def initEnviroment(self):
print('Initialize env')
# update to your
env = EnvironmentWrapper()
env.initEnviroment(UNITY_ENVIROMENT)
print('observation space:', env.state_dim)
print('action space:', env.action_dim)
# get the default brain
print('Env init done')
self.config.model_name = MODEL
self.config.env = env
def func():
env = gym.make(gym_id)
env = TimeLimit(env, max_episode_steps = args.max_episode_len)
env = EnvironmentWrapper(env.env, normOb=normOb, rewardNormalization=rewardNormalization, clipOb=clipOb, clipRew=clipRew, **kwargs)
env.seed(args.seed)
env.action_space.seed(args.seed)
env.observation_space.seed(args.seed)
return env
rewardNormalization = None
normOb = False
if args.plus:
if args.plus_returns:
rewardNormalization = "returns"
if args.plus_plus:
if args.plus_plus_observation_normalization:
normOb = True
if args.plus_plus_observation_clipping > 0:
clipOb = args.plus_plus_observation_clipping
if args.plus_plus_reward_clipping > 0:
clipRew = args.plus_plus_reward_clipping
env = EnvironmentWrapper(env.env,
normOb=normOb,
rewardNormalization=rewardNormalization,
clipOb=clipOb,
clipRew=clipRew,
gamma=args.gamma)
np.random.seed(args.seed)
env.seed(args.seed)
env.action_space.seed(args.seed)
env.observation_space.seed(args.seed)
tf.set_random_seed(args.seed)
discreteActionsSpace = utils.is_discrete(env)
inputLength = env.observation_space.shape[0]
outputLength = env.action_space.n if discreteActionsSpace else env.action_space.shape[
0]
help='gradient l2 norm, negative value to turn it off')
args = parser.parse_args()
dtype = tf.float32
dtypeNp = np.float32
if not args.seed:
args.seed = int(time.time())
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
env = gym.make(args.gym_id)
if not args.minimal:
env = EnvironmentWrapper(env.env, normOb=True, rewardNormalization="returns", clipOb=10., clipRew=10., episodicMeanVarObs=False, episodicMeanVarRew=False, gamma=args.gamma)
else:
env = EnvironmentWrapper(env.env, normOb=False, rewardNormalization=None, clipOb=1000000., clipRew=1000000)
np.random.seed(args.seed)
env.seed(args.seed)
env.action_space.seed(args.seed)
env.observation_space.seed(args.seed)
tf.set_random_seed(args.seed)
discreteActionsSpace = utils.is_discrete(env)
inputLength = env.observation_space.shape[0]
outputLength = env.action_space.n if discreteActionsSpace else env.action_space.shape[0]
#summeries placeholders and summery scalar objects
args = parser.parse_args()
if not args.seed:
args.seed = int(time.time())
if args.buffer_size == -1:
args.buffer_size = args.total_train_steps
print(
"\nBuffer size not specified. Taking value of {} which is the same as total_train_steps, as suggested by the paper\n"
.format(args.buffer_size))
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
env = gym.make(args.gym_id)
env = EnvironmentWrapper(env.env, args.norm_obs, args.norm_rew,
args.clip_obs, args.clip_rew)
np.random.seed(args.seed)
env.seed(args.seed)
env.action_space.seed(args.seed)
env.observation_space.seed(args.seed)
tf.set_random_seed(args.seed)
if utils.is_discrete(env):
exit("TD3 can only be applied to continuous action space environments")
inputLength = env.observation_space.shape[0]
outputLength = env.action_space.shape[0]
#summeries placeholders and summery scalar objects
epRewPh = tf.placeholder(tf.float32,
shape=None,