Skip to content

ResumeSettings

Full path: schola.scripts.rllib.train.settings.ResumeSettings

Dataclass for resume settings used in the RLlib training process. This class defines the parameters for resuming training from a saved checkpoint. This allows you to continue training from a previously saved model checkpoint instead of starting from scratch. This is useful for long training jobs or if you want to experiment with different hyperparameters without losing progress.

ResumeSettings(resume_from = None, reset_timestep = False)

Parameters

  • resume_from (Annotated)

  • reset_timestep (bool)

Methods

init

__init__(resume_from = None, reset_timestep = False)

Parameters

  • resume_from (Annotated)

  • reset_timestep (bool)

Attributes

reset_timestep

reset_timestep

Whether to treat -timesteps as additional steps to run beyond the checkpoint rather than an absolute lifetime cap. When False (default), -timesteps is the absolute num_env_steps_sampled_lifetime stop target, so the same command can resume training without modification. When True, the restored step count is added to -timesteps, matching the behaviour of SB3’s reset_num_timesteps=True.

resume_from

resume_from : Path to a trusted RLlib checkpoint to resume training from

the checkpoint directory that contains rllib_checkpoint.json (for example …/checkpoint_000000). You can also pass a single checkpoint file for older layouts. If set to None, training will start from scratch. Only use checkpoints from trusted sources because RLlib checkpoint metadata is read with pickle.

name

name