framework.middleware.ckpt_handler

ckpt_handler

CkptSaver

class ding.framework.middleware.ckpt_handler.CkptSaver(*args, **kwargs)

Overview:

The class used to save checkpoint data.

__call__(ctx: OnlineRLContext | OfflineRLContext) → None

Overview:

The method used to save checkpoint data. The checkpoint data will be saved in a file in following 3 cases: - When a multiple of self.train_freq iterations have elapsed since the beginning of training; - When the evaluation episode return is the best so far; - When task.finish is True.

Input of ctx:

• train_iter (int): Number of training iteration, i.e. the number of updating policy related network.

• eval_value (float): The episode return of current iteration.

__init__(policy: Policy, save_dir: str, train_freq: int | None = None, save_finish: bool = True)

Overview:

Initialize the CkptSaver.

Arguments:

• policy (Policy): Policy used to save the checkpoint.

• save_dir (str): The directory path to save ckpt.

• train_freq (int): Number of training iterations between each saving checkpoint data.

• save_finish (bool): Whether save final ckpt when task.finish = True.