p.add_argument('--do_not_save', action='store_true')
p.add_argument('--learning_freq', type=int, default=50)
p.add_argument('--log_every_t_iter', type=int, default=50)
p.add_argument('--max_gradient', type=float, default=10.0)
p.add_argument('--n_iter', type=int, default=10000)
p.add_argument('--seed', type=int, default=0)
p.add_argument('--wait_until_rbuffer', type=int, default=1000)
args = p.parse_args()
# Handle the log directory and save the arguments.
logdir = 'out/' + args.envname + '/seed' + str(args.seed).zfill(2)
if args.do_not_save:
logdir = None
logz.configure_output_dir(logdir)
if logdir is not None:
with open(logdir + '/args.pkl', 'wb') as f:
pickle.dump(args, f)
print("Saving in logdir: {}".format(logdir))
# Other stuff for seeding and getting things set up.
tf.set_random_seed(args.seed)
np.random.seed(args.seed)
env = gym.make(args.envname)
test_env = gym.make(args.envname)
tf_config = tf.ConfigProto(inter_op_parallelism_threads=1,
intra_op_parallelism_threads=1)
sess = tf.Session(config=tf_config)
ddpg = DDPGAgent(sess, env, test_env, args)
ddpg.train()
episode_reward = 0
for step in range(500):
if episode >= 45:
env.render()
#action = agent.get_action(state, ou_noise)
action = agent.get_action(state, ou_noise)
#print("action = ", action)
new_state, reward, done, _ = env.step(action)
#print("new state =", new_state)
#new_state = new_state['observation']
#print("new state =", new_state)
agent.memory.push(state, action, reward, new_state, done)
if len(agent.memory) > batch_size:
agent.train(batch_size)
state = new_state
episode_reward += reward
if done:
if episode == 0:
sys.stdout.write(
"episode: {}, reward: {}, average _reward: {} \n".format(
episode, np.round(episode_reward, decimals=2), "nan"))
else:
sys.stdout.write(
"episode: {}, reward: {}, average _reward: {} \n".format(
episode, np.round(episode_reward, decimals=2),
np.mean(rewards[-10:])))
break
