def make_optimizer(ob_size, action_size, args):
''' dm_control optimizer requires different treatment
'''
optimizer_tasks = multiprocessing.JoinableQueue()
optimizer_results = multiprocessing.Queue()
optimizer_agent = optimization_agent.optimization_agent(
args,
ob_size, action_size,
optimizer_tasks, optimizer_results
)
optimizer_agent.start()
optimizer_tasks.put(parallel_util.START_SIGNAL)
optimizer_tasks.join()
starting_weights = optimizer_results.get()
return optimizer_tasks, optimizer_results, optimizer_agent, starting_weights
viz_item = ['avg_reward', 'entropy', 'kl', 'surr_loss',
'vf_loss', 'weight_l2_loss', 'learning_rate']
viz_win = {}
for item in viz_item:
viz_win[item] = None
if not args.dm:
args.max_pathlength = gym.spec(args.task).timestep_limit
learner_env = gym.make(args.task)
optimizer_tasks = multiprocessing.JoinableQueue()
optimizer_results = multiprocessing.Queue()
optimizer_agent = optimization_agent.optimization_agent(
args,
learner_env.observation_space.shape[0],
learner_env.action_space.shape[0],
optimizer_tasks,
optimizer_results
)
optimizer_agent.start()
# the rollouts agents
rollout_agent = rollout_master_agent.parallel_rollout_master_agent(
args,
learner_env.observation_space.shape[0],
learner_env.action_space.shape[0]
)
# start the training and rollouting process
optimizer_tasks.put(parallel_util.START_SIGNAL)
optimizer_tasks.join()