def hierarchy_sac(
variant,
env_class,
env_kwargs=None,
observation_key="observation",
):
# initialize tensorflow and the multiprocessing interface
maybe_initialize_process()
# run an experiment with multiple agents
if env_kwargs is None:
env_kwargs = {}
# initialize the environment to track the cardinality of actions
env = NormalizedEnv(env_class, **env_kwargs)
action_dim = env.action_space.low.size
observation_dim = env.observation_space.spaces[observation_key].low.size
# create a replay buffer to store data
replay_buffer = StepReplayBuffer(max_num_steps=variant["max_num_steps"])
# create a logging instance
logger = TensorboardLogger(replay_buffer, variant["logging_dir"])
# a dict to store models for saving to the disk
models_dict = {}
# build a hierarchical agent that uses sac
levels = []
for level in range(variant["num_hierarchy_levels"]):
# create policies for each level in the hierarchy
policy = TanhGaussian(
dense(observation_dim + (0 if level == 0 else observation_dim),
2 * (observation_dim if level == 0 else action_dim),
hidden_size=variant["hidden_size"],
num_hidden_layers=variant["num_hidden_layers"]),
optimizer_kwargs=dict(
learning_rate=variant["policy_learning_rate"]),
tau=variant["tau"],
std=None)
# create critics for each level in the hierarchy
qf1 = Gaussian(
dense(observation_dim + (0 if level == 0 else observation_dim) +
(observation_dim if level == 0 else action_dim),
1,
hidden_size=variant["hidden_size"],
num_hidden_layers=variant["num_hidden_layers"]),
optimizer_kwargs=dict(learning_rate=variant["qf_learning_rate"]),
tau=variant["tau"],
std=1.0)
target_qf1 = qf1.clone()
# create critics for each level in the hierarchy
qf2 = Gaussian(
dense(observation_dim + (0 if level == 0 else observation_dim) +
(observation_dim if level == 0 else action_dim),
1,
hidden_size=variant["hidden_size"],
num_hidden_layers=variant["num_hidden_layers"]),
optimizer_kwargs=dict(learning_rate=variant["qf_learning_rate"]),
tau=variant["tau"],
std=1.0)
target_qf2 = qf2.clone()
# relabel the rewards of the lower level policies
relabeled_buffer = (GoalConditioned(
replay_buffer, reward_scale=0.0, goal_conditioned_scale=1.0)
if level > 0 else replay_buffer)
# train the agent using soft actor critic
algorithm = SAC(policy,
qf1,
qf2,
target_qf1,
target_qf2,
relabeled_buffer,
reward_scale=variant["reward_scale"],
discount=variant["discount"],
initial_alpha=variant["initial_alpha"],
alpha_optimizer_kwargs=dict(
learning_rate=variant["policy_learning_rate"]),
target_entropy=(-action_dim),
observation_key=observation_key,
batch_size=variant["batch_size"],
logger=logger,
logging_prefix="sac_level{}/".format(level))
# create a single agent to manage the hierarchy
levels.append(
PolicyAgent(
policy,
time_skip=variant["time_skip"]**(
variant["num_hierarchy_levels"] - 1 - level),
goal_skip=(variant["time_skip"]**(
variant["num_hierarchy_levels"] -
level) if level > 0 else variant["max_path_length"]),
algorithm=algorithm,
observation_key=observation_key))
models_dict["policy_level{}".format(level)] = policy
models_dict["qf1_level{}".format(level)] = qf1
models_dict["qf2_level{}".format(level)] = qf2
models_dict["target_qf1_level{}".format(level)] = target_qf1
models_dict["target_qf2_level{}".format(level)] = target_qf2
# create a hierarchy agent using the list of agents
agent = HierarchyAgent(levels)
# create a saver to record training progress to the disk
saver = LocalSaver(replay_buffer, variant["logging_dir"], **models_dict)
# make a sampler to collect data to warm up the hierarchy
sampler = ParallelSampler(env,
agent,
max_path_length=variant["max_path_length"],
num_workers=variant["num_workers"])
# collect more training samples
sampler.set_weights(agent.get_weights())
paths, returns, num_steps = sampler.collect(
variant["num_warm_up_steps"],
deterministic=False,
keep_data=True,
workers_to_use=variant["num_workers"])
# insert the samples into the replay buffer
for o, a, r in paths:
replay_buffer.insert_path(o, a, r)
# train for a specified number of iterations
for iteration in range(variant["num_epochs"]):
if iteration % variant["num_epochs_per_eval"] == 0:
# evaluate the policy at this step
sampler.set_weights(agent.get_weights())
paths, eval_returns, num_steps = sampler.collect(
variant["num_steps_per_eval"],
deterministic=True,
keep_data=False,
workers_to_use=variant["num_workers"])
logger.record("eval_mean_return", np.mean(eval_returns))
# save the replay buffer and the policies
saver.save()
# collect more training samples
sampler.set_weights(agent.get_weights())
paths, train_returns, num_steps = sampler.collect(
variant["num_steps_per_epoch"],
deterministic=False,
keep_data=True,
workers_to_use=1)
logger.record("train_mean_return", np.mean(train_returns))
# insert the samples into the replay buffer
for o, a, r in paths:
replay_buffer.insert_path(o, a, r)
# train once each for the number of steps collected
for i in range(num_steps):
agent.train()
def ddpg(
variant,
env_class,
env_kwargs=None,
observation_key="observation",
):
# initialize tensorflow and the multiprocessing interface
maybe_initialize_process()
# run an experiment with multiple agents
if env_kwargs is None:
env_kwargs = {}
# initialize the environment to track the cardinality of actions
env = NormalizedEnv(env_class, **env_kwargs)
action_dim = env.action_space.low.size
observation_dim = env.observation_space.spaces[observation_key].low.size
# create a replay buffer to store data
replay_buffer = StepReplayBuffer(max_num_steps=variant["max_num_steps"])
# create a logging instance
logger = TensorboardLogger(replay_buffer, variant["logging_dir"])
# create policies for each level in the hierarchy
policy = Gaussian(
dense(observation_dim,
action_dim,
hidden_size=variant["hidden_size"],
num_hidden_layers=variant["num_hidden_layers"],
output_activation="tanh"),
optimizer_kwargs=dict(learning_rate=variant["policy_learning_rate"]),
tau=variant["tau"],
std=variant["exploration_noise_std"])
target_policy = policy.clone()
# create critics for each level in the hierarchy
qf = Gaussian(
dense(observation_dim + action_dim,
1,
hidden_size=variant["hidden_size"],
num_hidden_layers=variant["num_hidden_layers"]),
optimizer_kwargs=dict(learning_rate=variant["qf_learning_rate"]),
tau=variant["tau"],
std=1.0)
target_qf = qf.clone()
# train the agent using soft actor critic
algorithm = DDPG(policy,
target_policy,
qf,
target_qf,
replay_buffer,
reward_scale=variant["reward_scale"],
discount=variant["discount"],
observation_key=observation_key,
batch_size=variant["batch_size"],
logger=logger,
logging_prefix="ddpg/")
# create a single agent to manage the hierarchy
agent = PolicyAgent(policy,
algorithm=algorithm,
observation_key=observation_key)
# create a saver to record training progress to the disk
saver = LocalSaver(replay_buffer,
variant["logging_dir"],
policy=policy,
target_policy=target_policy,
qf=qf,
target_qf=target_qf)
# load the networks if already trained
saver.load()
# make a sampler to collect data to warm up the hierarchy
sampler = ParallelSampler(env,
agent,
max_path_length=variant["max_path_length"],
num_workers=variant["num_workers"])
# collect more training samples
sampler.set_weights(agent.get_weights())
paths, returns, num_steps = sampler.collect(
variant["num_warm_up_steps"],
deterministic=False,
keep_data=True,
workers_to_use=variant["num_workers"])
# insert the samples into the replay buffer
for o, a, r in paths:
replay_buffer.insert_path(o, a, r)
# train for a specified number of iterations
for iteration in range(variant["num_epochs"]):
if iteration % variant["num_epochs_per_eval"] == 0:
# evaluate the policy at this step
sampler.set_weights(agent.get_weights())
paths, eval_returns, num_steps = sampler.collect(
variant["num_steps_per_eval"],
deterministic=True,
keep_data=False,
workers_to_use=variant["num_workers"])
logger.record("eval_mean_return", np.mean(eval_returns))
# save the replay buffer and the policies
saver.save()
# collect more training samples
sampler.set_weights(agent.get_weights())
paths, train_returns, num_steps = sampler.collect(
variant["num_steps_per_epoch"],
deterministic=False,
keep_data=True,
workers_to_use=1)
logger.record("train_mean_return", np.mean(train_returns))
# insert the samples into the replay buffer
for o, a, r in paths:
replay_buffer.insert_path(o, a, r)
# train once each for the number of steps collected
for i in range(num_steps):
agent.train()
def ppo(
variant,
env_class,
env_kwargs=None,
observation_key="observation",
):
# initialize tensorflow and the multiprocessing interface
maybe_initialize_process()
# run an experiment with multiple agents
if env_kwargs is None:
env_kwargs = {}
# initialize the environment to track the cardinality of actions
env = NormalizedEnv(env_class, **env_kwargs)
action_dim = env.action_space.low.size
observation_dim = env.observation_space.spaces[observation_key].low.size
# create a replay buffer to store data
replay_buffer = PathReplayBuffer(
max_path_length=variant["max_path_length"],
max_num_paths=variant["max_num_paths"])
# create a logging instance
logger = TensorboardLogger(replay_buffer, variant["logging_dir"])
# create policies for each level in the hierarchy
policy = Gaussian(
dense(observation_dim,
action_dim * 2,
hidden_size=variant["hidden_size"],
num_hidden_layers=variant["num_hidden_layers"]),
optimizer_kwargs=dict(learning_rate=variant["policy_learning_rate"]),
std=None)
old_policy = policy.clone()
# create critics for each level in the hierarchy
vf = Gaussian(
dense(observation_dim,
1,
hidden_size=variant["hidden_size"],
num_hidden_layers=variant["num_hidden_layers"]),
optimizer_kwargs=dict(learning_rate=variant["vf_learning_rate"]),
std=1.0)
# train the agent using soft actor critic
algorithm = PPO(
policy,
old_policy,
vf,
replay_buffer,
reward_scale=variant["reward_scale"],
discount=variant["discount"],
epsilon=variant["epsilon"],
lamb=variant["lamb"],
off_policy_updates=variant["off_policy_updates"],
critic_updates=variant["critic_updates"],
observation_key=observation_key,
batch_size=-1, # sample everything in the buffer
logger=logger,
logging_prefix="ppo/")
# create a single agent to manage the hierarchy
agent = PolicyAgent(policy,
algorithm=algorithm,
observation_key=observation_key)
# create a saver to record training progress to the disk
saver = LocalSaver(replay_buffer,
variant["logging_dir"],
policy=policy,
old_policy=old_policy,
vf=vf)
# load the networks if already trained
saver.load()
# make a sampler to collect data to warm up the hierarchy
sampler = ParallelSampler(env,
agent,
max_path_length=variant["max_path_length"],
num_workers=variant["num_workers"])
# train for a specified number of iterations
for iteration in range(variant["num_epochs"]):
# discard all previous samples for on policy learning
replay_buffer.empty()
if iteration % variant["num_epochs_per_eval"] == 0:
# evaluate the policy at this step
sampler.set_weights(agent.get_weights())
paths, eval_returns, num_steps = sampler.collect(
variant["num_steps_per_eval"],
deterministic=True,
keep_data=False,
workers_to_use=variant["num_workers"])
logger.record("eval_mean_return", np.mean(eval_returns))
# save the replay buffer and the policies
saver.save()
# collect more training samples
sampler.set_weights(agent.get_weights())
paths, train_returns, num_steps = sampler.collect(
variant["num_steps_per_epoch"],
deterministic=False,
keep_data=True,
workers_to_use=variant["num_workers"])
logger.record("train_mean_return", np.mean(train_returns))
# insert the samples into the replay buffer
for o, a, r in paths:
replay_buffer.insert_path(o, a, r)
# train once with the on policy data
agent.train()