_name: str) → None

Implements the weight loading logic.