rewardNormalization = "returns"
env = DummyVecEnv([
makeEnvLambda(args.gym_id,
args.seed,
normOb=normOb,
rewardNormalization=rewardNormalization,
clipOb=clipOb,
clipRew=clipRew,
gamma=args.gamma)
])
np.random.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
epRewTestPh = tf.placeholder(
tf.float32,
shape=None,
name='episode_test_real_reward_latest_mean_summary')
epRewTrainPh = tf.placeholder(
tf.float32,
shape=None,
name='episode_train_real_reward_latest_mean_summary')
epTotalRewPh = tf.placeholder(dtype,
"\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,
name='episode_reward_summary')
epRewLatestMeanPh = tf.placeholder(
tf.float32, shape=None, name='episode_reward_latest_mean_summary')
epLenPh = tf.placeholder(tf.float32,
shape=None,
name='episode_length_summary')
expVarPh = tf.placeholder(tf.float32,