def _init(self, config, env_creator):
validate_config(config)
env_context = EnvContext(config["env_config"] or {}, worker_index=0)
env = env_creator(env_context)
self._policy_class = get_policy_class(config)
self.policy = self._policy_class(env.observation_space,
env.action_space, config)
self.optimizer = optimizers.SGD(self.policy, config["sgd_stepsize"])
self.rollouts_used = config["rollouts_used"]
self.num_rollouts = config["num_rollouts"]
self.report_length = config["report_length"]
# Create the shared noise table.
logger.info("Creating shared noise table.")
noise_id = create_shared_noise.remote(config["noise_size"])
self.noise = SharedNoiseTable(ray.get(noise_id))
# Create the actors.
logger.info("Creating actors.")
self.workers = [
Worker.remote(config, env_creator, noise_id, idx + 1)
for idx in range(config["num_workers"])
]
self.episodes_so_far = 0
self.reward_list = []
self.tstart = time.time()
def setup(self, config):
# Setup our config: Merge the user-supplied config (which could
# be a partial config dict with the class' default).
self.config = self.merge_trainer_configs(self.get_default_config(),
config,
self._allow_unknown_configs)
# Validate our config dict.
self.validate_config(self.config)
# Generate `self.env_creator` callable to create an env instance.
self.env_creator = self._get_env_creator_from_env_id(self._env_id)
# Generate the local env.
env_context = EnvContext(self.config["env_config"] or {},
worker_index=0)
env = self.env_creator(env_context)
self.callbacks = self.config["callbacks"]()
self._policy_class = get_policy_class(self.config)
self.policy = self._policy_class(env.observation_space,
env.action_space, self.config)
self.optimizer = optimizers.SGD(self.policy,
self.config["sgd_stepsize"])
self.rollouts_used = self.config["rollouts_used"]
self.num_rollouts = self.config["num_rollouts"]
self.report_length = self.config["report_length"]
# Create the shared noise table.
logger.info("Creating shared noise table.")
noise_id = create_shared_noise.remote(self.config["noise_size"])
self.noise = SharedNoiseTable(ray.get(noise_id))
# Create the actors.
logger.info("Creating actors.")
self.workers = [
Worker.remote(self.config, self.env_creator, noise_id, idx + 1)
for idx in range(self.config["num_workers"])
]
self.episodes_so_far = 0
self.reward_list = []
self.tstart = time.time()
def _init(self, config, env_creator):
policy_params = {"action_noise_std": 0.01}
env = env_creator(config["env_config"])
from ray.rllib import models
preprocessor = models.ModelCatalog.get_preprocessor(env)
self.sess = utils.make_session(single_threaded=False)
self.policy = GenericGaussianPolicy(self.sess, env.action_space,
env.observation_space,
preprocessor,
config["observation_filter"],
config["model"], **policy_params)
if config["optimizer_type"] == "adam":
self.optimizer = optimizers.Adam(self.policy, config["stepsize"])
elif config["optimizer_type"] == "sgd":
self.optimizer = optimizers.SGD(self.policy, config["stepsize"])
else:
raise ValueError("optimizer must in [adam, sgd].")
self.report_length = config["report_length"]
# Create the shared noise table.
logger.info("Creating shared noise table.")
noise_id = create_shared_noise.remote(config["noise_size"])
self.noise = SharedNoiseTable(ray.get(noise_id))
# Create the actors.
logger.info("Creating actors.")
self._workers = [
Worker.remote(config, policy_params, env_creator, noise_id)
for _ in range(config["num_workers"])
]
self.episodes_so_far = 0
self.reward_list = []
self.tstart = time.time()
