ray.tune.schedulers.PopulationBasedTrainingReplay#

class ray.tune.schedulers.PopulationBasedTrainingReplay(policy_file: str)[source]#

Bases: FIFOScheduler

Replays a Population Based Training run.

Population Based Training does not return a single hyperparameter configuration, but rather a schedule of configurations. For instance, PBT might discover that a larger learning rate leads to good results in the first training iterations, but that a smaller learning rate is preferable later.

This scheduler enables replaying these parameter schedules from a finished PBT run. This requires that population based training has been run with log_config=True, which is the default setting.

The scheduler will only accept and train a single trial. It will start with the initial config of the existing trial and update the config according to the schedule.

Parameters:: policy_file – The PBT policy file. Usually this is stored in ~/ray_results/experiment_name/pbt_policy_xxx.txt where xxx is the trial ID.

Example:

# Replaying a result from ray.tune.examples.pbt_convnet_example
from ray import train, tune

from ray.tune.examples.pbt_convnet_example import PytorchTrainable
from ray.tune.schedulers import PopulationBasedTrainingReplay

replay = PopulationBasedTrainingReplay(
    "~/ray_results/pbt_test/pbt_policy_XXXXX_00001.txt")

tuner = tune.Tuner(
    PytorchTrainable,
    run_config=train.RunConfig(
        stop={"training_iteration": 100}
    ),
    tune_config=tune.TuneConfig(
        scheduler=replay,
    ),
)
tuner.fit()

Methods

`restore`	Restore trial scheduler from checkpoint.
`save`	Save trial scheduler to a checkpoint
`set_search_properties`	Pass search properties to scheduler.

Attributes

`CONTINUE`	Status for continuing trial execution
`NOOP`
`PAUSE`	Status for pausing trial execution
`STOP`	Status for stopping trial execution
`metric`
`supports_buffered_results`