def finish_actors(self):
for _ in range(self.num_actors):
self.command_queue.put('end', True)
for pi, p in enumerate(self.processes):
p.terminate()
p.join()
print('processes %d closed' % pi)
nb_inputs = env.observation_space.sample().shape[0]
nb_outputs = env.action_space.sample().shape[0] * 2
policy = Policy(nb_inputs, nb_outputs, hp.env_name, hp.normal, args)
hp.nb_directions = int(policy.theta.shape[0] * policy.theta.shape[1])
hp.nb_best_directions = int(hp.nb_directions / 2)
normalizer = Normalizer(nb_inputs)
print("start training")
train(env, policy, normalizer, hp, parentPipes, args)
if args.mp:
for parentPipe in parentPipes:
parentPipe.send([_CLOSE, "pay2"])
for p in processes:
p.join()
# --------------------------------------------------------------------------------
# STOCH2 Test
# env = sv.StochBulletEnv(render = True, gait = 'trot')
# hp = HyperParameters()
# nb_inputs = env.observation_space.shape[0]
# nb_outputs = env.action_space.shape[0] * 2
# args = 0
# policy = Policy(nb_inputs, nb_outputs, hp.env_name, 0, args)
# normalizer = Normalizer(nb_inputs)
# deltas = policy.sample_deltas()
# state = env.reset()
# i = 0
# hp.noise = 0.2
