def __init__(self,
config,
policy_params,
env_name,
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 = gym.make(env_name)
self.sess = utils.make_session(single_threaded=True)
self.policy = policies.MujocoPolicy(self.env.observation_space,
self.env.action_space,
**policy_params)
tf_util.initialize()
self.rs = np.random.RandomState()
assert self.policy.needs_ob_stat == (self.config.calc_obstat_prob != 0)
# Create the shared noise table.
print("Creating shared noise table.")
noise_array = create_shared_noise()
noise = SharedNoiseTable(noise_array)
# Create the workers.
print("Creating workers.")
workers = [
Worker(config, policy_params, env_name, noise_array)
for _ in range(num_workers)
]
env = gym.make(env_name)
sess = utils.make_session(single_threaded=False)
policy = policies.MujocoPolicy(env.observation_space, env.action_space,
**policy_params)
tf_util.initialize()
optimizer = optimizers.Adam(policy, stepsize)
ob_stat = utils.RunningStat(env.observation_space.shape, eps=1e-2)
episodes_so_far = 0
timesteps_so_far = 0
tstart = time.time()
while True:
step_tstart = time.time()
theta = policy.get_trainable_flat()
assert theta.dtype == np.float32
# These rollouts could be done in parallel.