def __init__(self, config, action_bound):
super(DQNModel, self).__init__(config=config)
self.proposed_action_list = []
self.action_bound = action_bound
action_list = []
for i in range(len(action_bound[0])):
low = action_bound[0][i]
high = action_bound[1][i]
action_list.append(
np.arange(start=low,
stop=high,
step=(high - low) /
self.config.config_dict['ACTION_SPLIT_COUNT']))
action_iterator = itertools.product(*action_list)
self.action_selection_list = []
for action_sample in action_iterator:
self.action_selection_list.append(tf.constant(action_sample))
self.reward_input = tf.placeholder(shape=[None, 1], dtype=tf.float32)
self.state_input = tf.placeholder(
shape=[None] + list(self.config.config_dict['STATE_SPACE']),
dtype=tf.float32)
self.next_state_input = tf.placeholder(
shape=[None] + list(self.config.config_dict['STATE_SPACE']),
dtype=tf.float32)
self.action_input = tf.placeholder(
shape=[None] + list(self.config.config_dict['ACTION_SPACE']),
dtype=tf.float32)
self.done_input = tf.placeholder(shape=[None, 1], dtype=tf.bool)
self.input = tf.concat([self.state_input, self.action_input])
self.done = tf.cast(self.done_input, dtype=tf.float32)
self.q_value_list = []
var_list = None
for action_sample in self.action_selection_list:
q_net, q_output, var_list = NetworkCreator.create_network(
input=tf.concat(self.state_input, action_sample),
network_config=self.config.config_dict['NET_CONFIG'],
net_name=self.config.config_dict['NAME'])
self.q_value_list.append(q_output)
self.var_list = var_list
self.target_q_value_list = []
for action_sample in self.action_selection_list:
q_net, q_output, var_list = NetworkCreator.create_network(
input=tf.concat(self.next_state_input, action_sample),
network_config=self.config.config_dict['NET_CONFIG'],
net_name='TARGET' + self.config.config_dict['NAME'])
self.target_var_list.append(q_output)
self.target_var_list = var_list
self.loss, self.optimizer, self.optimize = self.create_training_method(
)
self.update_target_q_op = self.create_target_q_update()
self.memory = Memory(
limit=1e100,
action_shape=self.config.config_dict['ACTION_SPACE'],
observation_shape=self.config.config_dict['STATE_SPACE'])
self.sess = tf.get_default_session()
def __init__(self, limit, env):
self.limit = limit
self.env = env
self.memory = Memory(
limit=self.limit,
action_shape=self.env.action_space.shape,
observation_shape=self.env.observation_space.shape)
self.file_dir = None
class Expert:
def __init__(self, limit, env):
self.limit = limit
self.env = env
self.memory = Memory(
limit=self.limit,
action_shape=self.env.action_space.shape,
observation_shape=self.env.observation_space.shape)
self.file_dir = None
def load_file(self, file_dir):
self.file_dir = file_dir
expert_file = open(self.file_dir, 'rb')
expert_data = pickle.load(expert_file)
expert_file.close()
for episode_sample in expert_data:
for step_sample in episode_sample:
self.memory.append(step_sample[0], step_sample[1],
step_sample[2], step_sample[3],
step_sample[4])
def sample(self, batch_size):
return self.memory.sample(batch_size)
def set_tf(self, actor, critic, obs_rms, ret_rms, observation_range,
return_range):
self.expert_state = tf.placeholder(tf.float32,
shape=(None, ) +
self.env.observation_space.shape,
name='expert_state')
self.expert_action = tf.placeholder(tf.float32,
shape=(None, ) +
self.env.action_space.shape,
name='expert_action')
normalized_state = tf.clip_by_value(
normalize(self.expert_state, obs_rms), observation_range[0],
observation_range[1])
expert_actor = actor(normalized_state, reuse=True)
normalized_q_with_expert_data = critic(normalized_state,
self.expert_action,
reuse=True)
normalized_q_with_expert_actor = critic(normalized_state,
expert_actor,
reuse=True)
self.Q_with_expert_data = denormalize(
tf.clip_by_value(normalized_q_with_expert_data, return_range[0],
return_range[1]), ret_rms)
self.Q_with_expert_actor = denormalize(
tf.clip_by_value(normalized_q_with_expert_actor, return_range[0],
return_range[1]), ret_rms)
self.critic_loss = tf.reduce_mean(
tf.nn.softplus(self.Q_with_expert_actor - self.Q_with_expert_data))
self.actor_loss = -tf.reduce_mean(self.Q_with_expert_actor)
def __init__(self, config, action_bound, obs_bound):
super().__init__(config=config)
self.obs_dim = self.config.config_dict['STATE_SPACE']
self.obs_dim = self.obs_dim[0] + 1
self.act_dim = self.config.config_dict['ACTION_SPACE'][0]
with tf.variable_scope(name_or_scope=self.config.config_dict['NAME']):
self.scaler = Scaler(self.obs_dim)
self.val_func = NNValueFunction(
self.obs_dim,
hid1_mult=self.config.config_dict['HIDDEN_MULTIPLE'],
name_scope=self.config.config_dict['NAME'])
self.policy = Policy(
self.obs_dim,
self.act_dim,
kl_targ=self.config.config_dict['KL_TARG'],
hid1_mult=self.config.config_dict['HIDDEN_MULTIPLE'],
policy_logvar=self.config.config_dict['POLICY_LOGVAR'],
name_scope=self.config.config_dict['NAME'])
self._real_trajectories = {
'observes': [],
'actions': [],
'rewards': [],
'unscaled_obs': []
}
self._cyber_trajectories = {
'observes': [],
'actions': [],
'rewards': [],
'unscaled_obs': []
}
self._real_trajectories_memory = deque(
maxlen=self.config.config_dict['EPISODE_REAL_MEMORY_SIZE'])
self._cyber_trajectories_memory = deque(
maxlen=self.config.config_dict['EPISODE_CYBER_MEMORY_SIZE'])
self._real_step_count = 0.0
self._cyber_step_count = 0.0
self.action_low = action_bound[0]
self.action_high = action_bound[1]
self._env_status = None
self.real_data_memory = Memory(
limit=10000,
action_shape=self.config.config_dict['ACTION_SPACE'],
observation_shape=self.config.config_dict['STATE_SPACE'])
self.simulation_data_memory = Memory(
limit=10000,
action_shape=self.config.config_dict['ACTION_SPACE'],
observation_shape=self.config.config_dict['STATE_SPACE'])
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
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.
# TODO: Change back to 1e6
memory = Memory(limit=int(1e2), state_shape=env.state_space.shape, 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()
kwargs.pop('state_shape')
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 train(self, env, nb_steps):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(0.2),
desired_action_stddev=float(0.2))
# Configure components.
memory = Memory(limit=int(1e6),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=True)
actor = Actor(nb_actions, layer_norm=True)
# Seed everything to make things reproducible.
seed = self.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()
#load_state("D:\project\osim-rl-helper\ddpg.pkl")
training.train(env=env,
param_noise=param_noise,
restore=True,
action_noise=action_noise,
actor=actor,
critic=critic,
memory=memory,
nb_epochs=1,
nb_epoch_cycles=1,
render_eval=False,
reward_scale=1.0,
render=False,
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=nb_steps,
nb_rollout_steps=5,
nb_eval_steps=5,
batch_size=64)
#save_state("D:\project\osim-rl-helper\ddpg.pkl")
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
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(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
logger.configure(dir='/home/vaisakhs_shaj/Desktop/DeepReinforcementLearning/5_Deep_Deterministic_Policy_Gradients/LOGS/OSIM')
# Create envs.
env = ProstheticsEnv(visualize=True)
env.change_model(model = '2D', difficulty = 0, prosthetic = True, seed=seed)
#env.seed(seed)
#env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
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(2e6), 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 + 2000000 * 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 setup(self,
obs_shape,
nb_actions,
action_spec,
noise_type,
gamma=1.,
tau=0.01,
layer_norm=True):
super(DDPGAgent, self).setup(obs_shape, nb_actions, action_spec,
noise_type, gamma, tau, layer_norm)
self.action_spec_internal = action_spec
self.obs_dim = obs_shape
action_noise = None
param_noise = None
# Parse noise_type
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.memory = Memory(limit=int(500),
action_shape=(nb_actions, ),
observation_shape=obs_shape)
self.critic = Critic(layer_norm=layer_norm, hidden_size=128)
self.actor = Actor(nb_actions, layer_norm=layer_norm, hidden_size=128)
tf.reset_default_graph()
# max_action = env.action_space.high
self.ddpg = DDPG(actor=self.actor,
critic=self.critic,
memory=self.memory,
observation_shape=obs_shape,
action_shape=(nb_actions, ),
gamma=gamma,
tau=tau,
action_noise=action_noise,
param_noise=param_noise)
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 train_ddpg(env, N_episodes):
param_noise = None
nb_actions = env.action_space.shape[-1]
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
# 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)
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
param_noise = None
# 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
nb_actions = env.action_space.shape[-1]
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions),
sigma=np.ones(nb_actions))
# Configure components.
memory = Memory(limit=int(1e6),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
actor = Actor(nb_actions, layer_norm=layer_norm)
critic = Critic(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,
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 main():
args = parse_args()
# create the environment
env = gym.make("kuka-v0") # <-- this we need to create
env.init_bullet(render=True)
# create the learning agent
# model = deepq.models.mlp([16, 16])
ob_dim = env.observation_space.shape[0]
ac_dim = env.action_space.n
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
noise_type = 'adaptive-param_0.2'
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))
# policy = GaussianMlpPolicy(ob_dim, ac_dim)
# vf = NeuralNetValueFunction(ob_dim, ac_dim)
# learn(env, policy=policy, vf=vf,
# gamma=0.99, lam=0.97, timesteps_per_batch=2500,
# desired_kl=0.002,
# num_timesteps=1000, animate=False)
# Configure components.
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(layer_norm=False)
actor = Actor(nb_actions, layer_norm=False)
training.train(env=env, param_noise=param_noise,
action_noise=action_noise, actor=actor, critic=critic, memory=memory, **args)
env.close()
def __init__(self, hyperparams, dX, dU):
"""Initializes the policy.
Args:
hyperparams: Dictionary of hyperparameters.
dX: Dimension of state space.
dU: Dimension of action space.
"""
PolicyOpt.__init__(self, hyperparams, dX, dU)
self.dX = dX
self.dU = dU
self.epochs = hyperparams['epochs']
self.param_noise_adaption_interval = hyperparams[
'param_noise_adaption_interval']
set_global_seeds(hyperparams['seed'])
# Initialize DDPG policy
self.pol = DDPG(Actor(dU,
network=hyperparams['network'],
**hyperparams['network_kwargs']),
Critic(network=hyperparams['network'],
**hyperparams['network_kwargs']),
Memory(limit=hyperparams['memory_limit'],
action_shape=(dU, ),
observation_shape=(dX, )),
observation_shape=(dX, ),
action_shape=(dU, ),
param_noise=AdaptiveParamNoiseSpec(
initial_stddev=0.2, desired_action_stddev=0.2),
**hyperparams['ddpg_kwargs'])
sess = get_session()
self.pol.initialize(sess)
sess.graph.finalize()
self.policy = self # Act method is contained in this class
class DQNModel(TensorflowBasedModel):
key_list = Config.load_json(file_path=None)
def __init__(self, config, action_bound):
super(DQNModel, self).__init__(config=config)
self.proposed_action_list = []
self.action_bound = action_bound
action_list = []
for i in range(len(action_bound[0])):
low = action_bound[0][i]
high = action_bound[1][i]
action_list.append(
np.arange(start=low,
stop=high,
step=(high - low) /
self.config.config_dict['ACTION_SPLIT_COUNT']))
action_iterator = itertools.product(*action_list)
self.action_selection_list = []
for action_sample in action_iterator:
self.action_selection_list.append(tf.constant(action_sample))
self.reward_input = tf.placeholder(shape=[None, 1], dtype=tf.float32)
self.state_input = tf.placeholder(
shape=[None] + list(self.config.config_dict['STATE_SPACE']),
dtype=tf.float32)
self.next_state_input = tf.placeholder(
shape=[None] + list(self.config.config_dict['STATE_SPACE']),
dtype=tf.float32)
self.action_input = tf.placeholder(
shape=[None] + list(self.config.config_dict['ACTION_SPACE']),
dtype=tf.float32)
self.done_input = tf.placeholder(shape=[None, 1], dtype=tf.bool)
self.input = tf.concat([self.state_input, self.action_input])
self.done = tf.cast(self.done_input, dtype=tf.float32)
self.q_value_list = []
var_list = None
for action_sample in self.action_selection_list:
q_net, q_output, var_list = NetworkCreator.create_network(
input=tf.concat(self.state_input, action_sample),
network_config=self.config.config_dict['NET_CONFIG'],
net_name=self.config.config_dict['NAME'])
self.q_value_list.append(q_output)
self.var_list = var_list
self.target_q_value_list = []
for action_sample in self.action_selection_list:
q_net, q_output, var_list = NetworkCreator.create_network(
input=tf.concat(self.next_state_input, action_sample),
network_config=self.config.config_dict['NET_CONFIG'],
net_name='TARGET' + self.config.config_dict['NAME'])
self.target_var_list.append(q_output)
self.target_var_list = var_list
self.loss, self.optimizer, self.optimize = self.create_training_method(
)
self.update_target_q_op = self.create_target_q_update()
self.memory = Memory(
limit=1e100,
action_shape=self.config.config_dict['ACTION_SPACE'],
observation_shape=self.config.config_dict['STATE_SPACE'])
self.sess = tf.get_default_session()
def update(self):
for i in range(self.config.config_dict['ITERATION_EVER_EPOCH']):
batch_data = self.memory.sample(
batch_size=self.config.config_dict['BATCH_SIZE'])
loss = self.sess.run(fetches=[self.loss, self.optimize],
feed_dict={
self.reward_input: batch_data['rewards'],
self.action_input: batch_data['actions'],
self.state_input: batch_data['obs0'],
self.done_input: batch_data['terminals1']
})
def predict(self, obs, q_value):
pass
def print_log_queue(self, status):
self.status = status
while self.log_queue.qsize() > 0:
log = self.log_queue.get()
print("%s: Critic loss %f: " %
(self.name, log[self.name + '_CRITIC']))
log['INDEX'] = self.log_print_count
self.log_file_content.append(log)
self.log_print_count += 1
def create_training_method(self):
l1_l2 = tfcontrib.layers.l1_l2_regularizer()
loss = tf.reduce_sum((self.predict_q_value - self.q_output) ** 2) + \
tfcontrib.layers.apply_regularization(l1_l2, weights_list=self.var_list)
optimizer = tf.train.AdadeltaOptimizer(
learning_rate=self.config.config_dict['LEARNING_RATE'])
optimize_op = optimizer.minimize(loss=loss, var_list=self.var_list)
return loss, optimizer, optimize_op
def create_predict_q_value_op(self):
predict_q_value = (1. - self.done) * self.config.config_dict['DISCOUNT'] * self.target_q_output \
+ self.reward_input
return predict_q_value
def create_target_q_update(self):
op = []
for var, target_var in zip(self.var_list, self.target_var_list):
ref_val = self.config.config_dict['DECAY'] * target_var + (
1.0 - self.config.config_dict['DECAY']) * var
op.append(tf.assign(ref_val, var))
return op
def store_one_sample(self, state, next_state, action, reward, done, *arg,
**kwargs):
self.memory.append(obs0=state,
obs1=next_state,
action=action,
reward=reward,
terminal1=done)
_, 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))
nb_actions = env.action_space.shape[-1]
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)
tf.reset_default_graph()
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,
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
# env = CartpoleSwingupEnvX()
env = experiments[env_id]['env_call']()
if experiments[env_id]['normalize_env']:
env = normalize(env)
env = bench.Monitor(
env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
# env = gym.wrappers.Monitor(env, log_dir, video_callable=False,
# force=True)
gym.logger.setLevel(logging.WARN)
if evaluation and rank == 0:
eval_env = experiments[env_id]['env_call']()
if experiments[env_id]['normalize_env']:
eval_env = normalize(eval_env)
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))
dataPrimary = pd.read_csv("data_p/monitorChange.csv")
dataProgress = pd.read_csv("data_p/progress.csv")
action_shape = (1, )
nb_action = 1
observation_shape = (3, )
t_train_time = 10000
t_test_time = 10000
network = 'mlp'
action_noise = None
param_noise = None
popart = False,
load_path = 'ddpg_model'
load_path = None
memory = Memory(limit=int(1e6),
action_shape=action_shape,
observation_shape=observation_shape)
critic = Critic(network=network)
actor = Actor(nb_action, network=network)
agent = DDPG(actor,
critic,
memory,
observation_shape,
action_shape,
gamma=0.99,
tau=0.01,
normalize_returns=False,
normalize_observations=True,
batch_size=32,
action_noise=action_noise,
def run(env_id, seed, noise_type, layer_norm, evaluation, perform, use_expert,
expert_dir, use_trpo_expert, expert_limit, **kwargs):
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
gym.logger.setLevel(logging.WARN)
if evaluation and perform:
perform = False
if evaluation and rank == 0 or perform:
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(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)
if use_expert:
expert = Expert(limit=expert_limit, env=env)
if expert_dir is None:
expert_dir = os.path.join('./expert',
env.env.spec.id) + '/expert.pkl'
expert.load_file(expert_dir)
elif use_trpo_expert:
assert expert_dir is not None
expert = Expert(limit=expert_limit, env=env)
expert.load_file_trpo(expert_dir)
else:
expert = None
# 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,
perform=perform,
expert=expert,
**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 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]
env.seed(os.getpid())
set_global_seeds(os.getpid())
num_traj = 0
# 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=None,
param_noise=None,
critic_l2_reg=self.critic_l2_reg,
enable_popart=self.popart,
clip_norm=self.clip_norm,
reward_scale=self.reward_scale)
# Build the testing logic fn
testing_fn = make_testing_fn(agent, env, self.episode_length,
self.action_repeat, self.max_action,
self.nb_episodes)
# 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:
message, actor_ws, global_step = self.inputQ.get(
) # Pop message
if message == 'test':
# Set weights
set_parameters(actor_ws)
# Do testing
rewards, step_times, distances, episode_lengths = testing_fn(
)
self.outputQ.put((rewards, step_times, distances,
episode_lengths, global_step))
# 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 main():
with U.single_threaded_session() as sess:
batch_size = 64
current_noise_type = 'adaptive-param_0.2'
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
param_noise_adaption_interval = 2
env = gym.make("Pendulum-v0")
nb_actions = env.action_space.shape[-1]
layer_norm = True
# 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 = int(1000000 * np.random.rand())
logger.info('seed={}, logdir={}'.format(seed, logger.get_dir()))
tf.set_random_seed(seed)
np.random.seed(seed)
random.seed(seed)
env.seed(seed)
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,
batch_size=batch_size, param_noise=param_noise)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
for t in itertools.count():
episode_rewards = []
done = False
while not done:
env.render()
# Take action and update exploration to the newest value
action, q = agent.pi(obs, apply_noise=True, compute_Q=True)
new_obs, rew, done, _ = env.step(max_action * action)
# Book-keeping.
agent.store_transition(obs, action, rew, new_obs, done)
obs = new_obs
episode_rewards.append(rew)
if done:
agent.reset()
obs = env.reset()
nb_train_steps = 100
epoch_adaptive_distances = []
epoch_critic_losses = []
epoch_actor_losses = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t % 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()
if t % 10 == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", len(episode_rewards))
logger.record_tabular("mean episode reward", round(np.mean(episode_rewards), 1))
logger.record_tabular('train/loss_actor', round(np.mean(epoch_actor_losses)))
logger.record_tabular('train/loss_critic', round(np.mean(epoch_critic_losses)))
logger.record_tabular('train/param_noise_distance', round(np.mean(epoch_adaptive_distances)))
logger.dump_tabular()
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 run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
logging.basicConfig(filename='noGazebo_ddpg.log',
level=logging.DEBUG,
filemode="w")
logging.getLogger().addHandler(logging.StreamHandler())
# Configure logger for the process with rank 0 (main-process?)
# MPI = Message Passing Interface, for parallel computing; rank = process identifier within a group of processes
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
# Disable logging for rank != 0 to avoid noise.
logging.debug(
"I'm MPI worker {} and I guess I just log nothing".format(rank))
logger.set_level(logger.DISABLED)
logging.disable(logging.CRITICAL)
logging.info(
"********************************************* Starting RL algorithm *********************************************"
)
now = datetime.datetime.now()
logging.info(now.isoformat())
# Create envs.
env = gym.make(env_id)
env = bench.Monitor(env,
logger.get_dir()
and os.path.join(logger.get_dir(), str(rank)),
allow_early_resets=True)
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[0]
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. (initialize memory, critic & actor objects)
logging.info("action space of env: {}".format(env.action_space)) # Box(2,)
logging.info("observation space of env: {}".format(
env.observation_space)) # Box(51200,)
memory = Memory(limit=int(1e4),
action_shape=(env.action_space.shape[0], ),
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)
# Train the RL algorithm
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)
# Training is done
env.close()
if eval_env is not None:
eval_env.close()
logger.info('total runtime: {}s'.format(time.time() - start_time))
now = datetime.datetime.now()
logging.info(now.isoformat())
logging.info(
"********************************************* End of RL algorithm *********************************************"
)
return True
def __init__(self, config, action_bound, obs_bound):
super(DDPGModelNew, self).__init__(config=config)
self.action_noise = None
self.para_noise = None
if self.config.config_dict['NOISE_FLAG']:
nb_actions = self.config.config_dict['ACTION_SPACE']
noise_type = self.config.config_dict['NOISE_TYPE']
action_noise = None
param_noise = 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))
self.action_noise = action_noise
self.para_noise = param_noise
actor = Actor(nb_actions=self.config.config_dict['ACTION_SPACE'][0],
layer_norm=self.config.config_dict['LAYER_NORM_FLAG'],
net_config=self.config.config_dict['ACTOR_LAYER_CONFIG'],
action_low=action_bound[0],
action_high=action_bound[1])
critic = Critic(net_config=self.config.config_dict['CRITIC_LAYER_CONFIG'])
self.real_data_memory = Memory(limit=int(1e5),
action_shape=self.config.config_dict['ACTION_SPACE'],
observation_shape=self.config.config_dict['STATE_SPACE'])
self.simulation_data_memory = Memory(limit=int(1e5),
action_shape=self.config.config_dict['ACTION_SPACE'],
observation_shape=self.config.config_dict['STATE_SPACE'])
# TODO deal with obs range
self.ddpg_model = baseline_ddpg(actor=actor,
critic=critic,
memory=self.real_data_memory,
observation_shape=self.config.config_dict['STATE_SPACE'],
action_shape=self.config.config_dict['ACTION_SPACE'],
param_noise=self.para_noise,
action_noise=self.action_noise,
gamma=self.config.config_dict['GAMMA'],
tau=self.config.config_dict['TAU'],
action_range=action_bound,
return_range=(-np.inf, np.inf),
normalize_observations=False,
actor_lr=self.config.config_dict['ACTOR_LEARNING_RATE'],
critic_lr=self.config.config_dict['CRITIC_LEARNING_RATE'],
critic_l2_reg=self.config.config_dict['CRITIC_L2_REG'],
batch_size=self.config.config_dict['BATCH_SIZE'],
observation_range=(-np.inf, np.inf))
self.ddpg_model.sess = tf.get_default_session()
var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
self.var_list = []
for var in var_list:
# TODO THIS MAY LEAD TO SOME BUGS IN THE FUTURE
if 'actor' in var.name or 'critic' in var.name or 'obs' in var.name:
self.var_list.append(var)
self.variables_initializer = tf.variables_initializer(var_list=self.var_list)
self._env_status = None
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 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 run(env_id, seed, noise_type, num_cpu, layer_norm, logdir, gym_monitor, evaluation, bind_to_core, **kwargs):
kwargs['logdir'] = logdir
whoami = mpi_fork(num_cpu, bind_to_core=bind_to_core)
if whoami == 'parent':
sys.exit(0)
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
# Write to temp directory for all non-master workers.
actual_dir = None
Logger.CURRENT.close()
Logger.CURRENT = Logger(dir=mkdtemp(), output_formats=[])
logger.set_level(logger.DISABLED)
# Create envs.
if rank == 0:
env = gym.make(env_id)
if gym_monitor and logdir:
env = gym.wrappers.Monitor(env, os.path.join(logdir, 'gym_train'), force=True)
env = SimpleMonitor(env)
if evaluation:
eval_env = gym.make(env_id)
if gym_monitor and logdir:
eval_env = gym.wrappers.Monitor(eval_env, os.path.join(logdir, 'gym_eval'), force=True)
eval_env = SimpleMonitor(eval_env)
else:
eval_env = None
else:
env = gym.make(env_id)
if evaluation:
eval_env = gym.make(env_id)
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()
Logger.CURRENT.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
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 run(env_id, seed, noise_type, layer_norm, evaluation, outdir, no_hyp,
**kwargs):
params = locals()
# Configure things.
# rank = MPI.COMM_WORLD.Get_rank()
# if rank != 0: logger.set_level(logger.DISABLED)
rank = 0
# Create envs.
env = make_env(env_id)
weight_file = kwargs.pop('weight_file')
if not weight_file:
outdir = exp_utils.prepare_exp_dirs(params, outdir, env_id)
else:
outdir = exp_utils.prepare_exp_dirs(params, outdir, env_id, 'eval')
logger.configure(outdir)
os.makedirs(outdir, exist_ok=True)
env = bench.Monitor(env, os.path.join(outdir, "%i.monitor.json" % rank))
gym.logger.setLevel(logging.WARN)
logger.info('Output directory:{}, env:{}, no_hyp:{}'.format(
outdir, env_id, no_hyp))
if evaluation:
eval_env = make_env(env_id)
eval_env.seed(42)
eval_env = bench.Monitor(eval_env,
os.path.join(logger.get_dir(), 'gym_eval'),
allow_early_resets=True)
# 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(1e5),
action_shape=env.action_space.shape,
observation_shape=env.observation_space.shape)
# critic = models.ConvCritic(layer_norm=layer_norm)
# actor = models.ConvActor(nb_actions, layer_norm=layer_norm, no_hyp=no_hyp)
critic = Critic(layer_norm=layer_norm)
actor = Actor(nb_actions, layer_norm=layer_norm, no_hyp=no_hyp)
# 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()
if weight_file:
evaluate(
env,
nb_episodes=kwargs.get('nb_epochs', 100),
reward_scale=kwargs.get('reward_scale'),
render=kwargs.get('render'),
param_noise=None,
action_noise=None,
actor=actor,
critic=critic,
critic_l2_reg=kwargs.get('critic_l2_reg'),
memory=memory,
weight_file=weight_file,
)
else:
training.train(env=env,
eval_env=eval_env,
param_noise=param_noise,
action_noise=action_noise,
actor=actor,
critic=critic,
memory=memory,
outdir=outdir,
no_hyp=no_hyp,
**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))
