ray.tune.Trainable — Ray 2.45.0 (original) (raw)

class ray.tune.Trainable(config: Dict[str, Any] = None, logger_creator: Callable[[Dict[str, Any]], Logger] = None, storage: StorageContext | None = None)[source]#

Abstract class for trainable models, functions, etc.

A call to train() on a trainable will execute one logical iteration of training. As a rule of thumb, the execution time of one train call should be large enough to avoid overheads (i.e. more than a few seconds), but short enough to report progress periodically (i.e. at most a few minutes).

Calling save() should save the training state of a trainable to disk, and restore(path) should restore a trainable to the given state.

Generally you only need to implement setup, step,save_checkpoint, and load_checkpoint when subclassing Trainable.

Other implementation methods that may be helpful to override arelog_result, reset_config, cleanup, and _export_model.

Tune will convert this class into a Ray actor, which runs on a separate process. By default, Tune will also change the current working directory of this process to its corresponding trial-level log directory self.logdir. This is designed so that different trials that run on the same physical node won’t accidentally write to the same location and overstep each other.

The behavior of changing the working directory can be disabled by setting theRAY_CHDIR_TO_TRIAL_DIR=0 environment variable. This allows access to files in the original working directory, but relative paths should be used for read only purposes, and you must make sure that the directory is synced on all nodes if running on multiple machines.

The TUNE_ORIG_WORKING_DIR environment variable was the original workaround for accessing paths relative to the original working directory. This environment variable is deprecated, and the RAY_CHDIR_TO_TRIAL_DIR environment variable described above should be used instead.

This class supports checkpointing to and restoring from remote storage.

Methods

Attributes