def _init_noise(self, noise_type):
action_noise = None
param_noise = None
if noise_type is not None:
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(self.action_shape[-1]),
sigma=float(stddev) * np.ones(self.action_shape[-1]))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(self.action_shape[-1]),
sigma=float(stddev) * np.ones(self.action_shape[-1]))
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
return action_noise, param_noise
def run_baselines(env, seed, log_dir):
'''
Create baselines model and training.
Replace the ddpg and its training with the algorithm you want to run.
:param env: Environment of the task.
:param seed: Random seed for the trial.
:param log_dir: Log dir path.
:return
'''
rank = MPI.COMM_WORLD.Get_rank()
seed = seed + 1000000 * rank
set_global_seeds(seed)
env.seed(seed)
# Set up logger for baselines
configure(dir=log_dir, format_strs=['stdout', 'log', 'csv', 'tensorboard'])
baselines_logger.info('rank {}: seed={}, logdir={}'.format(
rank, seed, baselines_logger.get_dir()))
# Set up params for baselines ddpg
nb_actions = env.action_space.shape[-1]
layer_norm = False
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions),
sigma=float(params['sigma']) *
np.ones(nb_actions))
memory = Memory(limit=params['replay_buffer_size'],
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)
training.train(env=env,
eval_env=None,
param_noise=None,
action_noise=action_noise,
actor=actor,
critic=critic,
memory=memory,
nb_epochs=params['n_epochs'],
nb_epoch_cycles=params['n_epoch_cycles'],
render_eval=False,
reward_scale=1.,
render=False,
normalize_returns=False,
normalize_observations=False,
critic_l2_reg=0,
actor_lr=params['policy_lr'],
critic_lr=params['qf_lr'],
popart=False,
gamma=params['discount'],
clip_norm=None,
nb_train_steps=params['n_train_steps'],
nb_rollout_steps=params['n_rollout_steps'],
nb_eval_steps=100,
batch_size=64)
return osp.join(log_dir, 'progress.csv')
def run(cfg, seed, noise_type, layer_norm, evaluation, architecture, **kwargs):
if MPI.COMM_WORLD.Get_rank() == 0:
dir_path = os.path.dirname(os.path.realpath(__file__))
logger.configure(dir_path, ['stdout'])
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = GRLEnv(cfg)
gym.logger.setLevel(logging.WARN)
env = MyMonitor(env, os.path.join(logger.get_dir(), kwargs['output']))
# 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, theta = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), dt=0.03,
sigma=float(stddev) * np.ones(nb_actions),
theta=float(theta) * 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 = MyCritic(layer_norm=layer_norm, architecture=architecture)
actor = MyActor(nb_actions, layer_norm=layer_norm, architecture=architecture)
# 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)
# Disable logging for rank != 0 to avoid noise.
if rank == 0:
start_time = time.time()
training.train(env=env, param_noise=param_noise, action_noise=action_noise,
actor=actor, critic=critic, memory=memory, **kwargs)
env.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
def main():
steering_angles = np.array([-0.7, -0.5, -0.25, 0.0, 0.25, 0.5, 0.7])
env = AirSimGym(continuous=False, off_road_dist=2.9, max_speed=4.5, scale_reward=True, steering_angles=steering_angles)
env = DummyVecEnv([lambda: env])
# the noise objects for DDPG
n_actions = env.action_space.shape[-1]
param_noise = None
action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=float(0.5) * np.ones(n_actions))
model = DDPG(MlpPolicy, env, verbose=1, param_noise=param_noise, action_noise=action_noise)
model.learn(total_timesteps=40000)
del model # remove to demonstrate saving and loading
obs = env.reset()
while True:
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
def learn(
network,
env,
data_path='',
model_path='./model/',
model_name='ddpg_none_fuzzy_150',
file_name='test',
model_based=False,
memory_extend=False,
model_type='linear',
restore=False,
dyna_learning=False,
seed=None,
nb_epochs=5, # with default settings, perform 1M steps total
nb_sample_cycle=5,
nb_epoch_cycles=150,
nb_rollout_steps=400,
nb_model_learning=10,
nb_sample_steps=50,
nb_samples_extend=5,
reward_scale=1.0,
noise_type='normal_0.2', #'adaptive-param_0.2', ou_0.2, normal_0.2
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
batch_size=32, # per MPI worker
tau=0.01,
param_noise_adaption_interval=50,
**network_kwargs):
nb_actions = env.action_space.shape[0]
memory = Memory(limit=int(1e5),
action_shape=env.action_space.shape[0],
observation_shape=env.observation_space.shape)
if model_based:
""" store fake_data"""
fake_memory = Memory(limit=int(1e5),
action_shape=env.action_space.shape[0],
observation_shape=env.observation_space.shape)
""" select model or not """
if model_type == 'gp':
kernel = ConstantKernel(1.0, (1e-3, 1e3)) * RBF(10, (1e-2, 1e2))
dynamic_model = GaussianProcessRegressor(kernel=kernel)
reward_model = GaussianProcessRegressor(kernel=kernel)
elif model_type == 'linear':
dynamic_model = LinearRegression()
reward_model = LinearRegression()
elif model_type == 'mlp':
dynamic_model = MLPRegressor(hidden_layer_sizes=(100, ),
activation='relu',
solver='adam',
alpha=0.0001,
batch_size='auto',
learning_rate='constant',
learning_rate_init=0.001,
power_t=0.5,
max_iter=200,
shuffle=True,
random_state=None,
tol=0.0001,
verbose=False,
warm_start=False,
momentum=0.9,
nesterovs_momentum=True,
early_stopping=False,
validation_fraction=0.1,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08)
reward_model = MLPRegressor(hidden_layer_sizes=(100, ),
activation='relu',
solver='adam',
alpha=0.0001,
batch_size='auto',
learning_rate='constant',
learning_rate_init=0.001,
power_t=0.5,
max_iter=200,
shuffle=True,
random_state=None,
tol=0.0001,
verbose=False,
warm_start=False,
momentum=0.9,
nesterovs_momentum=True,
early_stopping=False,
validation_fraction=0.1,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08)
else:
logger.info(
"You need to give the model_type to fit the dynamic and reward!!!"
)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
""" set noise """
action_noise = None
param_noise = None
if noise_type is not None:
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))
"""action scale"""
max_action = env.action_high_bound
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
""" agent ddpg """
agent = DDPG(actor,
critic,
memory,
env.observation_space.shape,
env.action_space.shape[0],
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
sess = U.get_session()
if restore:
agent.restore(sess, model_path, model_name)
else:
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
episodes = 0
epochs_rewards = np.zeros((nb_epochs, nb_epoch_cycles), dtype=np.float32)
epochs_times = np.zeros((nb_epochs, nb_epoch_cycles), dtype=np.float32)
epochs_steps = np.zeros((nb_epochs, nb_epoch_cycles), dtype=np.float32)
epochs_states = []
for epoch in range(nb_epochs):
logger.info(
"======================== The {} epoch start !!! ========================="
.format(epoch))
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_episode_times = []
epoch_actions = []
epoch_episode_states = []
epoch_qs = []
epoch_episodes = 0
for cycle in range(nb_epoch_cycles):
start_time = time.time()
obs, state, done = env.reset()
obs_reset = cp.deepcopy(obs)
episode_reward = 0.
episode_step = 0
episode_states = []
logger.info(
"================== The {} episode start !!! ==================="
.format(cycle))
for t_rollout in range(nb_rollout_steps):
logger.info(
"================== The {} steps finish !!! ==================="
.format(t_rollout))
""" Predict next action """
action, q, _, _ = agent.step(obs,
stddev,
apply_noise=True,
compute_Q=True)
new_obs, next_state, r, done, safe_or_not, final_action = env.step(
max_action * action, t_rollout)
if safe_or_not is False:
break
episode_reward += r
episode_step += 1
episode_states.append([
cp.deepcopy(state),
cp.deepcopy(final_action),
np.array(cp.deepcopy(r)),
cp.deepcopy(next_state)
])
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(obs, action, r, new_obs, done)
obs = new_obs
state = next_state
if done:
break
""" extend the memory """
if model_based and cycle > (nb_model_learning +
1) and memory_extend:
pred_x = np.zeros((1, 18), dtype=np.float32)
for j in range(nb_samples_extend):
m_action, _, _, _ = agent.step(obs,
stddev,
apply_noise=True,
compute_Q=False)
pred_x[:, :12] = obs
pred_x[:, 12:] = m_action
m_new_obs = dynamic_model.predict(pred_x)[0]
""" get real reward """
# state = env.inverse_state(m_new_obs)
# m_reward = env.get_reward(state, m_action)
m_reward = reward_model.predict(pred_x)[0]
agent.store_transition(obs, m_action, m_reward,
m_new_obs, done)
""" generate new data and fit model"""
if model_based and cycle > nb_model_learning:
logger.info(
"============================== Model Fit !!! ==============================="
)
input_x = np.concatenate(
(memory.observations0.data[:memory.nb_entries],
memory.actions.data[:memory.nb_entries]),
axis=1)
input_y_obs = memory.observations1.data[:memory.nb_entries]
input_y_reward = memory.rewards.data[:memory.nb_entries]
dynamic_model.fit(input_x, input_y_obs)
reward_model.fit(input_x, input_y_reward)
if dyna_learning:
logger.info(
"========================= Collect data !!! ================================="
)
pred_obs = np.zeros((1, 18), dtype=np.float32)
for sample_index in range(nb_sample_cycle):
fake_obs = obs_reset
for t_episode in range(nb_sample_steps):
fake_action, _, _, _ = agent.step(fake_obs,
stddev,
apply_noise=True,
compute_Q=False)
pred_obs[:, :12] = fake_obs
pred_obs[:, 12:] = fake_action
next_fake_obs = dynamic_model.predict(pred_obs)[0]
fake_reward = reward_model.predict(pred_obs)[0]
# next_fake_obs = dynamic_model.predict(np.concatenate((fake_obs, fake_action)))[0]
# fake_reward = reward_model.predict(np.concatenate((fake_obs, fake_action)))[0]
fake_obs = next_fake_obs
fake_terminals = False
fake_memory.append(fake_obs, fake_action,
fake_reward, next_fake_obs,
fake_terminals)
""" noise decay """
stddev = float(stddev) * 0.95
duration = time.time() - start_time
epoch_episode_rewards.append(episode_reward)
epoch_episode_steps.append(episode_step)
epoch_episode_times.append(cp.deepcopy(duration))
epoch_episode_states.append(cp.deepcopy(episode_states))
epochs_rewards[epoch, cycle] = episode_reward
epochs_steps[epoch, cycle] = episode_step
epochs_times[epoch, cycle] = cp.deepcopy(duration)
logger.info(
"============================= The Episode_Times:: {}!!! ============================"
.format(epoch_episode_rewards))
logger.info(
"============================= The Episode_Times:: {}!!! ============================"
.format(epoch_episode_times))
epoch_episodes += 1
episodes += 1
""" Training process """
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
logger.info("")
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
""" planning training """
if model_based and cycle > (nb_model_learning +
1) and dyna_learning:
for t_train in range(nb_train_steps):
# setting for adapt param noise, if necessary.
if fake_memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
batch = fake_memory.sample(batch_size=batch_size)
fake_cl, fake_al = agent.train_fake_data(batch)
epoch_critic_losses.append(fake_cl)
epoch_actor_losses.append(fake_al)
agent.update_target_net()
epochs_states.append(cp.deepcopy(epoch_episode_states))
# # save data
np.save(
data_path + 'train_reward_' + algorithm_name + '_' + noise_type +
file_name, epochs_rewards)
np.save(
data_path + 'train_step_' + algorithm_name + '_' + noise_type +
file_name, epochs_steps)
np.save(
data_path + 'train_states_' + algorithm_name + '_' + noise_type +
file_name, epochs_states)
np.save(
data_path + 'train_times_' + algorithm_name + '_' + noise_type +
file_name, epochs_times)
# # agent save
agent.store(model_path + 'train_model_' + algorithm_name + '_' +
noise_type + file_name)
def learn(
network,
env,
seed=None,
total_timesteps=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
batch_size=64, # per MPI worker
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50,
**network_kwargs):
set_global_seeds(seed)
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles *
nb_rollout_steps)
else:
nb_epochs = 500
rank = MPI.COMM_WORLD.Get_rank()
nb_actions = env.action_space.shape[-1]
assert (np.abs(env.action_space.low) == env.action_space.high
).all() # we assume symmetric actions.
memory = Memory(limit=int(1e6),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
action_noise = None
param_noise = None
if noise_type is not None:
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))
max_action = env.action_space.high
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor,
critic,
memory,
env.observation_space.shape,
env.action_space.shape,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
sess = U.get_session()
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nenvs = obs.shape[0]
episode_reward = np.zeros(nenvs, dtype=np.float32) #vector
episode_step = np.zeros(nenvs, dtype=int) # vector
episodes = 0 #scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
if nenvs > 1:
# if simulating multiple envs in parallel, impossible to reset agent at the end of the episode in each
# of the environments, so resetting here instead
agent.reset()
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q, _, _ = agent.step(obs,
apply_noise=True,
compute_Q=True)
# Execute next action.
if rank == 0 and render:
env.render()
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
new_obs, r, done, info = env.step(
max_action * action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
# note these outputs are batched from vecenv
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(
obs, action, r, new_obs, done
) #the batched data will be unrolled in memory.py's append.
obs = new_obs
for d in range(len(done)):
if done[d]:
# Episode done.
epoch_episode_rewards.append(episode_reward[d])
episode_rewards_history.append(episode_reward[d])
epoch_episode_steps.append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if nenvs == 1:
agent.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
nenvs_eval = eval_obs.shape[0]
eval_episode_reward = np.zeros(nenvs_eval, dtype=np.float32)
for t_rollout in range(nb_eval_steps):
eval_action, eval_q, _, _ = agent.step(eval_obs,
apply_noise=False,
compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(
max_action * eval_action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
for d in range(len(eval_done)):
if eval_done[d]:
eval_episode_rewards.append(eval_episode_reward[d])
eval_episode_rewards_history.append(
eval_episode_reward[d])
eval_episode_reward[d] = 0.0
mpi_size = MPI.COMM_WORLD.Get_size()
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(
episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = np.mean(
epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(
eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s' % x)
combined_stats_sums = MPI.COMM_WORLD.allreduce(
np.array(
[np.array(x).flatten()[0] for x in combined_stats.values()]))
combined_stats = {
k: v / mpi_size
for (k, v) in zip(combined_stats.keys(), combined_stats_sums)
}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'),
'wb') as f:
pickle.dump(eval_env.get_state(), f)
return agent
def __init__(self, network, env, gamma=1, tau=0.01, total_timesteps=1e6,
normalize_observations=True, normalize_returns=False, enable_popart=False,
noise_type='adaptive-param_0.2', clip_norm=None, reward_scale=1.,
batch_size=128, l2_reg_coef=0.2, actor_lr=1e-4, critic_lr=1e-3,
observation_range=(-5., 5.), action_range=(-1., 1.), return_range=(-np.inf, np.inf),
**network_kwargs):
# logger.info('Using agent with the following configuration:')
# logger.info(str(self.__dict__.items()))
observation_shape = env.observation_space.shape
action_shape = env.action_space.shape
# Inputs.
self.obs0 = tf.placeholder(tf.float32, shape=(None,) + observation_shape, name='obs0')
self.obs1 = tf.placeholder(tf.float32, shape=(None,) + observation_shape, name='obs1')
self.terminals1 = tf.placeholder(tf.float32, shape=(None, 1), name='terminals1')
self.rewards = tf.placeholder(tf.float32, shape=(None, 1), name='rewards')
self.actions = tf.placeholder(tf.float32, shape=(None,) + action_shape, name='actions')
self.critic_target = tf.placeholder(tf.float32, shape=(None, 1), name='critic_target')
self.param_noise_stddev = tf.placeholder(tf.float32, shape=(), name='param_noise_stddev')
# Parameters.
self.env = env
self.gamma = gamma
self.tau = tau
self.total_timesteps = total_timesteps
self.normalize_observations = normalize_observations
self.normalize_returns = normalize_returns
self.enable_popart = enable_popart
self.clip_norm = clip_norm
self.reward_scale = reward_scale
self.action_range = action_range
self.return_range = return_range
self.observation_range = observation_range
self.batch_size = batch_size
self.actor_lr = actor_lr
self.critic_lr = critic_lr
self.l2_reg_coef = l2_reg_coef
self.stats_sample = None
self.action_noise = None
self.param_noise = None
nb_actions = self.env.action_space.shape[-1]
if noise_type is not None:
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('_')
self.param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
self.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('_')
self.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))
assert (np.abs(env.action_space.low) == env.action_space.high).all() # we assume symmetric actions.
self.memory = Memory(limit=int(1e6), action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
self.critic = Critic(network=network, **network_kwargs)
self.actor = Actor(nb_actions, network=network, **network_kwargs)
# Observation normalization.
if self.normalize_observations:
with tf.variable_scope('obs_rms'):
self.obs_rms = RunningMeanStd(shape=observation_shape)
else:
self.obs_rms = None
normalized_obs0 = tf.clip_by_value(normalize(self.obs0, self.obs_rms),
self.observation_range[0], self.observation_range[1])
normalized_obs1 = tf.clip_by_value(normalize(self.obs1, self.obs_rms),
self.observation_range[0], self.observation_range[1])
# Return normalization.
if self.normalize_returns:
with tf.variable_scope('ret_rms'):
self.ret_rms = RunningMeanStd()
else:
self.ret_rms = None
# Create target networks.
target_actor = copy(self.actor)
target_actor.name = 'target_actor'
self.target_actor = target_actor
target_critic = copy(self.critic)
target_critic.name = 'target_critic'
self.target_critic = target_critic
# Create networks and core TF parts that are shared across setup parts.
self.actor_tf = self.actor(normalized_obs0)
self.normalized_critic_tf = self.critic(normalized_obs0, self.actions)
self.critic_tf = denormalize(
tf.clip_by_value(self.normalized_critic_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
self.normalized_critic_with_actor_tf = self.critic(normalized_obs0, self.actor_tf, reuse=True)
self.critic_with_actor_tf = denormalize(
tf.clip_by_value(self.normalized_critic_with_actor_tf, self.return_range[0], self.return_range[1]),
self.ret_rms)
Q_obs1 = denormalize(target_critic(normalized_obs1, target_actor(normalized_obs1)), self.ret_rms)
self.target_Q = self.rewards + (1. - self.terminals1) * gamma * Q_obs1
# Set up parts.
if self.param_noise is not None:
self.setup_param_noise(normalized_obs0)
self.setup_actor_optimizer()
self.setup_critic_optimizer()
if self.normalize_returns and self.enable_popart:
self.setup_popart()
self.setup_stats()
self.setup_target_network_updates()
self.initial_state = None # recurrent architectures not supported yet
self.def_path_pre = os.path.dirname(os.path.abspath(__file__)) + '/tmp/'
class Model(object):
def __init__(self, network, env, gamma=1, tau=0.01, total_timesteps=1e6,
normalize_observations=True, normalize_returns=False, enable_popart=False,
noise_type='adaptive-param_0.2', clip_norm=None, reward_scale=1.,
batch_size=128, l2_reg_coef=0.2, actor_lr=1e-4, critic_lr=1e-3,
observation_range=(-5., 5.), action_range=(-1., 1.), return_range=(-np.inf, np.inf),
**network_kwargs):
# logger.info('Using agent with the following configuration:')
# logger.info(str(self.__dict__.items()))
observation_shape = env.observation_space.shape
action_shape = env.action_space.shape
# Inputs.
self.obs0 = tf.placeholder(tf.float32, shape=(None,) + observation_shape, name='obs0')
self.obs1 = tf.placeholder(tf.float32, shape=(None,) + observation_shape, name='obs1')
self.terminals1 = tf.placeholder(tf.float32, shape=(None, 1), name='terminals1')
self.rewards = tf.placeholder(tf.float32, shape=(None, 1), name='rewards')
self.actions = tf.placeholder(tf.float32, shape=(None,) + action_shape, name='actions')
self.critic_target = tf.placeholder(tf.float32, shape=(None, 1), name='critic_target')
self.param_noise_stddev = tf.placeholder(tf.float32, shape=(), name='param_noise_stddev')
# Parameters.
self.env = env
self.gamma = gamma
self.tau = tau
self.total_timesteps = total_timesteps
self.normalize_observations = normalize_observations
self.normalize_returns = normalize_returns
self.enable_popart = enable_popart
self.clip_norm = clip_norm
self.reward_scale = reward_scale
self.action_range = action_range
self.return_range = return_range
self.observation_range = observation_range
self.batch_size = batch_size
self.actor_lr = actor_lr
self.critic_lr = critic_lr
self.l2_reg_coef = l2_reg_coef
self.stats_sample = None
self.action_noise = None
self.param_noise = None
nb_actions = self.env.action_space.shape[-1]
if noise_type is not None:
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('_')
self.param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
self.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('_')
self.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))
assert (np.abs(env.action_space.low) == env.action_space.high).all() # we assume symmetric actions.
self.memory = Memory(limit=int(1e6), action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
self.critic = Critic(network=network, **network_kwargs)
self.actor = Actor(nb_actions, network=network, **network_kwargs)
# Observation normalization.
if self.normalize_observations:
with tf.variable_scope('obs_rms'):
self.obs_rms = RunningMeanStd(shape=observation_shape)
else:
self.obs_rms = None
normalized_obs0 = tf.clip_by_value(normalize(self.obs0, self.obs_rms),
self.observation_range[0], self.observation_range[1])
normalized_obs1 = tf.clip_by_value(normalize(self.obs1, self.obs_rms),
self.observation_range[0], self.observation_range[1])
# Return normalization.
if self.normalize_returns:
with tf.variable_scope('ret_rms'):
self.ret_rms = RunningMeanStd()
else:
self.ret_rms = None
# Create target networks.
target_actor = copy(self.actor)
target_actor.name = 'target_actor'
self.target_actor = target_actor
target_critic = copy(self.critic)
target_critic.name = 'target_critic'
self.target_critic = target_critic
# Create networks and core TF parts that are shared across setup parts.
self.actor_tf = self.actor(normalized_obs0)
self.normalized_critic_tf = self.critic(normalized_obs0, self.actions)
self.critic_tf = denormalize(
tf.clip_by_value(self.normalized_critic_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
self.normalized_critic_with_actor_tf = self.critic(normalized_obs0, self.actor_tf, reuse=True)
self.critic_with_actor_tf = denormalize(
tf.clip_by_value(self.normalized_critic_with_actor_tf, self.return_range[0], self.return_range[1]),
self.ret_rms)
Q_obs1 = denormalize(target_critic(normalized_obs1, target_actor(normalized_obs1)), self.ret_rms)
self.target_Q = self.rewards + (1. - self.terminals1) * gamma * Q_obs1
# Set up parts.
if self.param_noise is not None:
self.setup_param_noise(normalized_obs0)
self.setup_actor_optimizer()
self.setup_critic_optimizer()
if self.normalize_returns and self.enable_popart:
self.setup_popart()
self.setup_stats()
self.setup_target_network_updates()
self.initial_state = None # recurrent architectures not supported yet
self.def_path_pre = os.path.dirname(os.path.abspath(__file__)) + '/tmp/'
def setup_target_network_updates(self):
actor_init_updates, actor_soft_updates = get_target_updates(self.actor.vars, self.target_actor.vars, self.tau)
critic_init_updates, critic_soft_updates = get_target_updates(self.critic.vars, self.target_critic.vars,
self.tau)
self.target_init_updates = [actor_init_updates, critic_init_updates]
self.target_soft_updates = [actor_soft_updates, critic_soft_updates]
def setup_param_noise(self, normalized_obs0):
assert self.param_noise is not None
# Configure perturbed actor.
param_noise_actor = copy(self.actor)
param_noise_actor.name = 'param_noise_actor'
self.perturbed_actor_tf = param_noise_actor(normalized_obs0)
logger.info('setting up param noise')
self.perturb_policy_ops = get_perturbed_actor_updates(self.actor, param_noise_actor, self.param_noise_stddev)
# Configure separate copy for stddev adoption.
adaptive_param_noise_actor = copy(self.actor)
adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
adaptive_actor_tf = adaptive_param_noise_actor(normalized_obs0)
self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(self.actor, adaptive_param_noise_actor,
self.param_noise_stddev)
self.adaptive_policy_distance = tf.sqrt(tf.reduce_mean(tf.square(self.actor_tf - adaptive_actor_tf)))
def setup_actor_optimizer(self):
logger.info('setting up actor optimizer')
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
actor_shapes = [var.get_shape().as_list() for var in self.actor.trainable_vars]
actor_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = U.flatgrad(self.actor_loss, self.actor.trainable_vars, clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def setup_critic_optimizer(self):
logger.info('setting up critic optimizer')
normalized_critic_target_tf = tf.clip_by_value(normalize(self.critic_target, self.ret_rms),
self.return_range[0], self.return_range[1])
self.critic_loss = tf.reduce_mean(tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
if self.l2_reg_coef > 0.:
critic_reg_vars = [var for var in self.critic.trainable_vars if
var.name.endswith('/w:0') and 'output' not in var.name]
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(self.l2_reg_coef))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.l2_reg_coef),
weights_list=critic_reg_vars
)
self.critic_loss += critic_reg
critic_shapes = [var.get_shape().as_list() for var in self.critic.trainable_vars]
critic_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = U.flatgrad(self.critic_loss, self.critic.trainable_vars, clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
beta1=0.9, beta2=0.999, epsilon=1e-08)
def setup_popart(self):
# See https://arxiv.org/pdf/1602.07714.pdf for details.
self.old_std = tf.placeholder(tf.float32, shape=[1], name='old_std')
new_std = self.ret_rms.std
self.old_mean = tf.placeholder(tf.float32, shape=[1], name='old_mean')
new_mean = self.ret_rms.mean
self.renormalize_Q_outputs_op = []
for vs in [self.critic.output_vars, self.target_critic.output_vars]:
assert len(vs) == 2
M, b = vs
assert 'kernel' in M.name
assert 'bias' in b.name
assert M.get_shape()[-1] == 1
assert b.get_shape()[-1] == 1
self.renormalize_Q_outputs_op += [M.assign(M * self.old_std / new_std)]
self.renormalize_Q_outputs_op += [b.assign((b * self.old_std + self.old_mean - new_mean) / new_std)]
def setup_stats(self):
ops = []
names = []
if self.normalize_returns:
ops += [self.ret_rms.mean, self.ret_rms.std]
names += ['ret_rms_mean', 'ret_rms_std']
if self.normalize_observations:
ops += [tf.reduce_mean(self.obs_rms.mean), tf.reduce_mean(self.obs_rms.std)]
names += ['obs_rms_mean', 'obs_rms_std']
ops += [tf.reduce_mean(self.critic_tf)]
names += ['reference_Q_mean']
ops += [reduce_std(self.critic_tf)]
names += ['reference_Q_std']
ops += [tf.reduce_mean(self.critic_with_actor_tf)]
names += ['reference_actor_Q_mean']
ops += [reduce_std(self.critic_with_actor_tf)]
names += ['reference_actor_Q_std']
ops += [tf.reduce_mean(self.actor_tf)]
names += ['reference_action_mean']
ops += [reduce_std(self.actor_tf)]
names += ['reference_action_std']
if self.param_noise:
ops += [tf.reduce_mean(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_mean']
ops += [reduce_std(self.perturbed_actor_tf)]
names += ['reference_perturbed_action_std']
self.stats_ops = ops
self.stats_names = names
def train_step(self, obs, apply_noise=True, compute_Q=True):
if self.param_noise is not None and apply_noise:
actor_tf = self.perturbed_actor_tf
else:
actor_tf = self.actor_tf
feed_dict = {self.obs0: U.adjust_shape(self.obs0, [obs])}
if compute_Q:
action, q = self.sess.run([actor_tf, self.critic_with_actor_tf], feed_dict=feed_dict)
else:
action = self.sess.run(actor_tf, feed_dict=feed_dict)
q = None
if self.action_noise is not None and apply_noise:
noise = self.action_noise()
assert noise.shape == action[0].shape
action += noise
action = np.clip(action, self.action_range[0], self.action_range[1])
return action, q, None, None
def step(self, obs, compute_Q=True):
feed_dict = {self.obs0: U.adjust_shape(self.obs0, [obs])}
if compute_Q:
action, q = self.sess.run([self.actor_tf, self.critic_with_actor_tf], feed_dict=feed_dict)
else:
action = self.sess.run(self.actor_tf, feed_dict=feed_dict)
q = None
action = np.clip(action, self.action_range[0], self.action_range[1])
return action, q, None, None
def store_transition(self, obs0, action, reward, obs1, terminal1):
reward *= self.reward_scale
B = obs0.shape[0]
for b in range(B):
self.memory.append(obs0[b], action[b], reward[b], obs1[b], terminal1[b])
if self.normalize_observations:
self.obs_rms.update(np.array([obs0[b]]))
def train(self):
# Get a batch.
batch = self.memory.sample(batch_size=self.batch_size)
if self.normalize_returns and self.enable_popart:
old_mean, old_std, target_Q = self.sess.run([self.ret_rms.mean, self.ret_rms.std, self.target_Q],
feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
self.ret_rms.update(target_Q.flatten())
self.sess.run(self.renormalize_Q_outputs_op, feed_dict={
self.old_std: np.array([old_std]),
self.old_mean: np.array([old_mean]),
})
# Run sanity check. Disabled by default since it slows down things considerably.
# print('running sanity check')
# target_Q_new, new_mean, new_std = self.sess.run([self.target_Q, self.ret_rms.mean, self.ret_rms.std], feed_dict={
# self.obs1: batch['obs1'],
# self.rewards: batch['rewards'],
# self.terminals1: batch['terminals1'].astype('float32'),
# })
# print(target_Q_new, target_Q, new_mean, new_std)
# assert (np.abs(target_Q - target_Q_new) < 1e-3).all()
else:
target_Q = self.sess.run(self.target_Q, feed_dict={
self.obs1: batch['obs1'],
self.rewards: batch['rewards'],
self.terminals1: batch['terminals1'].astype('float32'),
})
# Get all gradients and perform a synced update.
ops = [self.actor_grads, self.actor_loss, self.critic_grads, self.critic_loss]
actor_grads, actor_loss, critic_grads, critic_loss = self.sess.run(ops, feed_dict={
self.obs0: batch['obs0'],
self.actions: batch['actions'],
self.critic_target: target_Q,
})
self.actor_optimizer.update(actor_grads, stepsize=self.actor_lr)
self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
return critic_loss, actor_loss
def initialize(self, sess):
self.sess = sess
self.sess.run(tf.global_variables_initializer())
self.actor_optimizer.sync()
self.critic_optimizer.sync()
self.sess.run(self.target_init_updates)
def update_target_net(self):
self.sess.run(self.target_soft_updates)
def get_stats(self):
if self.stats_sample is None:
# Get a sample and keep that fixed for all further computations.
# This allows us to estimate the change in value for the same set of inputs.
self.stats_sample = self.memory.sample(batch_size=self.batch_size)
values = self.sess.run(self.stats_ops, feed_dict={
self.obs0: self.stats_sample['obs0'],
self.actions: self.stats_sample['actions'],
})
names = self.stats_names[:]
assert len(names) == len(values)
stats = dict(zip(names, values))
if self.param_noise is not None:
stats = {**stats, **self.param_noise.get_stats()}
return stats
def adapt_param_noise(self):
try:
from mpi4py import MPI
except ImportError:
MPI = None
if self.param_noise is None:
return 0.
# Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
batch = self.memory.sample(batch_size=self.batch_size)
self.sess.run(self.perturb_adaptive_policy_ops, feed_dict={
self.param_noise_stddev: self.param_noise.current_stddev,
})
distance = self.sess.run(self.adaptive_policy_distance, feed_dict={
self.obs0: batch['obs0'],
self.param_noise_stddev: self.param_noise.current_stddev,
})
if MPI is not None:
mean_distance = MPI.COMM_WORLD.allreduce(distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
else:
mean_distance = distance
self.param_noise.adapt(mean_distance)
return mean_distance
def reset(self):
# Reset internal state after an episode is complete.
if self.action_noise is not None:
self.action_noise.reset()
if self.param_noise is not None:
self.sess.run(self.perturb_policy_ops, feed_dict={
self.param_noise_stddev: self.param_noise.current_stddev,
})
def learn(self,
total_timesteps=None,
seed=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
render=False,
nb_train_steps=50, # per epoch cycle and MPI worker,
batch_size=64, # per MPI worker
param_noise_adaption_interval=50,):
set_global_seeds(seed)
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles * nb_rollout_steps)
else:
nb_epochs = 500
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
# eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
sess = U.get_session()
# Prepare everything.
self.initialize(sess)
sess.graph.finalize()
self.reset()
obs = self.env.reset()
# if eval_env is not None:
# eval_obs = eval_env.reset()
nenvs = obs.shape[0]
episode_reward = np.zeros(nenvs, dtype=np.float32) # vector
episode_step = np.zeros(nenvs, dtype=int) # vector
episodes = 0 # scalar
t = 0 # scalar
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
if nenvs > 1:
# if simulating multiple envs in parallel, impossible to reset agent at the end of the episode in each
# of the environments, so resetting here instead
self.reset()
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q, _, _ = self.train_step(obs, apply_noise=True, compute_Q=True)
# Execute next action.
if rank == 0 and render:
self.env.render()
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
# new_obs, r, done, info = self.env.step(max_action * action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
new_obs, r, done, info = self.env.step(action)
# note these outputs are batched from vecenv
t += 1
if rank == 0 and render:
self.env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
# the batched data will be unrolled in memory.py's append.
self.store_transition(obs, action, r, new_obs, done)
obs = new_obs
for d in range(len(done)):
if done[d]:
# Episode done.
epoch_episode_rewards.append(episode_reward[d])
episode_rewards_history.append(episode_reward[d])
epoch_episode_steps.append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if nenvs == 1:
self.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if self.memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = self.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = self.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
self.update_target_net()
#
# # Evaluate.
# eval_episode_rewards = []
# eval_qs = []
# if eval_env is not None:
# eval_obs = eval_env.reset()
# nenvs_eval = eval_obs.shape[0]
# eval_episode_reward = np.zeros(nenvs_eval, dtype=np.float32)
# for t_rollout in range(nb_eval_steps):
# eval_action, eval_q, _, _ = self.train_step(eval_obs, apply_noise=False, compute_Q=True)
# # eval_obs, eval_r, eval_done, eval_info = eval_env.step(
# # max_action * eval_action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
# eval_obs, eval_r, eval_done, eval_info = eval_env.step(eval_action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
#
# if render_eval:
# eval_env.render()
# eval_episode_reward += eval_r
#
# eval_qs.append(eval_q)
# for d in range(len(eval_done)):
# if eval_done[d]:
# eval_episode_rewards.append(eval_episode_reward[d])
# eval_episode_rewards_history.append(eval_episode_reward[d])
# eval_episode_reward[d] = 0.0
if MPI is not None:
mpi_size = MPI.COMM_WORLD.Get_size()
else:
mpi_size = 1
# save trainable variables
file_name = time.strftime('Y%YM%mD%d_h%Hm%Ms%S', time.localtime(time.time()))
model_save_path = self.def_path_pre + file_name
self.save(model_save_path)
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = self.get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_std'] = np.std(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(episode_rewards_history)
combined_stats['rollout/return_history_std'] = np.std(episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
# combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
# combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
# combined_stats['train/param_noise_distance'] = np.mean(epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics.
# if eval_env is not None:
# combined_stats['eval/return'] = eval_episode_rewards
# combined_stats['eval/return_history'] = np.mean(eval_episode_rewards_history)
# combined_stats['eval/Q'] = eval_qs
# combined_stats['eval/episodes'] = len(eval_episode_rewards)
combined_stats_sums = np.array([np.array(x).flatten()[0] for x in combined_stats.values()])
if MPI is not None:
combined_stats_sums = MPI.COMM_WORLD.allreduce(combined_stats_sums)
combined_stats = {k: v / mpi_size for (k, v) in zip(combined_stats.keys(), combined_stats_sums)}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(self.env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(self.env.get_state(), f)
# if eval_env and hasattr(eval_env, 'get_state'):
# with open(os.path.join(logdir, 'eval_env_state.pkl'), 'wb') as f:
# pickle.dump(eval_env.get_state(), f)
self.sess.graph._unsafe_unfinalize()
return self
def save(self, save_path=None):
save_variables(save_path=save_path, sess=self.sess)
print('save model variables to', save_path)
def load_newest(self, load_path=None):
file_list = os.listdir(self.def_path_pre)
file_list.sort(key=lambda x: os.path.getmtime(os.path.join(self.def_path_pre, x)))
if load_path is None:
load_path = os.path.join(self.def_path_pre, file_list[-1])
load_variables(load_path=load_path, sess=self.sess)
print('load_path: ', load_path)
def load_index(self, index, load_path=None):
file_list = os.listdir(self.def_path_pre)
file_list.sort(key=lambda x: os.path.getmtime(os.path.join(self.def_path_pre, x)), reverse=True)
if load_path is None:
load_path = os.path.join(self.def_path_pre, file_list[index])
load_variables(load_path=load_path, sess=self.sess)
print('load_path: ', load_path)
def __init__(
self,
env,
gamma,
total_timesteps,
network='mlp',
nb_rollout_steps=100,
reward_scale=1.0,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=False,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker, <- HERE!
nb_eval_steps=100,
buffer_size=1000000,
batch_size=64, # per MPI worker
tau=0.01,
param_noise_adaption_interval=50,
**network_kwargs):
# Adjusting hyper-parameters by considering the number of options policies to learn
num_options = env.get_number_of_options()
buffer_size = num_options * buffer_size
batch_size = num_options * batch_size
observation_space = env.option_observation_space
action_space = env.option_action_space
nb_actions = action_space.shape[-1]
assert (np.abs(action_space.low) == action_space.high
).all() # we assume symmetric actions.
memory = Memory(limit=buffer_size,
action_shape=action_space.shape,
observation_shape=observation_space.shape)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
action_noise = None
param_noise = None
if noise_type is not None:
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))
max_action = action_space.high
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor,
critic,
memory,
observation_space.shape,
action_space.shape,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
sess = U.get_session()
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
# Variables that are used during learning
self.agent = agent
self.memory = memory
self.max_action = max_action
self.batch_size = batch_size
self.nb_train_steps = nb_train_steps
self.nb_rollout_steps = nb_rollout_steps
self.param_noise_adaption_interval = param_noise_adaption_interval
def train(self,
env_fn,
num_timesteps,
noise_type,
layer_norm,
folder,
load_policy,
video_width,
video_height,
plot_rewards,
save_every=50,
seed=1234,
episode_length=1000,
pi_hid_size=150,
pi_num_hid_layers=3,
render_frames=_render_frames,
**kwargs):
num_cpu = self.workers
if sys.platform == 'darwin':
num_cpu //= 2
config = tf.ConfigProto(
allow_soft_placement=True,
intra_op_parallelism_threads=num_cpu,
inter_op_parallelism_threads=num_cpu)
if self.gpu_usage is None or self.gpu_usage <= 0.:
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
else:
config.gpu_options.allow_growth = True # pylint: disable=E1101
config.gpu_options.per_process_gpu_memory_fraction = self.gpu_usage / self.workers
tf.Session(config=config).__enter__()
worker_seed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(worker_seed)
tf.set_random_seed(worker_seed)
np.random.seed(worker_seed)
save_every = max(1, save_every)
env = env_fn()
env.seed(worker_seed)
rank = MPI.COMM_WORLD.Get_rank()
logger.info('rank {}: seed={}, logdir={}'.format(rank, worker_seed,
logger.get_dir()))
def policy_fn(name, ob_space, ac_space):
return mlp_policy.MlpPolicy(
name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=pi_hid_size,
num_hid_layers=pi_num_hid_layers)
env = bench.Monitor(
env,
logger.get_dir() and osp.join(logger.get_dir(), str(rank)),
allow_early_resets=True)
gym.logger.setLevel(logging.INFO)
that = self
iter_name = 'iters_so_far'
if self.method == 'sql':
iter_name = 'epoch'
# TODO replace with utils.create_callback(...)
def callback(locals, globals):
if that.method != "ddpg":
if load_policy is not None and locals[iter_name] == 0:
# noinspection PyBroadException
try:
utils.load_state(load_policy)
if MPI.COMM_WORLD.Get_rank() == 0:
logger.info("Loaded policy network weights from %s." % load_policy)
# save TensorFlow summary (contains at least the graph definition)
except:
logger.error("Failed to load policy network weights from %s." % load_policy)
if MPI.COMM_WORLD.Get_rank() == 0 and locals[iter_name] == 0:
_ = tf.summary.FileWriter(folder, tf.get_default_graph())
if MPI.COMM_WORLD.Get_rank() == 0 and locals[iter_name] % save_every == 0:
print('Saving video and checkpoint for policy at iteration %i...' %
locals[iter_name])
ob = env.reset()
images = []
rewards = []
max_reward = 1. # if any reward > 1, we have to rescale
lower_part = video_height // 5
for i in range(episode_length):
if that.method == "ddpg":
ac, _ = locals['agent'].pi(ob, apply_noise=False, compute_Q=False)
elif that.method == "sql":
ac, _ = locals['policy'].get_action(ob)
elif isinstance(locals['pi'], GaussianMlpPolicy):
ac, _, _ = locals['pi'].act(np.concatenate((ob, ob)))
else:
ac, _ = locals['pi'].act(False, ob)
ob, rew, new, _ = env.step(ac)
images.append(render_frames(env))
if plot_rewards:
rewards.append(rew)
max_reward = max(rew, max_reward)
if new:
break
orange = np.array([255, 163, 0])
red = np.array([255, 0, 0])
video = []
width_factor = 1. / episode_length * video_width
for i, imgs in enumerate(images):
for img in imgs:
img[-lower_part, :10] = orange
img[-lower_part, -10:] = orange
if episode_length < video_width:
p_rew_x = 0
for j, r in enumerate(rewards[:i]):
rew_x = int(j * width_factor)
if r < 0:
img[-1:, p_rew_x:rew_x] = red
img[-1:, p_rew_x:rew_x] = red
else:
rew_y = int(r / max_reward * lower_part)
img[-rew_y - 1:, p_rew_x:rew_x] = orange
img[-rew_y - 1:, p_rew_x:rew_x] = orange
p_rew_x = rew_x
else:
for j, r in enumerate(rewards[:i]):
rew_x = int(j * width_factor)
if r < 0:
img[-1:, rew_x] = red
img[-1:, rew_x] = red
else:
rew_y = int(r / max_reward * lower_part)
img[-rew_y - 1:, rew_x] = orange
img[-rew_y - 1:, rew_x] = orange
video.append(np.hstack(imgs))
imageio.mimsave(
os.path.join(folder, "videos", "%s_%s_iteration_%i.mp4" %
(that.environment, that.method, locals[iter_name])),
video,
fps=60)
env.reset()
if that.method != "ddpg":
utils.save_state(os.path.join(that.folder, "checkpoints", "%s_%i" %
(that.environment, locals[iter_name])))
if self.method == "ppo":
pposgd_simple.learn(
env,
policy_fn,
max_timesteps=int(num_timesteps),
timesteps_per_actorbatch=1024, # 256
clip_param=0.2,
entcoeff=0.01,
optim_epochs=4,
optim_stepsize=1e-3, # 1e-3
optim_batchsize=64,
gamma=0.99,
lam=0.95,
schedule='linear', # 'linear'
callback=callback)
elif self.method == "trpo":
trpo_mpi.learn(
env,
policy_fn,
max_timesteps=int(num_timesteps),
timesteps_per_batch=1024,
max_kl=0.1, # 0.01
cg_iters=10,
cg_damping=0.1,
gamma=0.99,
lam=0.98,
vf_iters=5,
vf_stepsize=1e-3,
callback=callback)
elif self.method == "acktr":
from algos.acktr import acktr
with tf.Session(config=tf.ConfigProto()):
ob_dim = env.observation_space.shape[0]
ac_dim = env.action_space.shape[0]
with tf.variable_scope("vf"):
vf = NeuralNetValueFunction(ob_dim, ac_dim)
with tf.variable_scope("pi"):
policy = GaussianMlpPolicy(ob_dim, ac_dim)
acktr.learn(
env,
pi=policy,
vf=vf,
gamma=0.99,
lam=0.97,
timesteps_per_batch=1024,
desired_kl=0.01, # 0.002
num_timesteps=num_timesteps,
animate=False,
callback=callback)
elif self.method == "ddpg":
from algos.ddpg import ddpg
# 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('_')
from baselines.ddpg.noise import AdaptiveParamNoiseSpec
param_noise = AdaptiveParamNoiseSpec(
initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
from baselines.ddpg.noise import NormalActionNoise
action_noise = NormalActionNoise(
mu=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
from baselines.ddpg.noise import OrnsteinUhlenbeckActionNoise
_, 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)
ddpg.train(
env=env,
eval_env=None,
param_noise=param_noise,
render=False,
render_eval=False,
action_noise=action_noise,
actor=actor,
critic=critic,
memory=memory,
callback=callback,
**kwargs)
elif self.method == "sql":
from softqlearning.algorithms import SQL
from softqlearning.misc.kernel import adaptive_isotropic_gaussian_kernel
from softqlearning.misc.utils import timestamp
from softqlearning.replay_buffers import SimpleReplayBuffer
from softqlearning.value_functions import NNQFunction
from softqlearning.policies import StochasticNNPolicy
from rllab.envs.gym_env import GymEnv
env = GymEnv(env)
variant = {
'seed': [1, 2, 3],
'policy_lr': 3E-4,
'qf_lr': 3E-4,
'discount': 0.99,
'layer_size': 128,
'batch_size': 128,
'max_pool_size': 1E6,
'n_train_repeat': 1,
'epoch_length': 1000,
'snapshot_mode': 'last',
'snapshot_gap': 100,
}
pool = SimpleReplayBuffer(
env_spec=env.spec,
max_replay_buffer_size=variant['max_pool_size'],
)
base_kwargs = dict(
min_pool_size=episode_length,
epoch_length=episode_length,
n_epochs=num_timesteps,
max_path_length=episode_length,
batch_size=variant['batch_size'],
n_train_repeat=variant['n_train_repeat'],
eval_render=False,
eval_n_episodes=1,
iter_callback=callback
)
qf = NNQFunction(
env_spec=env.spec,
hidden_layer_sizes=tuple([pi_hid_size] * pi_num_hid_layers),
)
pi_layers = tuple([pi_hid_size] * pi_num_hid_layers)
policy = StochasticNNPolicy(env_spec=env.spec, hidden_layer_sizes=pi_layers)
algorithm = SQL(
base_kwargs=base_kwargs,
env=env,
pool=pool,
qf=qf,
policy=policy,
kernel_fn=adaptive_isotropic_gaussian_kernel,
kernel_n_particles=32,
kernel_update_ratio=0.5,
value_n_particles=16,
td_target_update_interval=1000,
qf_lr=variant['qf_lr'],
policy_lr=variant['policy_lr'],
discount=variant['discount'],
reward_scale=1,
save_full_state=False,
)
algorithm.train()
else:
print('ERROR: Invalid "method" argument provided.', file=sys.stderr)
env.close()
def run(self):
"""Override Process.run()"""
# Create environment
env = create_environment(
action_repeat=self.action_repeat,
full=self.full,
exclude_centering_frame=self.exclude_centering_frame,
visualize=self.visualize,
fail_reward=self.fail_reward,
integrator_accuracy=self.integrator_accuracy)
nb_actions = env.action_space.shape[-1]
# keep tracks of the number of trajectory done
num_traj = 0
env.seed(os.getpid())
set_global_seeds(os.getpid())
# Create OU Noise
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions),
sigma=0.2,
theta=0.1)
# Allocate ReplayBuffer
memory = Memory(limit=int(1e6),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
# Create DPPG agent
agent = DDPG(self.actor,
self.critic,
memory,
env.observation_space.shape,
env.action_space.shape,
gamma=self.gamma,
tau=self.tau,
normalize_returns=self.normalize_returns,
normalize_observations=self.normalize_observations,
batch_size=self.batch_size,
action_noise=action_noise,
param_noise=self.param_noise,
critic_l2_reg=self.critic_l2_reg,
enable_popart=self.popart,
clip_norm=self.clip_norm,
reward_scale=self.reward_scale)
# Build the sampling logic fn
sampling_fn = make_sampling_fn(agent, env, self.episode_length,
self.action_repeat, self.max_action,
self.nb_episodes,
self.action_noise_prob)
# Start TF session
with U.single_threaded_session() as sess:
agent.initialize(sess)
set_parameters = U.SetFromFlat(self.actor.trainable_vars)
# Start sampling-worker loop.
while True:
# self.event.wait() # Wait for a new message
# self.event.clear() # Upon message receipt, mark as read
message, actor_ws = self.inputQ.get() # Pop message
if message == 'sample':
# Set weights
set_parameters(actor_ws)
# Do sampling
transitions = sampling_fn()
self.outputQ.put((self.process_index, transitions))
# update number of trajectories
num_traj += self.nb_episodes
# restore environment if needed
if num_traj >= self.max_env_traj:
env.restore()
num_traj = 0
elif message == 'exit':
print('[Worker {}] Exiting...'.format(os.getpid()))
env.close()
break
def learn(
network,
env,
seed=None,
total_timesteps=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
batch_size=64, # per MPI worker
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50,
shared_critic=False,
save_rate=1,
save_model=True,
save_actions=True,
restore=False,
saver=None,
**network_kwargs):
if save_model or save_actions:
assert saver is not None
set_global_seeds(seed)
if shared_critic:
raise NotImplementedError()
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles *
nb_rollout_steps)
else:
nb_epochs = 500
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
nb_actions_n = [
action_space.shape[-1] for action_space in env.action_space
]
assert np.array([(np.abs(action_space.low) == action_space.high)
for action_space in env.action_space
]).all() # we assume symmetric actions.
action_noise_n = []
param_noise_n = []
if noise_type is not None:
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_n = [
AdaptiveParamNoiseSpec(initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
for _ in range(env.n)
]
action_noise_n = [None for _ in range(env.n)]
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise_n = [None for _ in range(env.n)]
action_noise_n = [
NormalActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
for nb_actions in nb_actions_n
]
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise_n = [None for _ in range(env.n)]
action_noise_n = [
OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
for nb_actions in nb_actions_n
]
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
max_action_n = [action_space.high for action_space in env.action_space]
logger.info(
'scaling actions by {} before executing in env'.format(max_action_n))
agent = MADDPG(env,
network,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise_n=action_noise_n,
param_noise_n=param_noise_n,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale,
shared_critic=shared_critic)
if saver is not None and restore:
saver.load_model()
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = [deque(maxlen=100) for _ in range(env.n)]
# sess = U.get_session()
# Prepare everything.
agent.initialize()
# sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nveh = obs.shape[0]
episode_reward = np.zeros((nveh, 1), dtype=np.float32) # vector
episode_step = np.zeros(nveh, dtype=int) # vector
episodes = 0 # scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = [[] for _ in range(env.n)]
epoch_episode_steps = [[] for _ in range(env.n)]
epoch_actions = [[] for _ in range(env.n)]
epoch_qs = [[] for _ in range(env.n)]
for epoch in range(nb_epochs):
epoch_episodes = 0
for cycle in range(nb_epoch_cycles):
from glog import info
info("epoch %d, cycle %d" % (epoch, cycle))
# Perform rollouts.
# if nveh > 1:
# # if simulating multiple envs in parallel, impossible to reset agent at the end of the episode in each
# # of the environments, so resetting here instead
agent.reset()
env.reset()
for t_rollout in range(nb_rollout_steps):
# Predict next action.
# todo no compute Q for now
action_n, q_n, _, _ = agent.step(obs,
apply_noise=True,
compute_Q=True)
if cycle == 0 and save_actions:
saver.add_action(action_n)
# Execute next action.
if cycle == 0 and rank == 0 and render:
env.render()
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
# todo max_action not scale yet
new_obs, r, done, info = env.step(action_n)
# scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
# note these outputs are batched from vecenv
# set epoch and cycle for video recorder
if isinstance(env, VecVideoRecorder):
env.set_stample(epoch, cycle)
t += 1
episode_reward += r
episode_step += 1
# Book-keeping.
for i in range(env.n):
epoch_actions[i].append(action_n[i])
epoch_qs[i].append(q_n[i])
agent.store_transition(
obs, action_n, r, new_obs, done
) # the batched data will be unrolled in memory.py's append.
obs = new_obs
terminal = (t_rollout == (nb_rollout_steps - 1))
if any(done) or terminal:
agent.reset()
env.reset()
for d in range(len(done)):
if done[d] or terminal:
# Episode done.
epoch_episode_rewards[d].append(
episode_reward[d][0])
episode_rewards_history[d].append(
episode_reward[d][0])
epoch_episode_steps[d].append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if save_actions and cycle == 0:
saver.save_actions(epoch, cycle)
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if agent.memory_nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
nenvs_eval = eval_obs.shape[0]
eval_episode_reward = np.zeros(nenvs_eval, dtype=np.float32)
for t_rollout in range(nb_eval_steps):
eval_action, eval_q, _, _ = agent.step(eval_obs,
apply_noise=False,
compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(
max_action_n * eval_action
) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
for d in range(len(eval_done)):
if eval_done[d]:
eval_episode_rewards.append(eval_episode_reward[d])
eval_episode_rewards_history.append(
eval_episode_reward[d])
eval_episode_reward[d] = 0.0
if MPI is not None:
mpi_size = MPI.COMM_WORLD.Get_size()
else:
mpi_size = 1
if save_model and save_rate is not None and epoch % save_rate == 0:
saver.save_model()
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
# todo not record agent stats here
# stats = agent.get_stats()
# combined_stats = stats.copy()
# todo simplified log
combined_stats = {}
for i in range(env.n):
combined_stats['rollout/return_%d' % i] = np.mean(
epoch_episode_rewards[i])
combined_stats['rollout/return_std_%d' % i] = np.std(
epoch_episode_rewards[i])
combined_stats['rollout/return_history_%d' % i] = np.mean(
episode_rewards_history[i])
combined_stats['rollout/return_history_std_%d' % i] = np.std(
episode_rewards_history[i])
combined_stats['rollout/episode_steps_%d' % i] = np.mean(
epoch_episode_steps[i])
combined_stats['rollout/Q_mean_%d' % i] = np.mean(epoch_qs[i])
combined_stats['rollout/actions_mean_%d' % i] = np.mean(
epoch_actions[i])
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
# Evaluation statistics.
if eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(
eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s' % x)
# todo now only log mean reward
combined_stats_sums = np.array(
[np.array(x).flatten().mean() for x in combined_stats.values()])
if MPI is not None:
combined_stats_sums = MPI.COMM_WORLD.allreduce(combined_stats_sums)
combined_stats = {
k: v / mpi_size
for (k, v) in zip(combined_stats.keys(), combined_stats_sums)
}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'),
'wb') as f:
pickle.dump(eval_env.get_state(), f)
return agent
def learn(
network,
env,
seed=None,
total_timesteps=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=True,
actor_l2_reg=0.0,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=1000,
batch_size=64, # per MPI worker
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50,
rb_size=1e6,
save_interval=1,
pretrain_epochs=0,
load_path=None,
demos_path=None,
bc_teacher_lambda=0.0,
use_qfilter=False,
**network_kwargs):
"""Learns policy using DDPG, with vectorized environments.
If we pass other arguments that aren't specified here, they are considered
as network_kwargs.
Parameters
----------
noise_type: for noise to be added to the behavior policy. They are NOT
using the noise type from the paper but 'AdaptiveParamNoiseSpec'. I
_think_ that if one does the OU process, we get action noise, but not
parameter noise. Also, be sure to use `name_stdev` in that convention,
as the code will split the argument at the underscores.
actor_lr: 1e-4 (matches paper)
critic_lr: 1e-3 (matches paper)
critic_l2: 1e-2 (matches paper)
gamma: 0.99 (matches paper)
batch_size: 64 (matches paper for lower-dim env obs/states)
tau: 0.01 for soft target updates of actor and critic nets. Paper used 0.001.
nb_epoch_cycles: number of times we go through this cycle of: (1) get
rollouts with noise added to policy and apply to replay buffer, (2)
gradient updates for actor/critic, (3) evaluation rollouts (if any).
AFTER all of these cycles happen, THEN we log statistics.
nb_rollout_steps: number of steps in each parallel env we take with
exploration policy without training, so this is just to populate the
replay buffer. More parallel envs *should* mean that we get more
samples in the buffer between each gradient updates of the network, so
this might need to be environment *and* machine (# of CPUs) specific.
nb_train_steps: after doing `nb_rollout_steps` in each parallel env, we do
this many updates; each involves sampling from the replay buffer and
updating the actor and critic (via lagged target updates).
nb_eval_steps: 1000, I changed from the 100 as default. Using 1000 ensures
that fixed length envs like Ant-v2 can get one full episode (assuming
one parallel env) during evaluation stagtes.
eval_env: A separate environment for evaluation only, where no noise is
applied, similar to how rlkit does it.
save_interval: Frequency between saving.
"""
set_global_seeds(seed)
# Daniel: this helps to maintain compatibility with PPO2 code. For now
# we're ignoring it, but we should check that we're always clipping. I
# changed the nb_epochs to match with PPO2 in that we divide by nenvs.
if 'limit_act_range' in network_kwargs:
network_kwargs.pop('limit_act_range')
nenvs = env.num_envs
nbatchsize = nenvs * nb_epoch_cycles * nb_rollout_steps
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // nbatchsize
else:
nb_epochs = 500
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
# we assume symmetric actions.
nb_actions = env.action_space.shape[-1]
assert (np.abs(env.action_space.low) == env.action_space.high).all()
# Form XP (1M steps, same as in paper), and critic/actor networks.
# Daniel: force dtype here so we can use uint8 type images.
assert env.observation_space.low.dtype == env.observation_space.high.dtype
memory = Memory(limit=int(rb_size),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape,
dtype=env.observation_space.low.dtype)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
action_noise = None
param_noise = None
if noise_type is not None:
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))
#if nenvs > 1:
# Daniel: adding this to replace the former.
action_noise = NormalActionNoise(mu=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions),
shape=(nenvs, 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))
max_action = env.action_space.high
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
# The `learn` defaults above have priority over defaults in DDPG class.
agent = DDPG(actor,
critic,
memory,
env.observation_space.shape,
env.action_space.shape,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
actor_l2_reg=actor_l2_reg,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale,
bc_teacher_lambda=bc_teacher_lambda,
use_qfilter=use_qfilter)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
# Prepare everything.
sess = U.get_session()
agent.initialize(sess)
# --------------------------------------------------------------------------
# Daniel: similar as PPO2 code as `agent` is similar to `model` but has to
# be initialized explicitly above. Must call after `agent.load` gets
# created. Not sure if this works with parameter space noise or with
# normalization, but I don't plan to resume training (for now). It also has
# to be *before* the `graph.finalize()` because otherwise we get an error.
# --------------------------------------------------------------------------
if load_path is not None:
logger.info("\nInside ddpg, loading model from: {}".format(load_path))
agent.load(load_path)
# --------------------------------------------------------------------------
sess.graph.finalize()
agent.reset()
# --------------------------------------------------------------------------
# Daniel: populate replay buffer, followed by pretraining stage.
# But if load_path is not None, then doesn't make sense -- we want to load.
# We also don't need to do this if timesteps is 0 (e.g., for playing policy).
# --------------------------------------------------------------------------
if total_timesteps == 0:
return agent
assert seed == 1500, 'We normally want seed 1500, yet: {}'.format(seed)
if (demos_path is not None and load_path is None):
_ddpg_demos(demos_path, agent, memory)
assert memory.nb_entries == memory.nb_teach_entries, memory.nb_entries
checkdir = osp.join(logger.get_dir(), 'checkpoints')
os.makedirs(checkdir, exist_ok=True)
# Pretrain, based on their training code for some # of minibatches.
pt_actor_losses = []
pt_critic_losses = []
batches_per_ep = int(memory.nb_entries / batch_size)
logger.info(
'Running pre-training for {} epochs'.format(pretrain_epochs))
logger.info(' data size in memory: {}'.format(memory.nb_entries))
logger.info(' each batch: {}, epoch mbs: {}'.format(
batch_size, batches_per_ep))
pt_start = time.time()
for epoch in range(1, pretrain_epochs + 1):
c_losses = []
a_losses = []
for _ in range(batches_per_ep):
cl, al = agent.train(during_pretrain=True)
agent.update_target_net()
c_losses.append(cl)
a_losses.append(al)
pt_critic_losses.append(np.mean(c_losses))
pt_actor_losses.append(np.mean(a_losses))
# Check and save model occasionally.
if epoch == 1 or epoch % 5 == 0:
pt_time = (time.time() - pt_start) / 60.
logger.info(
' epoch done: {}, loss over past epoch: {:.4f}'.format(
str(epoch).zfill(4), pt_actor_losses[-1]))
logger.info(' critic loss over past epoch: {:.4f}'.format(
pt_critic_losses[-1]))
logger.info(' elapsed time: {:.1f}m'.format(pt_time))
savepath = osp.join(
checkdir, 'pretrain_epoch_{}'.format(str(epoch).zfill(4)))
logger.info('Saving model checkpoint to: ', savepath)
agent.save(savepath)
pt_time = (time.time() - pt_start) / 60.
logger.info('losses a: {}'.format(np.array(pt_actor_losses)))
logger.info('losses c: {}'.format(np.array(pt_critic_losses)))
logger.info('Finished loading teacher samples + pre-training.')
logger.info('Pre-training took {:.1f}m.\n'.format(pt_time))
# --------------------------------------------------------------------------
# Back to their code. For cloth, `env.reset()` takes a while so we put it here.
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nenvs = obs.shape[0]
# Daniel: Debugging/sanity checks.
_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
U.display_var_info(_variables)
logger.info("\nInside DDPG, about to start epochs")
logger.info(
"nbatchsize: {}, get this in buffer before DDPG updates".format(
nbatchsize))
logger.info(" i.e.: (nenv {}) * (cycles {}) * (nsteps {})".format(
nenvs, nb_epoch_cycles, nb_rollout_steps))
logger.info("nb_epochs: {}, number of cycles to use".format(nb_epochs))
logger.info("eval_env None? {}".format(eval_env is None))
logger.info("(end of debugging messages)\n")
# File paths.
checkdir = osp.join(logger.get_dir(), 'checkpoints')
action_dir = osp.join(logger.get_dir(), 'actions')
episode_dir = osp.join(logger.get_dir(), 'ep_all_infos')
os.makedirs(checkdir, exist_ok=True)
os.makedirs(action_dir, exist_ok=True)
os.makedirs(episode_dir, exist_ok=True)
# Daniel: use these two to store past 100 episode history. Report these stats!
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
all_eval_episode_rewards = []
# reward/step: cumulative quantities for each episode in vecenv.
# epoch_{actions,qs} will grow without bound, fyi.
episode_reward = np.zeros(nenvs, dtype=np.float32) #vector
episode_step = np.zeros(nenvs, dtype=int) # vector
episodes = 0 #scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
mb_actions = []
mb_epinfos = []
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
if nenvs > 1:
# if simulating multiple envs in parallel, impossible to reset
# agent at the end of the episode in each of the environments,
# so resetting here instead
agent.reset()
# Daniel: pure data collection (noise added) to populate replay buffer.
# No training until after this, and note the parallel stepping (VecEnv).
for t_rollout in range(nb_rollout_steps):
# Predict next action.
# action: (#_parallel_envs, ac_dim), q: (#_parallel_envs, 1)
action, q, _, _ = agent.step(obs,
apply_noise=True,
compute_Q=True)
# Execute next action.
if rank == 0 and render:
env.render()
# max_action is of dimension A, whereas action is dimension
# (nenvs, A) - the multiplication gets broadcasted to the batch
# scale for execution in env (as far as DDPG is concerned,
# every action is in [-1, 1])
new_obs, r, done, info = env.step(max_action * action)
r = r.astype(np.float32)
# note these outputs are batched from vecenv (first dim = batch).
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping. (Daniel: agent.train() doesn't require these two lists)
epoch_actions.append(action)
epoch_qs.append(q)
# Daniel: Same as PPO2 code.
mb_actions.append(action)
for inf in info:
maybeepinfo = inf.get('episode')
if maybeepinfo:
mb_epinfos.append(inf)
#the batched data will be unrolled in memory.py's append.
agent.store_transition(obs, action, r, new_obs, done)
obs = new_obs
for d in range(len(done)):
if done[d]:
# Episode done.
epoch_episode_rewards.append(
episode_reward[d]) # Entire history
episode_rewards_history.append(
episode_reward[d]) # Last 100 only
epoch_episode_steps.append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if nenvs == 1:
agent.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate. (Daniel: note that no noise is applied here.)
# Also it seems like episodes do not naturally reset before this starts?
# Also, unlike epoch_episode_reward, here we create eval_episode_reward here ...
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
logger.info('Now on the eval_env for {} steps...'.format(
nb_eval_steps))
nenvs_eval = eval_obs.shape[0]
eval_episode_reward = np.zeros(nenvs_eval, dtype=np.float32)
for t_rollout in range(nb_eval_steps):
eval_action, eval_q, _, _ = agent.step(eval_obs,
apply_noise=False,
compute_Q=True)
# scale for execution in env (for DDPG, every action is in [-1, 1])
eval_obs, eval_r, eval_done, eval_info = eval_env.step(
max_action * eval_action)
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
for d in range(len(eval_done)):
if eval_done[d]:
eval_episode_rewards.append(eval_episode_reward[d])
eval_episode_rewards_history.append(
eval_episode_reward[d])
all_eval_episode_rewards.append(
eval_episode_reward[d])
eval_episode_reward[
d] = 0.0 # Daniel: reset for next episode.
if MPI is not None:
mpi_size = MPI.COMM_WORLD.Get_size()
else:
mpi_size = 1
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.get_stats()
combined_stats = stats.copy()
combined_stats['memory/nb_entries'] = memory.nb_entries
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(
episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
combined_stats['train/param_noise_distance'] = np.mean(
epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics. (Daniel: use eval/return_history for plots)
if eval_env is not None:
combined_stats['eval/return'] = np.mean(eval_episode_rewards)
combined_stats['eval/return_history'] = np.mean(
eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s' % x)
combined_stats_sums = np.array(
[np.array(x).flatten()[0] for x in combined_stats.values()])
if MPI is not None:
combined_stats_sums = MPI.COMM_WORLD.allreduce(combined_stats_sums)
combined_stats = {
k: v / mpi_size
for (k, v) in zip(combined_stats.keys(), combined_stats_sums)
}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps_per_env'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(osp.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(osp.join(logdir, 'eval_env_state.pkl'), 'wb') as f:
pickle.dump(eval_env.get_state(), f)
# Daniel: arguable, we can save all episodes but hard if we don't know the steps.
#if eval_env:
# with open(os.path.join(logdir, 'all_eval_episode_rewards.pkl'), 'wb') as f:
# pickle.dump(all_eval_episode_rewards, f)
# Daniel: we can use cycle or epoch for this if condition ... kind of annoying but w/e.
if cycle % save_interval == 0:
logger.info('We are now saving stuff!!')
savepath = osp.join(checkdir, '%.5i' % epoch)
logger.info('Saving model checkpoint to: ', savepath)
agent.save(savepath)
# ------------------------------------------------------------------
# Daniel: extra stuff for debugging PPO on cloth, actions and infos for each episode.
mb_actions = _sf01(np.asarray(mb_actions))
act_savepath = osp.join(action_dir, 'actions_%.5i.pkl' % epoch)
epi_savepath = osp.join(episode_dir, 'infos_%.5i.pkl' % epoch)
with open(act_savepath, 'wb') as fh:
pickle.dump(mb_actions, fh)
with open(epi_savepath, 'wb') as fh:
pickle.dump(mb_epinfos, fh)
# Daniel: we were not resetting earlier. Actually there are other
# epoch_stats which we might consider resetting here?
epoch_episodes = 0
return agent
def run(args, seed, noise_type, layer_norm, evaluation, **kwargs):
import time
import os
import baselines.ddpg.training as training
from baselines.ddpg.models import Actor, Critic
from baselines.ddpg.memory import Memory
from baselines.ddpg.noise import AdaptiveParamNoiseSpec, NormalActionNoise, OrnsteinUhlenbeckActionNoise
import tensorflow as tf
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = common.make_env(args)
env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
if evaluation and rank==0:
eval_env = common.make_env(args)
eval_env = bench.Monitor(eval_env, os.path.join(logger.get_dir(), 'gym_eval'))
env = bench.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(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
"""
run the training of DDPG
:param env_id: (str) the environment ID
:param seed: (int) the initial random seed
:param noise_type: (str) the wanted noises ('adaptive-param', 'normal' or 'ou'), can use multiple noise type by
seperating them with commas
:param layer_norm: (bool) use layer normalization
:param evaluation: (bool) enable evaluation of DDPG training
:param kwargs: (dict) extra keywords for the training.train function
"""
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = bench.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 = bench.Monitor(eval_env,
os.path.join(logger.get_dir(), 'gym_eval'))
env = bench.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(mean=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(
mean=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 __init__(self,
env,
agent_index,
sess,
action_range=(-1., 1.),
reward_scale=0.1,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.975,
clip_norm=10,
batch_size=64,
memory_size=1e6,
tau=0.01,
normalize_returns=False,
normalize_observations=False,
noise_type="adaptive-param_0.1",
layer_norm=True,
nb_layers=2,
nb_neurons=64,
activation='tanh',
**network_kwargs):
super(DDPGAgent, self).__init__(agent_index)
# self.sess = sess
self.nb_actions = env.action_space[agent_index].n
print('agent action_space ' + str(env.action_space[agent_index].n))
self.state_size = env.observation_space[agent_index].shape
self.action_range = action_range
with tf.variable_scope('ddpg_' + str(agent_index)):
critic = Critic(name='critic_' + str(agent_index),
layer_norm=layer_norm,
nb_layers=nb_layers,
nb_neurons=nb_neurons)
actor = Actor(self.nb_actions,
name='actor_' + str(agent_index),
layer_norm=layer_norm,
nb_neurons=nb_neurons,
activation=activation)
memory = Memory(limit=int(memory_size),
action_shape=(self.nb_actions, ),
observation_shape=self.state_size)
action_noise = None
param_noise = None
if noise_type is not None:
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(self.nb_actions),
sigma=float(stddev) * np.ones(self.nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(self.nb_actions),
sigma=float(stddev) * np.ones(self.nb_actions),
dt=env.world.dt,
theta=0.1)
else:
raise RuntimeError('unknown noise type "{}"'.format(
current_noise_type))
self.agent = DDPG(actor,
critic,
memory,
self.state_size, (self.nb_actions, ),
action_range=self.action_range,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(self.agent.__dict__.items()))
self.agent.initialize(sess)
self.agent.reset()
def __init__(self, env_in):
EnvLearner.__init__(self, env_in)
# from baselines.ddpg.models import Actor, Critic
# Parse noise_type
action_noise = None
param_noise = None
noise_type = 'adaptive-param_0.2'
layer_norm = True
nb_actions = self.state_dim
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.
self.buff_len = 10
self.buffer = deque(self.buff_init * self.buff_len,
maxlen=self.buff_len)
obs_space = (self.buff_init[0].size * self.buff_len, )
self.memory = Memory(limit=int(1e6),
action_shape=env_in.observation_space.shape,
observation_shape=obs_space)
self.critic = models.Critic(layer_norm=layer_norm)
self.actor = models.Actor(nb_actions, layer_norm=layer_norm)
self.agent = DDPG(self.actor,
self.critic,
self.memory,
obs_space,
env_in.observation_space.shape,
gamma=0.99,
tau=0.01,
normalize_returns=False,
normalize_observations=True,
batch_size=64,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=1e-2,
actor_lr=1e-5,
critic_lr=1e-5,
enable_popart=False,
clip_norm=None,
reward_scale=1.)
def learn_setup(
network,
env,
seed=None,
total_timesteps=None,
iterations=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=None,
nb_rollout_steps=100,
n_episodes=None,
logspace=True,
n_steps_per_episode=None,
reward_threshold=0,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
noise_level="0.2",
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
exp_name="test",
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
batch_size=64, # per MPI worker
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50,
**network_kwargs):
if logspace:
actor_lr = 10**-actor_lr
critic_lr = 10**-critic_lr
batch_size = 2**int(batch_size)
if seed is None:
seed = 17
seed = int(seed)
tau = 10**-tau
set_global_seeds(seed)
if nb_epoch_cycles is None:
nb_epoch_cycles = n_episodes
nb_rollout_steps = n_steps_per_episode
else:
input("Not using automated interface? ")
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles *
nb_rollout_steps)
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
nb_actions = env.action_space.shape[-1]
assert (np.abs(env.action_space.low) == env.action_space.high
).all() # we assume symmetric actions.
memory = Memory(limit=int(1e5),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
action_noise = None
param_noise = None
if noise_type is not None:
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('_')
stddev = noise_level
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))
max_action = env.action_space.high
#print("actual max action", max_action)
max_action = 1
logger.info(
'scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor,
critic,
memory,
env.observation_space.shape,
env.action_space.shape,
gamma=gamma,
tau=tau,
normalize_returns=normalize_returns,
normalize_observations=normalize_observations,
batch_size=batch_size,
action_noise=action_noise,
param_noise=param_noise,
critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr,
critic_lr=critic_lr,
enable_popart=popart,
clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
sess = U.get_session()
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nenvs = obs.shape[0]
episode_reward = np.zeros(nenvs, dtype=np.float32) #vector
episode_step = np.zeros(nenvs, dtype=int) # vector
episodes = 0 #scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
local_variables = {
"epoch_episode_rewards": epoch_episode_rewards,
"epoch_episode_steps": epoch_episode_steps,
"batch_size": batch_size,
"eval_env": eval_env,
"reward_threshold": reward_threshold,
"epoch_actions": epoch_actions,
"nb_train_steps": nb_train_steps,
"epoch_qs": epoch_qs,
"start_time": start_time,
"epoch_episodes": [epoch_episodes],
"nb_epoch_cycles": nb_epoch_cycles,
"nb_rollout_steps": nb_rollout_steps,
"agent": agent,
"memory": memory,
"max_action": max_action,
"env": env,
"nenvs": nenvs,
"obs": [obs], #Forgive me 6.031
"t": [t],
"episode_reward": episode_reward,
"episode_rewards_history": episode_rewards_history,
"episode_step": episode_step,
"episodes": [episodes],
"rank": rank,
"param_noise_adaption_interval": param_noise_adaption_interval,
"noise_type": noise_type,
"render": render
}
return local_variables
