ImpalaRunner

class maze.train.trainers.impala.impala_runners.ImpalaRunner(state_dict_dump_file: str, dump_interval: int | None, spaces_config_dump_file: str, normalization_samples: int, eval_concurrency: int)

Common superclass for IMPALA runners, implementing the main training controls.

abstract create_distributed_eval_env(env_factory: Callable[[], StructuredEnv | StructuredEnvSpacesMixin], eval_concurrency: int, logging_prefix: str) → VectorEnv

The individual runners implement the setup of the distributed eval env

abstract create_distributed_rollout_actors(env_factory: Callable[[], StructuredEnv | StructuredEnvSpacesMixin | LogStatsEnv], policy: TorchPolicy, n_rollout_steps: int, n_actors: int, batch_size: int, queue_out_of_sync_factor: float, env_instance_seeds: List[int], agent_instance_seeds: List[int]) → DistributedActors

The individual runners implement the setup of the distributed training rollout actors

eval_concurrency: int

Number of concurrent evaluation envs

setup(cfg: omegaconf.DictConfig) → None

(overrides TrainingRunner)

See setup().