def __init__(self,
config,
policy_params,
env_creator,
noise,
min_task_runtime=0.2):
self.min_task_runtime = min_task_runtime
self.config = config
self.policy_params = policy_params
self.noise = SharedNoiseTable(noise)
self.env = env_creator()
self.preprocessor = ModelCatalog.get_preprocessor(
self.env.spec.id, self.env.observation_space.shape)
self.preprocessor_shape = self.preprocessor.transform_shape(
self.env.observation_space.shape)
self.sess = utils.make_session(single_threaded=True)
self.policy = policies.GenericPolicy(self.env.observation_space,
self.env.action_space,
self.preprocessor,
**policy_params)
tf_util.initialize()
self.rs = np.random.RandomState()
assert (self.policy.needs_ob_stat == (self.config["calc_obstat_prob"]
!= 0))
def _init(self):
policy_params = {"ac_noise_std": 0.01}
env = self.env_creator()
preprocessor = ModelCatalog.get_preprocessor(
env.spec.id, env.observation_space.shape)
preprocessor_shape = preprocessor.transform_shape(
env.observation_space.shape)
self.sess = utils.make_session(single_threaded=False)
self.policy = policies.GenericPolicy(env.observation_space,
env.action_space, preprocessor,
**policy_params)
tf_util.initialize()
self.optimizer = optimizers.Adam(self.policy, self.config["stepsize"])
self.ob_stat = utils.RunningStat(preprocessor_shape, eps=1e-2)
# Create the shared noise table.
print("Creating shared noise table.")
noise_id = create_shared_noise.remote()
self.noise = SharedNoiseTable(ray.get(noise_id))
# Create the actors.
print("Creating actors.")
self.workers = [
Worker.remote(self.config, policy_params, self.env_creator,
noise_id) for _ in range(self.config["num_workers"])
]
self.episodes_so_far = 0
self.timesteps_so_far = 0
self.tstart = time.time()