def test_ppo_get_value_estimates(mock_communicator, mock_launcher,
dummy_config):
tf.reset_default_graph()
mock_communicator.return_value = MockCommunicator(discrete_action=False,
visual_inputs=0)
env = UnityEnvironment(" ")
brain_infos = env.reset()
brain_info = brain_infos[env.external_brain_names[0]]
trainer_parameters = dummy_config
model_path = env.external_brain_names[0]
trainer_parameters["model_path"] = model_path
trainer_parameters["keep_checkpoints"] = 3
policy = PPOPolicy(0, env.brains[env.external_brain_names[0]],
trainer_parameters, False, False)
run_out = policy.get_value_estimates(brain_info, 0, done=False)
for key, val in run_out.items():
assert type(key) is str
assert type(val) is float
run_out = policy.get_value_estimates(brain_info, 0, done=True)
for key, val in run_out.items():
assert type(key) is str
assert val == 0.0
# Check if we ignore terminal states properly
policy.reward_signals["extrinsic"].use_terminal_states = False
run_out = policy.get_value_estimates(brain_info, 0, done=True)
for key, val in run_out.items():
assert type(key) is str
assert val != 0.0
env.close()
def test_ppo_get_value_estimates(mock_communicator, mock_launcher,
dummy_config):
tf.reset_default_graph()
brain_params = BrainParameters(
brain_name="test_brain",
vector_observation_space_size=1,
camera_resolutions=[],
vector_action_space_size=[2],
vector_action_descriptions=[],
vector_action_space_type=0,
)
dummy_config["summary_path"] = "./summaries/test_trainer_summary"
dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
policy = PPOPolicy(0, brain_params, dummy_config, False, False)
time_horizon = 15
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
run_out = policy.get_value_estimates(trajectory.next_obs,
"test_agent",
done=False)
for key, val in run_out.items():
assert type(key) is str
assert type(val) is float
run_out = policy.get_value_estimates(trajectory.next_obs,
"test_agent",
done=True)
for key, val in run_out.items():
assert type(key) is str
assert val == 0.0
# Check if we ignore terminal states properly
policy.reward_signals["extrinsic"].use_terminal_states = False
run_out = policy.get_value_estimates(trajectory.next_obs,
"test_agent",
done=True)
for key, val in run_out.items():
assert type(key) is str
assert val != 0.0
agentbuffer = trajectory.to_agentbuffer()
batched_values = policy.get_batched_value_estimates(agentbuffer)
for values in batched_values.values():
assert len(values) == 15
class PPOTrainer(RLTrainer):
"""The PPOTrainer is an implementation of the PPO algorithm."""
def __init__(
self,
brain,
reward_buff_cap,
trainer_parameters,
training,
load,
seed,
run_id,
multi_gpu,
):
"""
Responsible for collecting experiences and training PPO model.
:param trainer_parameters: The parameters for the trainer (dictionary).
:param reward_buff_cap: Max reward history to track in the reward buffer
:param training: Whether the trainer is set for training.
:param load: Whether the model should be loaded.
:param seed: The seed the model will be initialized with
:param run_id: The identifier of the current run
"""
super(PPOTrainer, self).__init__(brain, trainer_parameters, training,
run_id, reward_buff_cap)
self.param_keys = [
"batch_size",
"beta",
"buffer_size",
"epsilon",
"hidden_units",
"lambd",
"learning_rate",
"max_steps",
"normalize",
"num_epoch",
"num_layers",
"time_horizon",
"sequence_length",
"summary_freq",
"use_recurrent",
"summary_path",
"memory_size",
"model_path",
"reward_signals",
]
self.check_param_keys()
if multi_gpu and len(get_devices()) > 1:
self.ppo_policy = MultiGpuPPOPolicy(seed, brain,
trainer_parameters,
self.is_training, load)
else:
self.ppo_policy = PPOPolicy(seed, brain, trainer_parameters,
self.is_training, load)
self.policy = self.ppo_policy
for _reward_signal in self.policy.reward_signals.keys():
self.collected_rewards[_reward_signal] = {}
def process_experiences(self, current_info: BrainInfo,
next_info: BrainInfo) -> None:
"""
Checks agent histories for processing condition, and processes them as necessary.
Processing involves calculating value and advantage targets for model updating step.
:param current_info: current BrainInfo.
:param next_info: next BrainInfo.
"""
if self.is_training:
self.policy.update_normalization(next_info.vector_observations)
for l in range(len(next_info.agents)):
agent_actions = self.processing_buffer[
next_info.agents[l]]["actions"]
if (next_info.local_done[l] or len(agent_actions) >
self.trainer_parameters["time_horizon"]
) and len(agent_actions) > 0:
agent_id = next_info.agents[l]
if next_info.max_reached[l]:
bootstrapping_info = self.processing_buffer[
agent_id].last_brain_info
idx = bootstrapping_info.agents.index(agent_id)
else:
bootstrapping_info = next_info
idx = l
value_next = self.ppo_policy.get_value_estimates(
bootstrapping_info,
idx,
next_info.local_done[l] and not next_info.max_reached[l],
)
tmp_advantages = []
tmp_returns = []
for name in self.policy.reward_signals:
bootstrap_value = value_next[name]
local_rewards = self.processing_buffer[agent_id][
"{}_rewards".format(name)].get_batch()
local_value_estimates = self.processing_buffer[agent_id][
"{}_value_estimates".format(name)].get_batch()
local_advantage = get_gae(
rewards=local_rewards,
value_estimates=local_value_estimates,
value_next=bootstrap_value,
gamma=self.policy.reward_signals[name].gamma,
lambd=self.trainer_parameters["lambd"],
)
local_return = local_advantage + local_value_estimates
# This is later use as target for the different value estimates
self.processing_buffer[agent_id]["{}_returns".format(
name)].set(local_return)
self.processing_buffer[agent_id]["{}_advantage".format(
name)].set(local_advantage)
tmp_advantages.append(local_advantage)
tmp_returns.append(local_return)
global_advantages = list(
np.mean(np.array(tmp_advantages, dtype=np.float32),
axis=0))
global_returns = list(
np.mean(np.array(tmp_returns, dtype=np.float32), axis=0))
self.processing_buffer[agent_id]["advantages"].set(
global_advantages)
self.processing_buffer[agent_id]["discounted_returns"].set(
global_returns)
self.processing_buffer.append_to_update_buffer(
self.update_buffer,
agent_id,
batch_size=None,
training_length=self.policy.sequence_length,
)
self.processing_buffer[agent_id].reset_agent()
if next_info.local_done[l]:
self.stats["Environment/Episode Length"].append(
self.episode_steps.get(agent_id, 0))
self.episode_steps[agent_id] = 0
for name, rewards in self.collected_rewards.items():
if name == "environment":
self.cumulative_returns_since_policy_update.append(
rewards.get(agent_id, 0))
self.stats["Environment/Cumulative Reward"].append(
rewards.get(agent_id, 0))
self.reward_buffer.appendleft(
rewards.get(agent_id, 0))
rewards[agent_id] = 0
else:
self.stats[self.policy.reward_signals[name].
stat_name].append(
rewards.get(agent_id, 0))
rewards[agent_id] = 0
def add_policy_outputs(self, take_action_outputs: ActionInfoOutputs,
agent_id: str, agent_idx: int) -> None:
"""
Takes the output of the last action and store it into the training buffer.
"""
actions = take_action_outputs["action"]
if self.policy.use_continuous_act:
actions_pre = take_action_outputs["pre_action"]
self.processing_buffer[agent_id]["actions_pre"].append(
actions_pre[agent_idx])
a_dist = take_action_outputs["log_probs"]
# value is a dictionary from name of reward to value estimate of the value head
self.processing_buffer[agent_id]["actions"].append(actions[agent_idx])
self.processing_buffer[agent_id]["action_probs"].append(
a_dist[agent_idx])
def add_rewards_outputs(
self,
rewards_out: AllRewardsOutput,
values: Dict[str, np.ndarray],
agent_id: str,
agent_idx: int,
agent_next_idx: int,
) -> None:
"""
Takes the value output of the last action and store it into the training buffer.
"""
for name, reward_result in rewards_out.reward_signals.items():
# 0 because we use the scaled reward to train the agent
self.processing_buffer[agent_id]["{}_rewards".format(name)].append(
reward_result.scaled_reward[agent_next_idx])
self.processing_buffer[agent_id]["{}_value_estimates".format(
name)].append(values[name][agent_idx][0])
def is_ready_update(self):
"""
Returns whether or not the trainer has enough elements to run update model
:return: A boolean corresponding to whether or not update_model() can be run
"""
size_of_buffer = self.update_buffer.num_experiences
return size_of_buffer > self.trainer_parameters["buffer_size"]
def update_policy(self):
"""
Uses demonstration_buffer to update the policy.
The reward signal generators must be updated in this method at their own pace.
"""
buffer_length = self.update_buffer.num_experiences
self.trainer_metrics.start_policy_update_timer(
number_experiences=buffer_length,
mean_return=float(
np.mean(self.cumulative_returns_since_policy_update)),
)
self.cumulative_returns_since_policy_update.clear()
# Make sure batch_size is a multiple of sequence length. During training, we
# will need to reshape the data into a batch_size x sequence_length tensor.
batch_size = (self.trainer_parameters["batch_size"] -
self.trainer_parameters["batch_size"] %
self.policy.sequence_length)
# Make sure there is at least one sequence
batch_size = max(batch_size, self.policy.sequence_length)
n_sequences = max(
int(self.trainer_parameters["batch_size"] /
self.policy.sequence_length), 1)
advantages = self.update_buffer["advantages"].get_batch()
self.update_buffer["advantages"].set(
(advantages - advantages.mean()) / (advantages.std() + 1e-10))
num_epoch = self.trainer_parameters["num_epoch"]
batch_update_stats = defaultdict(list)
for _ in range(num_epoch):
self.update_buffer.shuffle(
sequence_length=self.policy.sequence_length)
buffer = self.update_buffer
max_num_batch = buffer_length // batch_size
for l in range(0, max_num_batch * batch_size, batch_size):
update_stats = self.policy.update(
buffer.make_mini_batch(l, l + batch_size), n_sequences)
for stat_name, value in update_stats.items():
batch_update_stats[stat_name].append(value)
for stat, stat_list in batch_update_stats.items():
self.stats[stat].append(np.mean(stat_list))
if self.policy.bc_module:
update_stats = self.policy.bc_module.update()
for stat, val in update_stats.items():
self.stats[stat].append(val)
self.clear_update_buffer()
self.trainer_metrics.end_policy_update()
class PPOTrainer(Trainer):
"""The PPOTrainer is an implementation of the PPO algorithm."""
def __init__(self, brain, reward_buff_cap, trainer_parameters, training,
load, seed, run_id):
"""
Responsible for collecting experiences and training PPO model.
:param trainer_parameters: The parameters for the trainer (dictionary).
:param reward_buff_cap: Max reward history to track in the reward buffer
:param training: Whether the trainer is set for training.
:param load: Whether the model should be loaded.
:param seed: The seed the model will be initialized with
:param run_id: The identifier of the current run
"""
super().__init__(brain, trainer_parameters, training, run_id,
reward_buff_cap)
self.param_keys = [
"batch_size",
"beta",
"buffer_size",
"epsilon",
"hidden_units",
"lambd",
"learning_rate",
"max_steps",
"normalize",
"num_epoch",
"num_layers",
"time_horizon",
"sequence_length",
"summary_freq",
"use_recurrent",
"summary_path",
"memory_size",
"model_path",
"reward_signals",
]
self.check_param_keys()
# Make sure we have at least one reward_signal
if not self.trainer_parameters["reward_signals"]:
raise UnityTrainerException(
"No reward signals were defined. At least one must be used with {}."
.format(self.__class__.__name__))
self.step = 0
self.policy = PPOPolicy(seed, brain, trainer_parameters,
self.is_training, load)
stats = defaultdict(list)
# collected_rewards is a dictionary from name of reward signal to a dictionary of agent_id to cumulative reward
# used for reporting only. We always want to report the environment reward to Tensorboard, regardless
# of what reward signals are actually present.
self.collected_rewards = {"environment": {}}
for _reward_signal in self.policy.reward_signals.keys():
self.collected_rewards[_reward_signal] = {}
self.stats = stats
self.training_buffer = Buffer()
self.episode_steps = {}
def __str__(self):
return """Hyperparameters for the {0} of brain {1}: \n{2}""".format(
self.__class__.__name__,
self.brain_name,
self.dict_to_str(self.trainer_parameters, 0),
)
@property
def parameters(self):
"""
Returns the trainer parameters of the trainer.
"""
return self.trainer_parameters
@property
def get_max_steps(self):
"""
Returns the maximum number of steps. Is used to know when the trainer should be stopped.
:return: The maximum number of steps of the trainer
"""
return float(self.trainer_parameters["max_steps"])
@property
def get_step(self):
"""
Returns the number of steps the trainer has performed
:return: the step count of the trainer
"""
return self.step
def increment_step(self, n_steps: int) -> None:
"""
Increment the step count of the trainer
:param n_steps: number of steps to increment the step count by
"""
self.step = self.policy.increment_step(n_steps)
def construct_curr_info(self, next_info: BrainInfo) -> BrainInfo:
"""
Constructs a BrainInfo which contains the most recent previous experiences for all agents
which correspond to the agents in a provided next_info.
:BrainInfo next_info: A t+1 BrainInfo.
:return: curr_info: Reconstructed BrainInfo to match agents of next_info.
"""
visual_observations: List[List[Any]] = [
[]
] # TODO add types to brain.py methods
vector_observations = []
text_observations = []
memories = []
rewards = []
local_dones = []
max_reacheds = []
agents = []
prev_vector_actions = []
prev_text_actions = []
action_masks = []
for agent_id in next_info.agents:
agent_brain_info = self.training_buffer[agent_id].last_brain_info
if agent_brain_info is None:
agent_brain_info = next_info
agent_index = agent_brain_info.agents.index(agent_id)
for i in range(len(next_info.visual_observations)):
visual_observations[i].append(
agent_brain_info.visual_observations[i][agent_index])
vector_observations.append(
agent_brain_info.vector_observations[agent_index])
text_observations.append(
agent_brain_info.text_observations[agent_index])
if self.policy.use_recurrent:
if len(agent_brain_info.memories) > 0:
memories.append(agent_brain_info.memories[agent_index])
else:
memories.append(self.policy.make_empty_memory(1))
rewards.append(agent_brain_info.rewards[agent_index])
local_dones.append(agent_brain_info.local_done[agent_index])
max_reacheds.append(agent_brain_info.max_reached[agent_index])
agents.append(agent_brain_info.agents[agent_index])
prev_vector_actions.append(
agent_brain_info.previous_vector_actions[agent_index])
prev_text_actions.append(
agent_brain_info.previous_text_actions[agent_index])
action_masks.append(agent_brain_info.action_masks[agent_index])
if self.policy.use_recurrent:
memories = np.vstack(memories)
curr_info = BrainInfo(
visual_observations,
vector_observations,
text_observations,
memories,
rewards,
agents,
local_dones,
prev_vector_actions,
prev_text_actions,
max_reacheds,
action_masks,
)
return curr_info
def add_experiences(
self,
curr_all_info: AllBrainInfo,
next_all_info: AllBrainInfo,
take_action_outputs: ActionInfoOutputs,
) -> None:
"""
Adds experiences to each agent's experience history.
:param curr_all_info: Dictionary of all current brains and corresponding BrainInfo.
:param next_all_info: Dictionary of all current brains and corresponding BrainInfo.
:param take_action_outputs: The outputs of the Policy's get_action method.
"""
self.trainer_metrics.start_experience_collection_timer()
if take_action_outputs:
self.stats["Policy/Entropy"].append(
take_action_outputs["entropy"].mean())
self.stats["Policy/Learning Rate"].append(
take_action_outputs["learning_rate"])
for name, signal in self.policy.reward_signals.items():
self.stats[signal.value_name].append(
np.mean(take_action_outputs["value"][name]))
curr_info = curr_all_info[self.brain_name]
next_info = next_all_info[self.brain_name]
for agent_id in curr_info.agents:
self.training_buffer[agent_id].last_brain_info = curr_info
self.training_buffer[
agent_id].last_take_action_outputs = take_action_outputs
if curr_info.agents != next_info.agents:
curr_to_use = self.construct_curr_info(next_info)
else:
curr_to_use = curr_info
tmp_rewards_dict = {}
for name, signal in self.policy.reward_signals.items():
tmp_rewards_dict[name] = signal.evaluate(curr_to_use, next_info)
for agent_id in next_info.agents:
stored_info = self.training_buffer[agent_id].last_brain_info
stored_take_action_outputs = self.training_buffer[
agent_id].last_take_action_outputs
if stored_info is not None:
idx = stored_info.agents.index(agent_id)
next_idx = next_info.agents.index(agent_id)
if not stored_info.local_done[idx]:
for i, _ in enumerate(stored_info.visual_observations):
self.training_buffer[agent_id][
"visual_obs%d" % i].append(
stored_info.visual_observations[i][idx])
self.training_buffer[agent_id][
"next_visual_obs%d" % i].append(
next_info.visual_observations[i][next_idx])
if self.policy.use_vec_obs:
self.training_buffer[agent_id]["vector_obs"].append(
stored_info.vector_observations[idx])
self.training_buffer[agent_id][
"next_vector_in"].append(
next_info.vector_observations[next_idx])
if self.policy.use_recurrent:
if stored_info.memories.shape[1] == 0:
stored_info.memories = np.zeros(
(len(stored_info.agents), self.policy.m_size))
self.training_buffer[agent_id]["memory"].append(
stored_info.memories[idx])
actions = stored_take_action_outputs["action"]
if self.policy.use_continuous_act:
actions_pre = stored_take_action_outputs["pre_action"]
self.training_buffer[agent_id]["actions_pre"].append(
actions_pre[idx])
epsilons = stored_take_action_outputs[
"random_normal_epsilon"]
self.training_buffer[agent_id][
"random_normal_epsilon"].append(epsilons[idx])
else:
self.training_buffer[agent_id]["action_mask"].append(
stored_info.action_masks[idx], padding_value=1)
a_dist = stored_take_action_outputs["log_probs"]
# value is a dictionary from name of reward to value estimate of the value head
value = stored_take_action_outputs["value"]
self.training_buffer[agent_id]["actions"].append(
actions[idx])
self.training_buffer[agent_id]["prev_action"].append(
stored_info.previous_vector_actions[idx])
self.training_buffer[agent_id]["masks"].append(1.0)
self.training_buffer[agent_id]["done"].append(
next_info.local_done[next_idx])
for name, reward_result in tmp_rewards_dict.items():
# 0 because we use the scaled reward to train the agent
self.training_buffer[agent_id]["{}_rewards".format(
name)].append(
reward_result.scaled_reward[next_idx])
self.training_buffer[agent_id][
"{}_value_estimates".format(name)].append(
value[name][idx][0])
self.training_buffer[agent_id]["action_probs"].append(
a_dist[idx])
for name, rewards in self.collected_rewards.items():
if agent_id not in rewards:
rewards[agent_id] = 0
if name == "environment":
# Report the reward from the environment
rewards[agent_id] += np.array(
next_info.rewards)[next_idx]
else:
# Report the reward signals
rewards[agent_id] += tmp_rewards_dict[
name].scaled_reward[next_idx]
if not next_info.local_done[next_idx]:
if agent_id not in self.episode_steps:
self.episode_steps[agent_id] = 0
self.episode_steps[agent_id] += 1
self.trainer_metrics.end_experience_collection_timer()
def process_experiences(self, current_info: AllBrainInfo,
new_info: AllBrainInfo) -> None:
"""
Checks agent histories for processing condition, and processes them as necessary.
Processing involves calculating value and advantage targets for model updating step.
:param current_info: Dictionary of all current brains and corresponding BrainInfo.
:param new_info: Dictionary of all next brains and corresponding BrainInfo.
"""
info = new_info[self.brain_name]
for l in range(len(info.agents)):
agent_actions = self.training_buffer[info.agents[l]]["actions"]
if (info.local_done[l] or len(agent_actions) >
self.trainer_parameters["time_horizon"]
) and len(agent_actions) > 0:
agent_id = info.agents[l]
if info.max_reached[l]:
bootstrapping_info = self.training_buffer[
agent_id].last_brain_info
idx = bootstrapping_info.agents.index(agent_id)
else:
bootstrapping_info = info
idx = l
value_next = self.policy.get_value_estimates(
bootstrapping_info,
idx,
info.local_done[l] and not info.max_reached[l],
)
tmp_advantages = []
tmp_returns = []
for name in self.policy.reward_signals:
bootstrap_value = value_next[name]
local_rewards = self.training_buffer[agent_id][
"{}_rewards".format(name)].get_batch()
local_value_estimates = self.training_buffer[agent_id][
"{}_value_estimates".format(name)].get_batch()
local_advantage = get_gae(
rewards=local_rewards,
value_estimates=local_value_estimates,
value_next=bootstrap_value,
gamma=self.policy.reward_signals[name].gamma,
lambd=self.trainer_parameters["lambd"],
)
local_return = local_advantage + local_value_estimates
# This is later use as target for the different value estimates
self.training_buffer[agent_id]["{}_returns".format(
name)].set(local_return)
self.training_buffer[agent_id]["{}_advantage".format(
name)].set(local_advantage)
tmp_advantages.append(local_advantage)
tmp_returns.append(local_return)
global_advantages = list(
np.mean(np.array(tmp_advantages), axis=0))
global_returns = list(np.mean(np.array(tmp_returns), axis=0))
self.training_buffer[agent_id]["advantages"].set(
global_advantages)
self.training_buffer[agent_id]["discounted_returns"].set(
global_returns)
self.training_buffer.append_update_buffer(
agent_id,
batch_size=None,
training_length=self.policy.sequence_length,
)
self.training_buffer[agent_id].reset_agent()
if info.local_done[l]:
self.stats["Environment/Episode Length"].append(
self.episode_steps.get(agent_id, 0))
self.episode_steps[agent_id] = 0
for name, rewards in self.collected_rewards.items():
if name == "environment":
self.cumulative_returns_since_policy_update.append(
rewards.get(agent_id, 0))
self.stats["Environment/Cumulative Reward"].append(
rewards.get(agent_id, 0))
self.reward_buffer.appendleft(
rewards.get(agent_id, 0))
rewards[agent_id] = 0
else:
self.stats[self.policy.reward_signals[name].
stat_name].append(
rewards.get(agent_id, 0))
rewards[agent_id] = 0
def end_episode(self):
"""
A signal that the Episode has ended. The buffer must be reset.
Get only called when the academy resets.
"""
self.training_buffer.reset_local_buffers()
for agent_id in self.episode_steps:
self.episode_steps[agent_id] = 0
for rewards in self.collected_rewards.values():
for agent_id in rewards:
rewards[agent_id] = 0
def is_ready_update(self):
"""
Returns whether or not the trainer has enough elements to run update model
:return: A boolean corresponding to whether or not update_model() can be run
"""
size_of_buffer = len(self.training_buffer.update_buffer["actions"])
return size_of_buffer > max(
int(self.trainer_parameters["buffer_size"] /
self.policy.sequence_length), 1)
def update_policy(self):
"""
Uses demonstration_buffer to update the policy.
The reward signal generators must be updated in this method at their own pace.
"""
self.trainer_metrics.start_policy_update_timer(
number_experiences=len(
self.training_buffer.update_buffer["actions"]),
mean_return=float(
np.mean(self.cumulative_returns_since_policy_update)),
)
self.cumulative_returns_since_policy_update = []
n_sequences = max(
int(self.trainer_parameters["batch_size"] /
self.policy.sequence_length), 1)
value_total, policy_total = [], []
advantages = self.training_buffer.update_buffer[
"advantages"].get_batch()
self.training_buffer.update_buffer["advantages"].set(
(advantages - advantages.mean()) / (advantages.std() + 1e-10))
num_epoch = self.trainer_parameters["num_epoch"]
for _ in range(num_epoch):
self.training_buffer.update_buffer.shuffle()
buffer = self.training_buffer.update_buffer
for l in range(
len(self.training_buffer.update_buffer["actions"]) //
n_sequences):
start = l * n_sequences
end = (l + 1) * n_sequences
run_out = self.policy.update(
buffer.make_mini_batch(start, end), n_sequences)
value_total.append(run_out["value_loss"])
policy_total.append(np.abs(run_out["policy_loss"]))
self.stats["Losses/Value Loss"].append(np.mean(value_total))
self.stats["Losses/Policy Loss"].append(np.mean(policy_total))
for _, reward_signal in self.policy.reward_signals.items():
update_stats = reward_signal.update(
self.training_buffer.update_buffer, n_sequences)
for stat, val in update_stats.items():
self.stats[stat].append(val)
if self.policy.bc_module:
update_stats = self.policy.bc_module.update()
for stat, val in update_stats.items():
self.stats[stat].append(val)
self.training_buffer.reset_update_buffer()
self.trainer_metrics.end_policy_update()
