def _enjoy():
# Launch the env with our helper function
env = launch_env()
print("Initialized environment")
# Wrappers
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env) # to make the images from 160x120x3 into 3x160x120
env = ActionWrapper(env)
env = DtRewardWrapper(env)
print("Initialized Wrappers")
state_dim = env.observation_space.shape
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
# Initialize policy
policy = DDPG(state_dim, action_dim, max_action, net_type="cnn")
policy.load(filename='ddpg', directory='reinforcement/pytorch/models/')
obs = env.reset()
done = False
while True:
while not done:
action = policy.predict(np.array(obs))
# Perform action
obs, reward, done, _ = env.step(action)
env.render()
done = False
obs = env.reset()
def _train(args):
if not os.path.exists("./results"):
os.makedirs("./results")
if not os.path.exists(args.model_dir):
os.makedirs(args.model_dir)
# Launch the env with our helper function
env = launch_env()
print("Initialized environment")
# Wrappers
env = ResizeWrapper(env)
env = NormalizeWrapper(env)
env = ImgWrapper(env) # to make the images from 160x120x3 into 3x160x120
env = ActionWrapper(env)
env = DtRewardWrapper(env)
print("Initialized Wrappers")
# Set seeds
seed(args.seed)
state_dim = env.observation_space.shape
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])
# Initialize policy
policy = DDPG(state_dim, action_dim, max_action, net_type="cnn")
replay_buffer = ReplayBuffer(args.replay_buffer_max_size)
print("Initialized DDPG")
# Evaluate untrained policy
evaluations = [evaluate_policy(env, policy)]
total_timesteps = 0
timesteps_since_eval = 0
episode_num = 0
done = True
episode_reward = None
env_counter = 0
reward = 0
episode_timesteps = 0
print("Starting training")
while total_timesteps < args.max_timesteps:
print("timestep: {} | reward: {}".format(total_timesteps, reward))
if done:
if total_timesteps != 0:
print(
("Total T: %d Episode Num: %d Episode T: %d Reward: %f") %
(total_timesteps, episode_num, episode_timesteps,
episode_reward))
policy.train(replay_buffer, episode_timesteps, args.batch_size,
args.discount, args.tau)
# Evaluate episode
if timesteps_since_eval >= args.eval_freq:
timesteps_since_eval %= args.eval_freq
evaluations.append(evaluate_policy(env, policy))
print("rewards at time {}: {}".format(
total_timesteps, evaluations[-1]))
if args.save_models:
policy.save(filename='{}_{}'.format(
'ddpg', total_timesteps),
directory=args.model_dir)
np.savez("./results/rewards.npz", evaluations)
# Reset environment
env_counter += 1
obs = env.reset()
done = False
episode_reward = 0
episode_timesteps = 0
episode_num += 1
# Select action randomly or according to policy
if total_timesteps < args.start_timesteps:
action = env.action_space.sample()
else:
action = policy.predict(np.array(obs))
if args.expl_noise != 0:
action = (action + np.random.normal(
0, args.expl_noise, size=env.action_space.shape[0])).clip(
env.action_space.low, env.action_space.high)
# Perform action
new_obs, reward, done, _ = env.step(action)
if episode_timesteps >= args.env_timesteps:
done = True
done_bool = 0 if episode_timesteps + 1 == args.env_timesteps else float(
done)
episode_reward += reward
# Store data in replay buffer
replay_buffer.add(obs, new_obs, action, reward, done_bool)
obs = new_obs
episode_timesteps += 1
total_timesteps += 1
timesteps_since_eval += 1
print("Training done, about to save..")
policy.save(filename='ddpg', directory=args.model_dir)
print("Finished saving..should return now!")