ESMasterRunner

class maze.train.trainers.es.es_runners.ESMasterRunner(state_dict_dump_file: str, dump_interval: Optional[int], spaces_config_dump_file: str, normalization_samples: int, shared_noise_table_size: int)

Baseclass of ES training master runners (serves as basis for dev and other runners).

abstract create_distributed_rollouts(env: Union[maze.core.env.structured_env.StructuredEnv, maze.core.env.structured_env_spaces_mixin.StructuredEnvSpacesMixin], shared_noise: maze.train.trainers.es.es_shared_noise_table.SharedNoiseTable, agent_instance_seed: int) → maze.train.trainers.es.distributed.es_distributed_rollouts.ESDistributedRollouts

Abstract method, derived runners like ESDevRunner return an appropriate rollout generator.

Parameters

  • env – The one and only environment.

  • shared_noise – Noise table to be shared by all workers.

  • agent_instance_seed – The agent seed to be used.

Returns

A newly instantiated rollout generator.

run(n_epochs: Optional[int] = None, distributed_rollouts: Optional[maze.train.trainers.es.distributed.es_distributed_rollouts.ESDistributedRollouts] = None, model_selection: Optional[maze.train.trainers.common.model_selection.model_selection_base.ModelSelectionBase] = None) → None

(overrides TrainingRunner)

See run(). :param distributed_rollouts: The distribution interface for experience collection. :param n_epochs: Number of epochs to train. :param model_selection: Optional model selection class, receives model evaluation results.

setup(cfg: omegaconf.DictConfig) → None

(overrides TrainingRunner)

Setup the training master node.

shared_noise_table_size: int

Number of float values in the deterministically generated pseudo-random table (250.000.000 x 32bit floats = 1GB)