Class BaseTrainProcess

cdslib/core/script/base_train.py:40–1202

r""" The base training procedure, including handling command line arguments through a yaml file, distributed data parallel, training loop, etc. It sets up a basic framework so that specific applications can easily extend. Example usage: .. code-block:: python if __name_

class BaseTrainProcess(ABC):
    r"""
    The base training procedure, including handling command line arguments
    through a yaml file, distributed data parallel, training loop, etc.

    It sets up a basic framework so that specific applications can easily extend.

    Example usage:
    .. code-block:: python
        if __name__ == '__main__':
            with TrainProcess() as trainer:
                trainer.run()


    Design logic:

    - To configure the settings, you can use the arguments of init or config file.
    But note that you should not use trainer file to configure the class.

    """

    def __init__(
            self,
            exp_tag: str = "exp",
            config_filename: str = None,
            trainer_filename: str = None,
            work_dir: str = ".",
            output_dir: str = None,
            rank: int = 0,
            n_gpus: int = 1,
            random_seed: int = 0,
            save_code: bool = True,
            exclude_dirs: T.List[str] = None,
            exp_tag_first: bool = False,
            open_tensorboard: bool = True,
            ddp_type: str = "ddp",
            start_epoch: int = 0,
            end_epoch: int = 1000000,
            max_train_epoch_batches: int = -1,
            max_valid_epoch_batches: int = -1,
            max_test_epoch_batches: int = -1,
            save_every_num_epoch: int = 1,
            validate_every_num_epoch: int = 1,
            test_every_num_epoch: int = 1,
            log_every_num_train_batch: int = 1,
            log_every_num_valid_batch: int = 1,
            log_every_num_test_batch: int = 1,
            visualize_every_num_train_batch: int = 1,
            visualize_every_num_valid_batch: int = 1,
            visualize_every_num_test_batch: int = 1,
            tensorboard_exe_path: str = "tensorboard",
            # overwrite_pretrained_names: T.List[str] = None,
            use_torchrun: bool = True,
            find_unused_parameters: bool = False,  # used for DDP, see _setup_for_distributed_learning
            *args,
            **kwargs,
    ):
        """