ray.rllib.algorithms.algorithm.Algorithm.save_to_path#

Algorithm.save_to_path(path: str | Path | None = None, *, state: Dict[str, Any] | None = None, filesystem: pyarrow.fs.FileSystem | None = None, use_msgpack: bool = False) → str#

Saves the state of the implementing class (or state) to path.

The state of the implementing class is always saved in the following format:

path/
    [component1]/
        [component1 subcomponentA]/
            ...
        [component1 subcomponentB]/
            ...
    [component2]/
            ...
    [cls.METADATA_FILE_NAME] (json)
    [cls.STATE_FILE_NAME] (pkl|msgpack)

The main logic is to loop through all subcomponents of this Checkpointable and call their respective save_to_path methods. Then save the remaining (non subcomponent) state to this Checkpointable’s STATE_FILE_NAME. In the exception that a component is a FaultTolerantActorManager instance, instead of calling save_to_path directly on that manager, the first healthy actor is interpreted as the component and its save_to_path method is called. Even if that actor is located on another node, the created file is automatically synced to the local node.

Parameters:

  • path – The path to the directory to save the state of the implementing class to. If path doesn’t exist or is None, then a new directory will be created (and returned).

  • state – An optional state dict to be used instead of getting a new state of the implementing class through self.get_state().

  • filesystem – PyArrow FileSystem to use to access data at the path. If not specified, this is inferred from the URI scheme of path.

  • use_msgpack – Whether the state file should be written using msgpack and msgpack_numpy (file extension is .msgpack), rather than pickle (file extension is .pkl).

Returns:

The path (str) where the state has been saved.