def advance(self, env: EnvManager) -> int:
with hierarchical_timer("env_step"):
new_step_infos = env.step()
for step_info in new_step_infos:
for brain_name, trainer in self.trainers.items():
if step_info.has_actions_for_brain(brain_name):
_processor = self.managers[brain_name].processor
_processor.add_experiences(
step_info.previous_all_brain_info[brain_name],
step_info.current_all_brain_info[brain_name],
step_info.brain_name_to_action_info[brain_name].
outputs,
)
for brain_name, trainer in self.trainers.items():
if self.train_model and trainer.get_step <= trainer.get_max_steps:
trainer.increment_step(len(new_step_infos))
if trainer.is_ready_update():
# Perform gradient descent with experience buffer
with hierarchical_timer("update_policy"):
trainer.update_policy()
env.set_policy(brain_name, trainer.policy)
else:
# Avoid memory leak during inference
# Eventually this whole block will take place in advance()
# But currently this only calls clear_update_buffer() in RLTrainer
# and nothing in the base class
trainer.advance()
return len(new_step_infos)
def _create_trainer_and_manager(self, env_manager: EnvManager,
name_behavior_id: str) -> None:
brain_name = BehaviorIdentifiers.from_name_behavior_id(
name_behavior_id).brain_name
try:
trainer = self.trainers[brain_name]
except KeyError:
trainer = self.trainer_factory.generate(brain_name)
self.trainers[brain_name] = trainer
self.logger.info(trainer)
if self.train_model:
trainer.write_tensorboard_text("Hyperparameters",
trainer.parameters)
# print("*/*/*/*/*/*/*/*/*/*/*/*/*")
# print(trainer)
# print("*/*/*/*/*/*/*/*/*/*/*/*/*")
policy = trainer.create_policy(
env_manager.external_brains[name_behavior_id])
trainer.add_policy(name_behavior_id, policy)
agent_manager = AgentManager(
policy,
name_behavior_id,
trainer.stats_reporter,
trainer.parameters.get("time_horizon", sys.maxsize),
)
env_manager.set_agent_manager(name_behavior_id, agent_manager)
env_manager.set_policy(name_behavior_id, policy)
self.brain_name_to_identifier[brain_name].add(name_behavior_id)
trainer.publish_policy_queue(agent_manager.policy_queue)
trainer.subscribe_trajectory_queue(agent_manager.trajectory_queue)
def start_trainer(self, trainer: Trainer, env_manager: EnvManager) -> None:
self.trainers[trainer.brain_name] = trainer
self.logger.info(trainer)
if self.train_model:
trainer.write_tensorboard_text("Hyperparameters",
trainer.parameters)
env_manager.set_policy(trainer.brain_name, trainer.policy)
def _create_trainer_and_manager(self, env_manager: EnvManager,
name_behavior_id: str) -> None:
parsed_behavior_id = BehaviorIdentifiers.from_name_behavior_id(
name_behavior_id)
brain_name = parsed_behavior_id.brain_name
try:
trainer = self.trainers[brain_name]
except KeyError:
trainer = self.trainer_factory.generate(brain_name)
self.trainers[brain_name] = trainer
policy = trainer.create_policy(
parsed_behavior_id, env_manager.external_brains[name_behavior_id])
trainer.add_policy(parsed_behavior_id, policy)
agent_manager = AgentManager(
policy,
name_behavior_id,
trainer.stats_reporter,
trainer.parameters.get("time_horizon", sys.maxsize),
)
env_manager.set_agent_manager(name_behavior_id, agent_manager)
env_manager.set_policy(name_behavior_id, policy)
self.brain_name_to_identifier[brain_name].add(name_behavior_id)
trainer.publish_policy_queue(agent_manager.policy_queue)
trainer.subscribe_trajectory_queue(agent_manager.trajectory_queue)
if trainer.threaded:
# Start trainer thread
trainerthread = threading.Thread(target=self.trainer_update_func,
args=(trainer, ),
daemon=True)
trainerthread.start()
self.trainer_threads.append(trainerthread)
def _get_and_process_experiences(self, env: EnvManager) -> int:
with hierarchical_timer("env_step"):
# Get new policies if found
for brain_name in self.trainers.keys():
for name_behavior_id in self.brain_name_to_identifier[
brain_name]:
try:
_policy = self.managers[
name_behavior_id].policy_queue.get_nowait()
env.set_policy(name_behavior_id, _policy)
except AgentManagerQueue.Empty:
pass
# Step the environment
new_step_infos = env.step()
# Add to AgentProcessor
for step_info in new_step_infos:
for name_behavior_id in step_info.name_behavior_ids:
if name_behavior_id not in self.managers:
self.logger.warning(
"Agent manager was not created for behavior id {}.".
format(name_behavior_id))
continue
self.managers[name_behavior_id].add_experiences(
step_info.previous_all_brain_info[name_behavior_id],
step_info.current_all_brain_info[name_behavior_id],
step_info.brain_name_to_action_info[name_behavior_id].
outputs,
)
return len(new_step_infos)
def _create_trainer_and_manager(
self, env_manager: EnvManager, name_behavior_id: str
) -> None:
parsed_behavior_id = BehaviorIdentifiers.from_name_behavior_id(name_behavior_id)
brain_name = parsed_behavior_id.brain_name
try:
trainer = self.trainers[brain_name]
except KeyError:
trainer = self.trainer_factory.generate(brain_name)
self.trainers[brain_name] = trainer
policy = trainer.create_policy(
parsed_behavior_id, env_manager.external_brains[name_behavior_id]
)
trainer.add_policy(parsed_behavior_id, policy)
agent_manager = AgentManager(
policy,
name_behavior_id,
trainer.stats_reporter,
trainer.parameters.get("time_horizon", sys.maxsize),
)
env_manager.set_agent_manager(name_behavior_id, agent_manager)
env_manager.set_policy(name_behavior_id, policy)
self.brain_name_to_identifier[brain_name].add(name_behavior_id)
trainer.publish_policy_queue(agent_manager.policy_queue)
trainer.subscribe_trajectory_queue(agent_manager.trajectory_queue)
def reset_env_if_ready(self, env: EnvManager) -> None:
# Get the sizes of the reward buffers.
reward_buff = {
k: list(t.reward_buffer)
for (k, t) in self.trainers.items()
}
curr_step = {k: int(t.step) for (k, t) in self.trainers.items()}
max_step = {
k: int(t.get_max_steps)
for (k, t) in self.trainers.items()
}
# Attempt to increment the lessons of the brains who
# were ready.
updated, param_must_reset = self.param_manager.update_lessons(
curr_step, max_step, reward_buff)
if updated:
for trainer in self.trainers.values():
trainer.reward_buffer.clear()
# If ghost trainer swapped teams
ghost_controller_reset = self.ghost_controller.should_reset()
if param_must_reset or ghost_controller_reset:
self._reset_env(env) # This reset also sends the new config to env
self.end_trainer_episodes()
elif updated:
env.set_env_parameters(self.param_manager.get_current_samplers())
def _reset_env(self, env: EnvManager) -> None:
"""Resets the environment.
Returns:
A Data structure corresponding to the initial reset state of the
environment.
"""
new_config = self.param_manager.get_current_samplers()
env.reset(config=new_config)
def _reset_env(self, env: EnvManager) -> None:
"""Resets the environment.
Returns:
A Data structure corresponding to the initial reset state of the
environment.
"""
new_meta_curriculum_config = (self.meta_curriculum.get_config()
if self.meta_curriculum else None)
env.reset(config=new_meta_curriculum_config)
def _reset_env(self, env_manager: EnvManager) -> None:
"""Resets the environment.
Returns:
A Data structure corresponding to the initial reset state of the
environment.
"""
new_config = self.param_manager.get_current_samplers()
env_manager.reset(config=new_config)
# Register any new behavior ids that were generated on the reset.
self._register_new_behaviors(env_manager, env_manager.first_step_infos)
def _reset_env(self, env_manager: EnvManager) -> None:
"""Resets the environment.
Returns:
A Data structure corresponding to the initial reset state of the
environment.
"""
sampled_reset_param = self.sampler_manager.sample_all()
new_meta_curriculum_config = (self.meta_curriculum.get_config()
if self.meta_curriculum else {})
sampled_reset_param.update(new_meta_curriculum_config)
env_manager.reset(config=sampled_reset_param)
def _get_and_process_experiences(self, env: EnvManager) -> int:
with hierarchical_timer("env_step"):
# Get new policies if found
for brain_name in self.trainers.keys():
for name_behavior_id in self.brain_name_to_identifier[
brain_name]:
try:
_policy = self.managers[
name_behavior_id].policy_queue.get_nowait()
env.set_policy(name_behavior_id, _policy)
except AgentManagerQueue.Empty:
pass
# Step the environment
new_step_infos = env.step()
# Add to AgentProcessor
num_step_infos = self._process_step_infos(new_step_infos)
return num_step_infos
def _create_trainer_and_manager(self, env_manager: EnvManager,
name_behavior_id: str) -> None:
parsed_behavior_id = BehaviorIdentifiers.from_name_behavior_id(
name_behavior_id)
brain_name = parsed_behavior_id.brain_name
trainerthread = None
if brain_name in self.trainers:
trainer = self.trainers[brain_name]
else:
trainer = self.trainer_factory.generate(brain_name)
self.trainers[brain_name] = trainer
if trainer.threaded:
# Only create trainer thread for new trainers
trainerthread = threading.Thread(
target=self.trainer_update_func,
args=(trainer, ),
daemon=True)
self.trainer_threads.append(trainerthread)
policy = trainer.create_policy(
parsed_behavior_id,
env_manager.training_behaviors[name_behavior_id],
create_graph=True,
)
trainer.add_policy(parsed_behavior_id, policy)
agent_manager = AgentManager(
policy,
name_behavior_id,
trainer.stats_reporter,
trainer.parameters.time_horizon,
threaded=trainer.threaded,
)
env_manager.set_agent_manager(name_behavior_id, agent_manager)
env_manager.set_policy(name_behavior_id, policy)
self.brain_name_to_identifier[brain_name].add(name_behavior_id)
trainer.publish_policy_queue(agent_manager.policy_queue)
trainer.subscribe_trajectory_queue(agent_manager.trajectory_queue)
# Only start new trainers
if trainerthread is not None:
trainerthread.start()
def advance(self, env: EnvManager) -> int:
with hierarchical_timer("env_step"):
new_step_infos = env.step()
for step_info in new_step_infos:
for name_behavior_id in step_info.name_behavior_ids:
if name_behavior_id not in self.managers:
self.logger.warning(
"Agent manager was not created for behavior id {}.".
format(name_behavior_id))
continue
_processor = self.managers[name_behavior_id].processor
_processor.add_experiences(
step_info.previous_all_brain_info[name_behavior_id],
step_info.current_all_brain_info[name_behavior_id],
step_info.brain_name_to_action_info[name_behavior_id].
outputs,
)
for brain_name, trainer in self.trainers.items():
if self.train_model and trainer.get_step <= trainer.get_max_steps:
n_steps = len(new_step_infos)
trainer.increment_step(n_steps)
for name_behavior_id in self.brain_name_to_identifier[
brain_name]:
trainer.get_policy(name_behavior_id).increment_step(
n_steps)
if trainer.is_ready_update():
# Perform gradient descent with experience buffer
with hierarchical_timer("update_policy"):
trainer.update_policy()
for name_behavior_id in self.brain_name_to_identifier[
brain_name]:
env.set_policy(name_behavior_id,
trainer.get_policy(name_behavior_id))
else:
# Avoid memory leak during inference
# Eventually this whole block will take place in advance()
# But currently this only calls clear_update_buffer() in RLTrainer
# and nothing in the base class
trainer.advance()
return len(new_step_infos)
def advance(self, env_manager: EnvManager) -> int:
# Get steps
with hierarchical_timer("env_step"):
new_step_infos = env_manager.get_steps()
self._register_new_behaviors(env_manager, new_step_infos)
num_steps = env_manager.process_steps(new_step_infos)
# Report current lesson for each environment parameter
for (
param_name,
lesson_number,
) in self.param_manager.get_current_lesson_number().items():
for trainer in self.trainers.values():
trainer.stats_reporter.set_stat(
f"Environment/Lesson Number/{param_name}", lesson_number)
for trainer in self.trainers.values():
if not trainer.threaded:
with hierarchical_timer("trainer_advance"):
trainer.advance()
return num_steps
def advance(self, env: EnvManager) -> int:
with hierarchical_timer("env_step"):
time_start_step = time()
new_step_infos = env.step()
delta_time_step = time() - time_start_step
for step_info in new_step_infos:
for brain_name, trainer in self.trainers.items():
if brain_name in self.trainer_metrics:
self.trainer_metrics[brain_name].add_delta_step(
delta_time_step)
if step_info.has_actions_for_brain(brain_name):
trainer.add_experiences(
step_info.previous_all_brain_info[brain_name],
step_info.current_all_brain_info[brain_name],
step_info.brain_name_to_action_info[brain_name].
outputs,
)
trainer.process_experiences(
step_info.previous_all_brain_info[brain_name],
step_info.current_all_brain_info[brain_name],
)
for brain_name, trainer in self.trainers.items():
if brain_name in self.trainer_metrics:
self.trainer_metrics[brain_name].add_delta_step(
delta_time_step)
if self.train_model and trainer.get_step <= trainer.get_max_steps:
trainer.increment_step(len(new_step_infos))
if trainer.is_ready_update():
# Perform gradient descent with experience buffer
with hierarchical_timer("update_policy"):
trainer.update_policy()
env.set_policy(brain_name, trainer.policy)
else:
# Avoid memory leak during inference
trainer.clear_update_buffer()
return len(new_step_infos)
def advance(self, env: EnvManager) -> int:
# Get steps
with hierarchical_timer("env_step"):
num_steps = env.advance()
# Report current lesson
if self.meta_curriculum:
for brain_name, curr in self.meta_curriculum.brains_to_curricula.items(
):
if brain_name in self.trainers:
self.trainers[brain_name].stats_reporter.set_stat(
"Environment/Lesson", curr.lesson_num)
# Advance trainers. This can be done in a separate loop in the future.
with hierarchical_timer("trainer_advance"):
for trainer in self.trainers.values():
trainer.advance()
return num_steps
def advance(self, env: EnvManager) -> int:
# Get steps
with hierarchical_timer("env_step"):
num_steps = env.advance()
# Report current lesson
if self.meta_curriculum:
for brain_name, curr in self.meta_curriculum.brains_to_curricula.items(
):
if brain_name in self.trainers:
self.trainers[brain_name].stats_reporter.set_stat(
"Environment/Lesson", curr.lesson_num)
for trainer in self.trainers.values():
if not trainer.threaded:
with hierarchical_timer("trainer_advance"):
trainer.advance()
return num_steps
def start_learning(self, env_manager: EnvManager) -> None:
self._create_model_path(self.model_path)
tf.reset_default_graph()
global_step = 0
last_brain_behavior_ids: Set[str] = set()
try:
self._reset_env(env_manager)
while self._not_done_training():
external_brain_behavior_ids = set(
env_manager.external_brains.keys())
new_behavior_ids = external_brain_behavior_ids - last_brain_behavior_ids
for name_behavior_id in new_behavior_ids:
try:
brain_name, _ = name_behavior_id.split("?")
except ValueError:
brain_name = name_behavior_id
try:
trainer = self.trainers[brain_name]
except KeyError:
trainer = self.trainer_factory.generate(brain_name)
self.trainers[brain_name] = trainer
self.logger.info(trainer)
if self.train_model:
trainer.write_tensorboard_text(
"Hyperparameters", trainer.parameters)
policy = trainer.create_policy(
env_manager.external_brains[name_behavior_id])
trainer.add_policy(name_behavior_id, policy)
env_manager.set_policy(name_behavior_id, policy)
self.brain_name_to_identifier[brain_name].add(
name_behavior_id)
agent_manager = AgentManager(processor=AgentProcessor(
trainer,
policy,
name_behavior_id,
trainer.stats_reporter,
trainer.parameters.get("time_horizon", sys.maxsize),
))
self.managers[name_behavior_id] = agent_manager
last_brain_behavior_ids = external_brain_behavior_ids
n_steps = self.advance(env_manager)
for _ in range(n_steps):
global_step += 1
self.reset_env_if_ready(env_manager, global_step)
if self._should_save_model(global_step):
# Save Tensorflow model
self._save_model()
self.write_to_tensorboard(global_step)
# Final save Tensorflow model
if global_step != 0 and self.train_model:
self._save_model()
except (KeyboardInterrupt, UnityCommunicationException):
if self.train_model:
self._save_model_when_interrupted()
pass
if self.train_model:
self._export_graph()
self._write_timing_tree()
