def main():
"""
Run the atari test
"""
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--num-timesteps', type=int, default=int(1e7))
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
policy = partial(CnnPolicy, dueling=args.dueling == 1)
# model = DQN(
# env=env,
# policy=policy,
# learning_rate=1e-4,
# buffer_size=10000,
# exploration_fraction=0.1,
# exploration_final_eps=0.01,
# train_freq=4,
# learning_starts=10000,
# target_network_update_freq=1000,
# gamma=0.99,
# prioritized_replay=bool(args.prioritized),
# prioritized_replay_alpha=args.prioritized_replay_alpha,
# )
model = DQN(
env=env,
policy_class=CnnPolicy,
learning_rate=1e-4,
buffer_size=10000,
double_q=False,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
dueling=True,
train_freq=4,
learning_starts=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
target_network_update_freq=1000,
model_path='atari_Breakout_duel'
)
# model.learn(total_timesteps=args.num_timesteps, seed=args.seed)
model.load('atari_Breakout_duel')
model.evaluate(100)
env.close()
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
# Create envs.
env = gym.make(env_id)
env = Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(0)))
if evaluation:
eval_env = gym.make(env_id)
eval_env = Monitor(eval_env, os.path.join(logger.get_dir(), 'gym_eval'))
env = Monitor(env, None)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
# Configure components.
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm)
# Seed everything to make things reproducible.
logger.info('seed={}, logdir={}'.format(seed, logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
start_time = time.time()
training.train(env=env, eval_env=eval_env, param_noise=param_noise,
action_noise=action_noise, actor=actor, critic=critic, memory=memory, **kwargs)
env.close()
if eval_env is not None:
eval_env.close()
logger.info('total runtime: {}s'.format(time.time() - start_time))
def main():
"""
Run the atari test
"""
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env',
help='environment ID',
default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--num-timesteps', type=int, default=int(1e7))
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env.action_space.seed(args.seed)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
model = DQN(env=env,
policy_class=CnnPolicy,
buffer_size=10000,
learning_rate=1e-4,
learning_starts=10000,
target_network_update_freq=1000,
train_freq=4,
exploration_final_eps=0.01,
exploration_fraction=0.1,
prioritized_replay=True,
model_path='atari_test_Breakout')
model.learn(total_timesteps=args.num_timesteps)
env.close()
def play_single_thread(
actor, critic, target_actor, target_critic, args, prepare_fn,
global_episode, global_update_step, episodes_queue,
best_reward):
workerseed = args.seed + 241 * args.thread
set_global_seeds(workerseed)
args.logdir = "{}/thread_{}".format(args.logdir, args.thread)
create_if_need(args.logdir)
act_fn, _, save_fn = prepare_fn(actor, critic, target_actor, target_critic, args)
logger = Logger(args.logdir)
env = create_env(args)
random_process = create_random_process(args)
epsilon_cycle_len = random.randint(args.epsilon_cycle_len // 2, args.epsilon_cycle_len * 2)
epsilon_decay_fn = create_decay_fn(
"cycle",
initial_value=args.initial_epsilon,
final_value=args.final_epsilon,
cycle_len=epsilon_cycle_len,
num_cycles=args.max_episodes // epsilon_cycle_len)
episode = 1
step = 0
start_time = time.time()
while global_episode.value < args.max_episodes * (args.num_threads - args.num_train_threads) \
and global_update_step.value < args.max_update_steps * args.num_train_threads:
if episode % 100 == 0:
env = create_env(args)
seed = random.randrange(2 ** 32 - 2)
epsilon = min(args.initial_epsilon, max(args.final_epsilon, epsilon_decay_fn(episode)))
episode_metrics = {
"reward": 0.0,
"step": 0,
"epsilon": epsilon
}
observation = env.reset(seed=seed, difficulty=args.difficulty)
random_process.reset_states()
done = False
replay = []
while not done:
action = act_fn(observation, noise=epsilon * random_process.sample())
next_observation, reward, done, _ = env.step(action)
replay.append((observation, action, reward, next_observation, done))
episode_metrics["reward"] += reward
episode_metrics["step"] += 1
observation = next_observation
episodes_queue.put(replay)
episode += 1
global_episode.value += 1
if episode_metrics["reward"] > best_reward.value:
best_reward.value = episode_metrics["reward"]
logger.scalar_summary("best reward", best_reward.value, episode)
if episode_metrics["reward"] > 15.0 * args.reward_scale:
save_fn(episode)
step += episode_metrics["step"]
elapsed_time = time.time() - start_time
for key, value in episode_metrics.items():
logger.scalar_summary(key, value, episode)
logger.scalar_summary(
"episode per minute",
episode / elapsed_time * 60,
episode)
logger.scalar_summary(
"step per second",
step / elapsed_time,
episode)
if elapsed_time > 86400 * args.max_train_days:
global_episode.value = args.max_episodes * (args.num_threads - args.num_train_threads) + 1
raise KeyboardInterrupt
def train_single_thread(
actor, critic, target_actor, target_critic, args, prepare_fn,
global_episode, global_update_step, episodes_queue):
workerseed = args.seed + 241 * args.thread
set_global_seeds(workerseed)
args.logdir = "{}/thread_{}".format(args.logdir, args.thread)
create_if_need(args.logdir)
_, update_fn, save_fn = prepare_fn(actor, critic, target_actor, target_critic, args)
logger = Logger(args.logdir)
buffer = create_buffer(args)
if args.prioritized_replay:
beta_deacy_fn = create_decay_fn(
"linear",
initial_value=args.prioritized_replay_beta0,
final_value=1.0,
max_step=args.max_update_steps)
actor_learning_rate_decay_fn = create_decay_fn(
"linear",
initial_value=args.actor_lr,
final_value=args.actor_lr_end,
max_step=args.max_update_steps)
critic_learning_rate_decay_fn = create_decay_fn(
"linear",
initial_value=args.critic_lr,
final_value=args.critic_lr_end,
max_step=args.max_update_steps)
update_step = 0
received_examples = 1 # just hack
while global_episode.value < args.max_episodes * (args.num_threads - args.num_train_threads) \
and global_update_step.value < args.max_update_steps * args.num_train_threads:
actor_lr = actor_learning_rate_decay_fn(update_step)
critic_lr = critic_learning_rate_decay_fn(update_step)
actor_lr = min(args.actor_lr, max(args.actor_lr_end, actor_lr))
critic_lr = min(args.critic_lr, max(args.critic_lr_end, critic_lr))
while True:
try:
replay = episodes_queue.get_nowait()
for (observation, action, reward, next_observation, done) in replay:
buffer.add(observation, action, reward, next_observation, done)
received_examples += len(replay)
except py_queue.Empty:
break
if len(buffer) >= args.train_steps:
if args.prioritized_replay:
beta = beta_deacy_fn(update_step)
beta = min(1.0, max(args.prioritized_replay_beta0, beta))
(tr_observations, tr_actions, tr_rewards, tr_next_observations, tr_dones,
weights, batch_idxes) = \
buffer.sample(
batch_size=args.batch_size,
beta=beta)
else:
(tr_observations, tr_actions, tr_rewards, tr_next_observations, tr_dones) = \
buffer.sample(batch_size=args.batch_size)
weights, batch_idxes = np.ones_like(tr_rewards), None
step_metrics, step_info = update_fn(
tr_observations, tr_actions, tr_rewards,
tr_next_observations, tr_dones,
weights, actor_lr, critic_lr)
update_step += 1
global_update_step.value += 1
if args.prioritized_replay:
new_priorities = np.abs(step_info["td_error"]) + 1e-6
buffer.update_priorities(batch_idxes, new_priorities)
for key, value in step_metrics.items():
value = to_numpy(value)[0]
logger.scalar_summary(key, value, update_step)
logger.scalar_summary("actor lr", actor_lr, update_step)
logger.scalar_summary("critic lr", critic_lr, update_step)
if update_step % args.save_step == 0:
save_fn(update_step)
else:
time.sleep(1)
logger.scalar_summary("buffer size", len(buffer), global_episode.value)
logger.scalar_summary(
"updates per example",
update_step * args.batch_size / received_examples,
global_episode.value)
save_fn(update_step)
raise KeyboardInterrupt
def train_multi_thread(actor, critic, target_actor, target_critic, args, prepare_fn, best_reward):
workerseed = args.seed + 241 * args.thread
set_global_seeds(workerseed)
args.logdir = "{}/thread_{}".format(args.logdir, args.thread)
create_if_need(args.logdir)
act_fn, update_fn, save_fn = prepare_fn(actor, critic, target_actor, target_critic, args)
logger = Logger(args.logdir)
buffer = create_buffer(args)
if args.prioritized_replay:
beta_deacy_fn = create_decay_fn(
"linear",
initial_value=args.prioritized_replay_beta0,
final_value=1.0,
max_step=args.max_episodes)
env = create_env(args)
random_process = create_random_process(args)
actor_learning_rate_decay_fn = create_decay_fn(
"linear",
initial_value=args.actor_lr,
final_value=args.actor_lr_end,
max_step=args.max_episodes)
critic_learning_rate_decay_fn = create_decay_fn(
"linear",
initial_value=args.critic_lr,
final_value=args.critic_lr_end,
max_step=args.max_episodes)
epsilon_cycle_len = random.randint(args.epsilon_cycle_len // 2, args.epsilon_cycle_len * 2)
epsilon_decay_fn = create_decay_fn(
"cycle",
initial_value=args.initial_epsilon,
final_value=args.final_epsilon,
cycle_len=epsilon_cycle_len,
num_cycles=args.max_episodes // epsilon_cycle_len)
episode = 0
step = 0
start_time = time.time()
while episode < args.max_episodes:
if episode % 100 == 0:
env = create_env(args)
seed = random.randrange(2 ** 32 - 2)
actor_lr = actor_learning_rate_decay_fn(episode)
critic_lr = critic_learning_rate_decay_fn(episode)
epsilon = min(args.initial_epsilon, max(args.final_epsilon, epsilon_decay_fn(episode)))
episode_metrics = {
"value_loss": 0.0,
"policy_loss": 0.0,
"reward": 0.0,
"step": 0,
"epsilon": epsilon
}
observation = env.reset(seed=seed, difficulty=args.difficulty)
random_process.reset_states()
done = False
while not done:
action = act_fn(observation, noise=epsilon*random_process.sample())
next_observation, reward, done, _ = env.step(action)
buffer.add(observation, action, reward, next_observation, done)
episode_metrics["reward"] += reward
episode_metrics["step"] += 1
if len(buffer) >= args.train_steps:
if args.prioritized_replay:
(tr_observations, tr_actions, tr_rewards, tr_next_observations, tr_dones,
weights, batch_idxes) = \
buffer.sample(batch_size=args.batch_size, beta=beta_deacy_fn(episode))
else:
(tr_observations, tr_actions, tr_rewards, tr_next_observations, tr_dones) = \
buffer.sample(batch_size=args.batch_size)
weights, batch_idxes = np.ones_like(tr_rewards), None
step_metrics, step_info = update_fn(
tr_observations, tr_actions, tr_rewards,
tr_next_observations, tr_dones,
weights, actor_lr, critic_lr)
if args.prioritized_replay:
new_priorities = np.abs(step_info["td_error"]) + 1e-6
buffer.update_priorities(batch_idxes, new_priorities)
for key, value in step_metrics.items():
value = to_numpy(value)[0]
episode_metrics[key] += value
observation = next_observation
episode += 1
if episode_metrics["reward"] > 15.0 * args.reward_scale \
and episode_metrics["reward"] > best_reward.value:
best_reward.value = episode_metrics["reward"]
logger.scalar_summary("best reward", best_reward.value, episode)
save_fn(episode)
step += episode_metrics["step"]
elapsed_time = time.time() - start_time
for key, value in episode_metrics.items():
value = value if "loss" not in key else value / episode_metrics["step"]
logger.scalar_summary(key, value, episode)
logger.scalar_summary(
"episode per minute",
episode / elapsed_time * 60,
episode)
logger.scalar_summary(
"step per second",
step / elapsed_time,
episode)
logger.scalar_summary("actor lr", actor_lr, episode)
logger.scalar_summary("critic lr", critic_lr, episode)
if episode % args.save_step == 0:
save_fn(episode)
if elapsed_time > 86400 * args.max_train_days:
episode = args.max_episodes + 1
save_fn(episode)
raise KeyboardInterrupt
def learn(env,
network,
seed=None,
lr=5e-5,
total_timesteps=100000,
buffer_size=500000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=1,
batch_size=32,
print_freq=10,
checkpoint_freq=100000,
checkpoint_path=None,
learning_starts=0,
gamma=0.99,
target_network_update_freq=10000,
prioritized_replay=True,
prioritized_replay_alpha=0.4,
prioritized_replay_beta0=0.6,
prioritized_replay_beta_iters=None,
prioritized_replay_eps=1e-3,
param_noise=False,
callback=None,
load_path=None,
load_idx=None,
demo_path=None,
n_step=10,
demo_prioritized_replay_eps=1.0,
pre_train_timesteps=750000,
epsilon_schedule="constant",
**network_kwargs):
# Create all the functions necessary to train the model
set_global_seeds(seed)
q_func = build_q_func(network, **network_kwargs)
with tf.device('/GPU:0'):
model = DQfD(q_func=q_func,
observation_shape=env.observation_space.shape,
num_actions=env.action_space.n,
lr=lr,
grad_norm_clipping=10,
gamma=gamma,
param_noise=param_noise)
# Load model from checkpoint
if load_path is not None:
load_path = osp.expanduser(load_path)
ckpt = tf.train.Checkpoint(model=model)
manager = tf.train.CheckpointManager(ckpt, load_path, max_to_keep=None)
if load_idx is None:
ckpt.restore(manager.latest_checkpoint)
print("Restoring from {}".format(manager.latest_checkpoint))
else:
ckpt.restore(manager.checkpoints[load_idx])
print("Restoring from {}".format(manager.checkpoints[load_idx]))
# Setup demo trajectory
assert demo_path is not None
with open(demo_path, "rb") as f:
trajectories = pickle.load(f)
# Create the replay buffer
replay_buffer = PrioritizedReplayBuffer(buffer_size,
prioritized_replay_alpha)
if prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=prioritized_replay_beta0,
final_p=1.0)
temp_buffer = deque(maxlen=n_step)
is_demo = True
for epi in trajectories:
for obs, action, rew, new_obs, done in epi:
obs, new_obs = np.expand_dims(
np.array(obs), axis=0), np.expand_dims(np.array(new_obs),
axis=0)
if n_step:
temp_buffer.append((obs, action, rew, new_obs, done, is_demo))
if len(temp_buffer) == n_step:
n_step_sample = get_n_step_sample(temp_buffer, gamma)
replay_buffer.demo_len += 1
replay_buffer.add(*n_step_sample)
else:
replay_buffer.demo_len += 1
replay_buffer.add(obs[0], action, rew, new_obs[0], float(done),
float(is_demo))
logger.log("trajectory length:", replay_buffer.demo_len)
# Create the schedule for exploration
if epsilon_schedule == "constant":
exploration = ConstantSchedule(exploration_final_eps)
else: # not used
exploration = LinearSchedule(schedule_timesteps=int(
exploration_fraction * total_timesteps),
initial_p=1.0,
final_p=exploration_final_eps)
model.update_target()
# ============================================== pre-training ======================================================
start = time()
num_episodes = 0
temp_buffer = deque(maxlen=n_step)
for t in tqdm(range(pre_train_timesteps)):
# sample and train
experience = replay_buffer.sample(batch_size,
beta=prioritized_replay_beta0)
batch_idxes = experience[-1]
if experience[6] is None: # for n_step = 0
obses_t, actions, rewards, obses_tp1, dones, is_demos = tuple(
map(tf.constant, experience[:6]))
obses_tpn, rewards_n, dones_n = None, None, None
weights = tf.constant(experience[-2])
else:
obses_t, actions, rewards, obses_tp1, dones, is_demos, obses_tpn, rewards_n, dones_n, weights = tuple(
map(tf.constant, experience[:-1]))
td_errors, n_td_errors, loss_dq, loss_n, loss_E, loss_l2, weighted_error = model.train(
obses_t, actions, rewards, obses_tp1, dones, is_demos, weights,
obses_tpn, rewards_n, dones_n)
# Update priorities
new_priorities = np.abs(td_errors) + np.abs(
n_td_errors) + demo_prioritized_replay_eps
replay_buffer.update_priorities(batch_idxes, new_priorities)
# Update target network periodically
if t > 0 and t % target_network_update_freq == 0:
model.update_target()
# Logging
elapsed_time = timedelta(time() - start)
if print_freq is not None and t % 10000 == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward", 0)
logger.record_tabular("max 100 episode reward", 0)
logger.record_tabular("min 100 episode reward", 0)
logger.record_tabular("demo sample rate", 1)
logger.record_tabular("epsilon", 0)
logger.record_tabular("loss_td", np.mean(loss_dq.numpy()))
logger.record_tabular("loss_n_td", np.mean(loss_n.numpy()))
logger.record_tabular("loss_margin", np.mean(loss_E.numpy()))
logger.record_tabular("loss_l2", np.mean(loss_l2.numpy()))
logger.record_tabular("losses_all", weighted_error.numpy())
logger.record_tabular("% time spent exploring",
int(100 * exploration.value(t)))
logger.record_tabular("pre_train", True)
logger.record_tabular("elapsed time", elapsed_time)
logger.dump_tabular()
# ============================================== exploring =========================================================
sample_counts = 0
demo_used_counts = 0
episode_rewards = deque(maxlen=100)
this_episode_reward = 0.
best_score = 0.
saved_mean_reward = None
is_demo = False
obs = env.reset()
# Always mimic the vectorized env
obs = np.expand_dims(np.array(obs), axis=0)
reset = True
for t in tqdm(range(total_timesteps)):
if callback is not None:
if callback(locals(), globals()):
break
kwargs = {}
if not param_noise:
update_eps = tf.constant(exploration.value(t))
update_param_noise_threshold = 0.
else: # not used
update_eps = tf.constant(0.)
update_param_noise_threshold = -np.log(1. - exploration.value(t) +
exploration.value(t) /
float(env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
action, epsilon, _, _ = model.step(tf.constant(obs),
update_eps=update_eps,
**kwargs)
action = action[0].numpy()
reset = False
new_obs, rew, done, _ = env.step(action)
# Store transition in the replay buffer.
new_obs = np.expand_dims(np.array(new_obs), axis=0)
if n_step:
temp_buffer.append((obs, action, rew, new_obs, done, is_demo))
if len(temp_buffer) == n_step:
n_step_sample = get_n_step_sample(temp_buffer, gamma)
replay_buffer.add(*n_step_sample)
else:
replay_buffer.add(obs[0], action, rew, new_obs[0], float(done), 0.)
obs = new_obs
# invert log scaled score for logging
this_episode_reward += np.sign(rew) * (np.exp(np.sign(rew) * rew) - 1.)
if done:
num_episodes += 1
obs = env.reset()
obs = np.expand_dims(np.array(obs), axis=0)
episode_rewards.append(this_episode_reward)
reset = True
if this_episode_reward > best_score:
best_score = this_episode_reward
ckpt = tf.train.Checkpoint(model=model)
manager = tf.train.CheckpointManager(ckpt,
'./best_model',
max_to_keep=1)
manager.save(t)
logger.log("saved best model")
this_episode_reward = 0.0
if t % train_freq == 0:
experience = replay_buffer.sample(batch_size,
beta=beta_schedule.value(t))
batch_idxes = experience[-1]
if experience[6] is None: # for n_step = 0
obses_t, actions, rewards, obses_tp1, dones, is_demos = tuple(
map(tf.constant, experience[:6]))
obses_tpn, rewards_n, dones_n = None, None, None
weights = tf.constant(experience[-2])
else:
obses_t, actions, rewards, obses_tp1, dones, is_demos, obses_tpn, rewards_n, dones_n, weights = tuple(
map(tf.constant, experience[:-1]))
td_errors, n_td_errors, loss_dq, loss_n, loss_E, loss_l2, weighted_error = model.train(
obses_t, actions, rewards, obses_tp1, dones, is_demos, weights,
obses_tpn, rewards_n, dones_n)
new_priorities = np.abs(td_errors) + np.abs(
n_td_errors
) + demo_prioritized_replay_eps * is_demos + prioritized_replay_eps * (
1. - is_demos)
replay_buffer.update_priorities(batch_idxes, new_priorities)
# for logging
sample_counts += batch_size
demo_used_counts += np.sum(is_demos)
if t % target_network_update_freq == 0:
# Update target network periodically.
model.update_target()
if t % checkpoint_freq == 0:
save_path = checkpoint_path
ckpt = tf.train.Checkpoint(model=model)
manager = tf.train.CheckpointManager(ckpt,
save_path,
max_to_keep=10)
manager.save(t)
logger.log("saved checkpoint")
elapsed_time = timedelta(time() - start)
if done and num_episodes > 0 and num_episodes % print_freq == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward",
np.mean(episode_rewards))
logger.record_tabular("max 100 episode reward",
np.max(episode_rewards))
logger.record_tabular("min 100 episode reward",
np.min(episode_rewards))
logger.record_tabular("demo sample rate",
demo_used_counts / sample_counts)
logger.record_tabular("epsilon", epsilon.numpy())
logger.record_tabular("loss_td", np.mean(loss_dq.numpy()))
logger.record_tabular("loss_n_td", np.mean(loss_n.numpy()))
logger.record_tabular("loss_margin", np.mean(loss_E.numpy()))
logger.record_tabular("loss_l2", np.mean(loss_l2.numpy()))
logger.record_tabular("losses_all", weighted_error.numpy())
logger.record_tabular("% time spent exploring",
int(100 * exploration.value(t)))
logger.record_tabular("pre_train", False)
logger.record_tabular("elapsed time", elapsed_time)
logger.dump_tabular()
return model
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
if evaluation and rank == 0:
eval_env = gym.make(env_id)
eval_env = Monitor(eval_env, os.path.join(logger.get_dir(), 'gym_eval'))
env = Monitor(env, None)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
# Configure components.
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm)
# Seed everything to make things reproducible.
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed, logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
if rank == 0:
start_time = time.time()
training.train(env=env, eval_env=eval_env, param_noise=param_noise,
action_noise=action_noise, actor=actor, critic=critic, memory=memory, **kwargs)
env.close()
if eval_env is not None:
eval_env.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
def run(mode, render, render_eval, verbose_eval, sanity_run, env_kwargs,
model_kwargs, train_kwargs):
if sanity_run:
# Mode to sanity check the basic code.
# Fixed seed and logging dir.
# Dynamic setting of nb_rollout_steps and nb_train_steps in training.train() is disabled.
print('SANITY CHECK MODE!!!')
# Configure MPI, logging, random seeds, etc.
mpi_rank = MPI.COMM_WORLD.Get_rank()
mpi_size = MPI.COMM_WORLD.Get_size()
if mpi_rank == 0:
logger.configure(dir='logs' if sanity_run else datetime.datetime.now().
strftime("train_%m%d_%H%M"))
logdir = logger.get_dir()
else:
logger.set_level(logger.DISABLED)
logdir = None
logdir = MPI.COMM_WORLD.bcast(logdir, root=0)
start_time = time.time()
# fixed seed when running sanity check, same seed hourly for training.
seed = 1000000 * mpi_rank
seed += int(start_time) // 3600 if not sanity_run else 0
seed_list = MPI.COMM_WORLD.gather(seed, root=0)
logger.info('mpi_size {}: seeds={}, logdir={}'.format(
mpi_size, seed_list, logger.get_dir()))
# Create envs.
envs = []
if mode in [MODE_TRAIN]:
train_env = cust_env.ProsEnvMon(
visualize=render,
seed=seed,
fn_step=None,
fn_epis=logdir and os.path.join(logdir, '%d' % mpi_rank),
reset_dflt_interval=2,
**env_kwargs)
logger.info('action, observation space:', train_env.action_space.shape,
train_env.observation_space.shape)
envs.append(train_env)
else:
train_env = None
# Always run eval_env, either in evaluation mode during MODE_TRAIN, or MODE_SAMPLE, MODE_TEST.
# Reset to random states (reset_dflt_interval=0) in MODE_SAMPLE ,
# Reset to default state (reset_dflt_interval=1) in evaluation of MODE_TRAIN, or MODE_TEST
reset_dflt_interval = 0 if mode in [MODE_SAMPLE] else 1
eval_env = cust_env.ProsEnvMon(
visualize=render_eval,
seed=seed,
fn_step=logdir and os.path.join(logdir, 'eval_step_%d.csv' % mpi_rank),
fn_epis=logdir and os.path.join(logdir, 'eval_%d' % mpi_rank),
reset_dflt_interval=reset_dflt_interval,
verbose=verbose_eval,
**env_kwargs)
envs.append(eval_env)
# Create DDPG agent
tf.reset_default_graph()
set_global_seeds(seed)
assert (eval_env is not None), 'Empty Eval Environment!'
action_range = (min(eval_env.action_space.low),
max(eval_env.action_space.high))
logger.info('\naction_range', action_range)
nb_demo_kine, nb_key_states = eval_env.obs_cust_params
agent = ddpg.DDPG(eval_env.observation_space.shape,
eval_env.action_space.shape,
nb_demo_kine,
nb_key_states,
action_range=action_range,
save_ckpt=mpi_rank == 0,
**model_kwargs)
logger.debug('Using agent with the following configuration:')
logger.debug(str(agent.__dict__.items()))
# Set up agent mimic reward interface, for environment
for env in envs:
env.set_agent_intf_fp(agent.get_mimic_rwd)
# Run..
logger.info('\nEnv params:', env_kwargs)
logger.info('Model params:', model_kwargs)
if mode == MODE_TRAIN:
logger.info('Start training', train_kwargs)
training.train(train_env,
eval_env,
agent,
render=render,
render_eval=render_eval,
sanity_run=sanity_run,
**train_kwargs)
elif mode == MODE_SAMPLE:
sampling.sample(eval_env, agent, render=render_eval, **train_kwargs)
else:
training.test(eval_env, agent, render_eval=render_eval, **train_kwargs)
# Close up.
if train_env:
train_env.close()
if eval_env:
eval_env.close()
mpi_complete(start_time, mpi_rank, mpi_size, non_blocking_mpi=True)
def train(args):
import baselines.baselines_common.tf_util as U
sess = U.single_threaded_session()
sess.__enter__()
if args.restore_args_from is not None:
args = restore_params(args)
rank = MPI.COMM_WORLD.Get_rank()
workerseed = args.seed + 241 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
def policy_fn(name, ob_space, ac_space):
return Actor(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=args.hid_size,
num_hid_layers=args.num_hid_layers,
noise_type=args.noise_type)
env = create_env(args)
env.seed(workerseed)
if rank == 0:
create_if_need(args.logdir)
with open("{}/args.json".format(args.logdir), "w") as fout:
json.dump(vars(args),
fout,
indent=4,
ensure_ascii=False,
sort_keys=True)
try:
args.thread = rank
if args.agent == "trpo":
trpo.learn(env,
policy_fn,
args,
timesteps_per_batch=1024,
gamma=args.gamma,
lam=0.98,
max_kl=0.01,
cg_iters=10,
cg_damping=0.1,
vf_iters=5,
vf_stepsize=1e-3)
elif args.agent == "ppo":
# optimal settings:
# timesteps_per_batch = optim_epochs * optim_batchsize
ppo.learn(env,
policy_fn,
args,
timesteps_per_batch=256,
gamma=args.gamma,
lam=0.95,
clip_param=0.2,
entcoeff=0.0,
optim_epochs=4,
optim_stepsize=3e-4,
optim_batchsize=64,
schedule='constant')
else:
raise NotImplementedError
except KeyboardInterrupt:
print("closing envs...")
env.close()