def _make_workers(self, env_creator, policy, config, num_workers):
"""Default factory method for a WorkerSet running under this Trainer.
Override this method by passing a custom `make_workers` into
`build_trainer`.
Args:
env_creator (callable): A function that return and Env given an env
config.
policy (class): The Policy class to use for creating the policies
of the workers.
config (dict): The Trainer's config.
num_workers (int): Number of remote rollout workers to create.
0 for local only.
remote_config_updates (Optional[List[dict]]): A list of config
dicts to update `config` with for each Worker (len must be
same as `num_workers`).
Returns:
WorkerSet: The created WorkerSet.
"""
return WorkerSet(env_creator,
policy,
config,
num_workers=num_workers,
logdir=self.logdir)
def _init(self, config: TrainerConfigDict, env_creator: EnvCreator):
# No `get_policy_class` function.
if get_policy_class is None:
# Default_policy must be provided (unless in multi-agent mode,
# where each policy can have its own default policy class).
if not config["multiagent"]["policies"]:
assert default_policy is not None
# Query the function for a class to use.
else:
self._policy_class = get_policy_class(config)
# If None returned, use default policy (must be provided).
if self._policy_class is None:
assert default_policy is not None
self._policy_class = default_policy
if before_init:
before_init(self)
# Creating all workers (excluding evaluation workers).
self.workers = WorkerSet(
env_creator=env_creator,
validate_env=validate_env,
policy_class=self._policy_class,
trainer_config=config,
num_workers=self.config["num_workers"],
)
self.train_exec_impl = self.execution_plan(
self.workers, config, **self._kwargs_for_execution_plan())
if after_init:
after_init(self)
def _make_workers(self, env_creator, policy, config, num_workers):
return WorkerSet(
env_creator,
policy,
config,
num_workers=num_workers,
logdir=self.logdir)
def __init__(self, num_sets, env_creator, policy, trainer_config=None,
num_workers_per_set=0, logdir=None, _setup=True):
self._worker_sets = {}
for i in range(num_sets):
self._worker_sets[i] = WorkerSet(
env_creator, policy, trainer_config, num_workers_per_set,
logdir, _setup)
