def __init__(
self,
brain_name: str,
trainer_parameters: dict,
training: bool,
run_id: str,
reward_buff_cap: int = 1,
):
"""
Responsible for collecting experiences and training a neural network model.
:BrainParameters brain: Brain to be trained.
:dict trainer_parameters: The parameters for the trainer (dictionary).
:bool training: Whether the trainer is set for training.
:str run_id: The identifier of the current run
:int reward_buff_cap:
"""
self.param_keys: List[str] = []
self.brain_name = brain_name
self.run_id = run_id
self.trainer_parameters = trainer_parameters
self.summary_path = trainer_parameters["summary_path"]
self.stats_reporter = StatsReporter(self.summary_path)
self.cumulative_returns_since_policy_update: List[float] = []
self.is_training = training
self._reward_buffer: Deque[float] = deque(maxlen=reward_buff_cap)
self.policy_queues: List[AgentManagerQueue[Policy]] = []
self.trajectory_queues: List[AgentManagerQueue[Trajectory]] = []
self.step: int = 0
self.training_start_time = time.time()
self.summary_freq = self.trainer_parameters["summary_freq"]
self.next_summary_step = self.summary_freq
def __init__(
self,
brain: BrainParameters,
trainer_parameters: dict,
training: bool,
run_id: str,
reward_buff_cap: int = 1,
):
"""
Responsible for collecting experiences and training a neural network model.
:BrainParameters brain: Brain to be trained.
:dict trainer_parameters: The parameters for the trainer (dictionary).
:bool training: Whether the trainer is set for training.
:str run_id: The identifier of the current run
:int reward_buff_cap:
"""
self.param_keys: List[str] = []
self.brain_name = brain.brain_name
self.run_id = run_id
self.trainer_parameters = trainer_parameters
self.summary_path = trainer_parameters["summary_path"]
self.stats_reporter = StatsReporter(self.summary_path)
self.cumulative_returns_since_policy_update: List[float] = []
self.is_training = training
self._reward_buffer: Deque[float] = deque(maxlen=reward_buff_cap)
self.policy: TFPolicy = None # type: ignore # this will always get set
self.step: int = 0
def _check_environment_trains(
env,
trainer_config,
reward_processor=default_reward_processor,
meta_curriculum=None,
success_threshold=0.9,
env_manager=None,
):
# Create controller and begin training.
with tempfile.TemporaryDirectory() as dir:
run_id = "id"
save_freq = 99999
seed = 1337
StatsReporter.writers.clear(
) # Clear StatsReporters so we don't write to file
debug_writer = DebugWriter()
StatsReporter.add_writer(debug_writer)
# Make sure threading is turned off for determinism
trainer_config["threading"] = False
if env_manager is None:
env_manager = SimpleEnvManager(env, FloatPropertiesChannel())
trainer_factory = TrainerFactory(
trainer_config=trainer_config,
summaries_dir=dir,
run_id=run_id,
model_path=dir,
keep_checkpoints=1,
train_model=True,
load_model=False,
seed=seed,
meta_curriculum=meta_curriculum,
multi_gpu=False,
)
tc = TrainerController(
trainer_factory=trainer_factory,
summaries_dir=dir,
model_path=dir,
run_id=run_id,
meta_curriculum=meta_curriculum,
train=True,
training_seed=seed,
sampler_manager=SamplerManager(None),
resampling_interval=None,
save_freq=save_freq,
)
# Begin training
tc.start_learning(env_manager)
if (success_threshold is not None
): # For tests where we are just checking setup and not reward
processed_rewards = [
reward_processor(rewards)
for rewards in env.final_rewards.values()
]
assert all(not math.isnan(reward) for reward in processed_rewards)
assert all(reward > success_threshold
for reward in processed_rewards)
def check_environment_trains(
env,
trainer_config,
reward_processor=default_reward_processor,
env_parameter_manager=None,
success_threshold=0.9,
env_manager=None,
training_seed=None,
):
if env_parameter_manager is None:
env_parameter_manager = EnvironmentParameterManager()
# Create controller and begin training.
with tempfile.TemporaryDirectory() as dir:
run_id = "id"
seed = 1337 if training_seed is None else training_seed
StatsReporter.writers.clear(
) # Clear StatsReporters so we don't write to file
debug_writer = DebugWriter()
StatsReporter.add_writer(debug_writer)
if env_manager is None:
env_manager = SimpleEnvManager(env, EnvironmentParametersChannel())
trainer_factory = TrainerFactory(
trainer_config=trainer_config,
output_path=dir,
train_model=True,
load_model=False,
seed=seed,
param_manager=env_parameter_manager,
multi_gpu=False,
)
tc = TrainerController(
trainer_factory=trainer_factory,
output_path=dir,
run_id=run_id,
param_manager=env_parameter_manager,
train=True,
training_seed=seed,
)
# Begin training
tc.start_learning(env_manager)
if (success_threshold is not None
): # For tests where we are just checking setup and not reward
processed_rewards = [
reward_processor(rewards)
for rewards in env.final_rewards.values()
]
assert all(not math.isnan(reward) for reward in processed_rewards)
assert all(reward > success_threshold
for reward in processed_rewards)
def test_agentprocessor(num_vis_obs):
policy = create_mock_policy()
trainer = mock.Mock()
processor = AgentProcessor(
trainer,
policy,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
fake_action_outputs = {
"action": [0.1, 0.1],
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"pre_action": [0.1, 0.1],
"log_probs": [0.1, 0.1],
}
mock_braininfo = mb.create_mock_braininfo(
num_agents=2,
num_vector_observations=8,
num_vector_acts=2,
num_vis_observations=num_vis_obs,
)
for i in range(5):
processor.add_experiences(mock_braininfo, mock_braininfo,
fake_action_outputs)
# Assert that two trajectories have been added to the Trainer
assert len(trainer.process_trajectory.call_args_list) == 2
# Assert that the trajectory is of length 5
trajectory = trainer.process_trajectory.call_args_list[0][0][0]
assert len(trajectory.steps) == 5
# Assert that the AgentProcessor is empty
assert len(processor.experience_buffers[0]) == 0
def test_agentprocessor(num_vis_obs):
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
fake_action_outputs = {
"action": ActionTuple(continuous=np.array([[0.1], [0.1]])),
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"log_probs": LogProbsTuple(continuous=np.array([[0.1], [0.1]])),
}
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=2,
observation_shapes=[(8,)] + num_vis_obs * [(84, 84, 3)],
action_spec=ActionSpec.create_continuous(2),
)
fake_action_info = ActionInfo(
action=ActionTuple(continuous=np.array([[0.1], [0.1]])),
env_action=ActionTuple(continuous=np.array([[0.1], [0.1]])),
value=[0.1, 0.1],
outputs=fake_action_outputs,
agent_ids=mock_decision_steps.agent_id,
)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
for _ in range(5):
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, fake_action_info
)
# Assert that two trajectories have been added to the Trainer
assert len(tqueue.put.call_args_list) == 2
# Assert that the trajectory is of length 5
trajectory = tqueue.put.call_args_list[0][0][0]
assert len(trajectory.steps) == 5
# Assert that the AgentProcessor is empty
assert len(processor.experience_buffers[0]) == 0
# Test empty steps
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=0,
observation_shapes=[(8,)] + num_vis_obs * [(84, 84, 3)],
action_spec=ActionSpec.create_continuous(2),
)
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
# Assert that the AgentProcessor is still empty
assert len(processor.experience_buffers[0]) == 0
def __init__(
self,
brain_name: str,
trainer_settings: TrainerSettings,
training: bool,
load: bool,
artifact_path: str,
reward_buff_cap: int = 1,
):
"""
Responsible for collecting experiences and training a neural network model.
:param brain_name: Brain name of brain to be trained.
:param trainer_settings: The parameters for the trainer (dictionary).
:param training: Whether the trainer is set for training.
:param artifact_path: The directory within which to store artifacts from this trainer
:param reward_buff_cap:
"""
self.brain_name = brain_name
self.trainer_settings = trainer_settings
self._threaded = trainer_settings.threaded
self._stats_reporter = StatsReporter(brain_name)
self.is_training = training
self.load = load
self._reward_buffer: Deque[float] = deque(maxlen=reward_buff_cap)
self.policy_queues: List[AgentManagerQueue[Policy]] = []
self.trajectory_queues: List[AgentManagerQueue[Trajectory]] = []
self.step: int = 0
self.artifact_path = artifact_path
self.summary_freq = self.trainer_settings.summary_freq
self.policies: Dict[str, Policy] = {}
def __init__(
self,
brain_name: str,
trainer_settings: TrainerSettings,
training: bool,
run_id: str,
reward_buff_cap: int = 1,
):
"""
Responsible for collecting experiences and training a neural network model.
:BrainParameters brain: Brain to be trained.
:dict trainer_settings: The parameters for the trainer (dictionary).
:bool training: Whether the trainer is set for training.
:str run_id: The identifier of the current run
:int reward_buff_cap:
"""
self.brain_name = brain_name
self.run_id = run_id
self.trainer_settings = trainer_settings
self._threaded = trainer_settings.threaded
self._stats_reporter = StatsReporter(brain_name)
self.is_training = training
self._reward_buffer: Deque[float] = deque(maxlen=reward_buff_cap)
self.policy_queues: List[AgentManagerQueue[Policy]] = []
self.trajectory_queues: List[AgentManagerQueue[Trajectory]] = []
self.step: int = 0
self.summary_freq = self.trainer_settings.summary_freq
def test_agent_deletion():
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
fake_action_outputs = {
"action": [0.1],
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"pre_action": [0.1],
"log_probs": [0.1],
}
mock_step = mb.create_mock_batchedstep(
num_agents=1,
num_vector_observations=8,
action_shape=[2],
num_vis_observations=0,
)
mock_done_step = mb.create_mock_batchedstep(
num_agents=1,
num_vector_observations=8,
action_shape=[2],
num_vis_observations=0,
done=True,
)
fake_action_info = ActionInfo(
action=[0.1],
value=[0.1],
outputs=fake_action_outputs,
agent_ids=mock_step.agent_id,
)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(mock_step, 0, ActionInfo.empty())
# Run 3 trajectories, with different workers (to simulate different agents)
add_calls = []
remove_calls = []
for _ep in range(3):
for _ in range(5):
processor.add_experiences(mock_step, _ep, fake_action_info)
add_calls.append(mock.call([get_global_agent_id(_ep, 0)], [0.1]))
processor.add_experiences(mock_done_step, _ep, fake_action_info)
# Make sure we don't add experiences from the prior agents after the done
remove_calls.append(mock.call([get_global_agent_id(_ep, 0)]))
policy.save_previous_action.assert_has_calls(add_calls)
policy.remove_previous_action.assert_has_calls(remove_calls)
# Check that there are no experiences left
assert len(processor.experience_buffers.keys()) == 0
assert len(processor.last_take_action_outputs.keys()) == 0
assert len(processor.episode_steps.keys()) == 0
assert len(processor.episode_rewards.keys()) == 0
def test_agent_manager_stats_report(aggregation_type):
stats_reporter = StatsReporter("recorder_name")
manager = AgentManager(None, "behaviorName", stats_reporter)
values = range(5)
env_stats = {"stat": [(i, aggregation_type) for i in values]}
manager.record_environment_stats(env_stats, 0)
summary = stats_reporter.get_stats_summaries("stat")
aggregation_result = {
StatsAggregationMethod.AVERAGE: sum(values) / len(values),
StatsAggregationMethod.MOST_RECENT: values[-1],
StatsAggregationMethod.SUM: sum(values),
StatsAggregationMethod.HISTOGRAM: sum(values) / len(values),
}
assert summary.aggregated_value == aggregation_result[aggregation_type]
stats_reporter.write_stats(0)
def test_group_statuses():
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=4,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
grouped=True,
)
fake_action_info = _create_action_info(4, mock_decision_steps.agent_id)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
for _ in range(2):
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, fake_action_info
)
# Make terminal steps for some dead agents
mock_decision_steps_2, mock_terminal_steps_2 = mb.create_mock_steps(
num_agents=2,
observation_specs=create_observation_specs_with_shapes([(8,)]),
action_spec=ActionSpec.create_continuous(2),
done=True,
grouped=True,
)
processor.add_experiences(
mock_decision_steps_2, mock_terminal_steps_2, 0, fake_action_info
)
fake_action_info = _create_action_info(4, mock_decision_steps.agent_id)
for _ in range(3):
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, fake_action_info
)
# Assert that four trajectories have been added to the Trainer
assert len(tqueue.put.call_args_list) == 4
# Last trajectory should be the longest
trajectory = tqueue.put.call_args_list[0][0][-1]
# Make sure trajectory has the right Groupmate Experiences
for step in trajectory.steps[0:3]:
assert len(step.group_status) == 3
# After 2 agents has died
for step in trajectory.steps[3:]:
assert len(step.group_status) == 1
def test_agentprocessor(num_vis_obs):
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
fake_action_outputs = {
"action": [0.1, 0.1],
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"pre_action": [0.1, 0.1],
"log_probs": [0.1, 0.1],
}
mock_step = mb.create_mock_batchedstep(
num_agents=2,
num_vector_observations=8,
action_shape=[2],
num_vis_observations=num_vis_obs,
)
fake_action_info = ActionInfo(
action=[0.1, 0.1],
value=[0.1, 0.1],
outputs=fake_action_outputs,
agent_ids=mock_step.agent_id,
)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(mock_step, 0, ActionInfo.empty())
for _ in range(5):
processor.add_experiences(mock_step, 0, fake_action_info)
# Assert that two trajectories have been added to the Trainer
assert len(tqueue.put.call_args_list) == 2
# Assert that the trajectory is of length 5
trajectory = tqueue.put.call_args_list[0][0][0]
assert len(trajectory.steps) == 5
# Assert that the AgentProcessor is empty
assert len(processor.experience_buffers[0]) == 0
# Test empty BatchedStepResult
mock_step = mb.create_mock_batchedstep(
num_agents=0,
num_vector_observations=8,
action_shape=[2],
num_vis_observations=num_vis_obs,
)
processor.add_experiences(mock_step, 0, ActionInfo([], [], {}, []))
# Assert that the AgentProcessor is still empty
assert len(processor.experience_buffers[0]) == 0
def test_agentprocessor(num_vis_obs):
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=2,
observation_specs=create_observation_specs_with_shapes(
[(8,)] + num_vis_obs * [(84, 84, 3)]
),
action_spec=ActionSpec.create_continuous(2),
)
fake_action_info = _create_action_info(2, mock_decision_steps.agent_id)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
for _ in range(5):
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, fake_action_info
)
# Assert that two trajectories have been added to the Trainer
assert len(tqueue.put.call_args_list) == 2
# Assert that the trajectory is of length 5
trajectory = tqueue.put.call_args_list[0][0][0]
assert len(trajectory.steps) == 5
# Make sure ungrouped agents don't have team obs
for step in trajectory.steps:
assert len(step.group_status) == 0
# Assert that the AgentProcessor is empty
assert len(processor._experience_buffers[0]) == 0
# Test empty steps
mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
num_agents=0,
observation_specs=create_observation_specs_with_shapes(
[(8,)] + num_vis_obs * [(84, 84, 3)]
),
action_spec=ActionSpec.create_continuous(2),
)
processor.add_experiences(
mock_decision_steps, mock_terminal_steps, 0, ActionInfo.empty()
)
# Assert that the AgentProcessor is still empty
assert len(processor._experience_buffers[0]) == 0
def test_end_episode():
policy = create_mock_policy()
tqueue = mock.Mock()
name_behavior_id = "test_brain_name"
processor = AgentProcessor(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
fake_action_outputs = {
"action": ActionTuple(continuous=np.array([[0.1]])),
"entropy": np.array([1.0], dtype=np.float32),
"learning_rate": 1.0,
"log_probs": LogProbsTuple(continuous=np.array([[0.1]])),
}
mock_decision_step, mock_terminal_step = mb.create_mock_steps(
num_agents=1,
observation_shapes=[(8,)],
action_spec=ActionSpec.create_continuous(2),
)
fake_action_info = ActionInfo(
action=ActionTuple(continuous=np.array([[0.1]])),
env_action=ActionTuple(continuous=np.array([[0.1]])),
value=[0.1],
outputs=fake_action_outputs,
agent_ids=mock_decision_step.agent_id,
)
processor.publish_trajectory_queue(tqueue)
# This is like the initial state after the env reset
processor.add_experiences(
mock_decision_step, mock_terminal_step, 0, ActionInfo.empty()
)
# Run 3 trajectories, with different workers (to simulate different agents)
remove_calls = []
for _ep in range(3):
remove_calls.append(mock.call([get_global_agent_id(_ep, 0)]))
for _ in range(5):
processor.add_experiences(
mock_decision_step, mock_terminal_step, _ep, fake_action_info
)
# Make sure we don't add experiences from the prior agents after the done
# Call end episode
processor.end_episode()
# Check that we removed every agent
policy.remove_previous_action.assert_has_calls(remove_calls)
# Check that there are no experiences left
assert len(processor.experience_buffers.keys()) == 0
assert len(processor.last_take_action_outputs.keys()) == 0
assert len(processor.episode_steps.keys()) == 0
assert len(processor.episode_rewards.keys()) == 0
def test_agent_manager():
policy = create_mock_policy()
name_behavior_id = "test_brain_name"
manager = AgentManager(
policy,
name_behavior_id,
max_trajectory_length=5,
stats_reporter=StatsReporter("testcat"),
)
assert len(manager.trajectory_queues) == 1
assert isinstance(manager.trajectory_queues[0], AgentManagerQueue)
def test_agent_manager_stats():
policy = mock.Mock()
stats_reporter = StatsReporter("FakeCategory")
writer = mock.Mock()
stats_reporter.add_writer(writer)
manager = AgentManager(policy, "MyBehavior", stats_reporter)
all_env_stats = [
{
"averaged": [(1.0, StatsAggregationMethod.AVERAGE)],
"most_recent": [(2.0, StatsAggregationMethod.MOST_RECENT)],
},
{
"averaged": [(3.0, StatsAggregationMethod.AVERAGE)],
"most_recent": [(4.0, StatsAggregationMethod.MOST_RECENT)],
},
]
for env_stats in all_env_stats:
manager.record_environment_stats(env_stats, worker_id=0)
expected_stats = {
"averaged": StatsSummary(mean=2.0, std=mock.ANY, num=2),
"most_recent": StatsSummary(mean=4.0, std=0.0, num=1),
}
stats_reporter.write_stats(123)
writer.write_stats.assert_any_call("FakeCategory", expected_stats, 123)
# clean up our Mock from the global list
StatsReporter.writers.remove(writer)
def test_stat_reporter_property():
# Test add_writer
mock_writer = mock.Mock()
StatsReporter.writers.clear()
StatsReporter.add_writer(mock_writer)
assert len(StatsReporter.writers) == 1
statsreporter1 = StatsReporter("category1")
# Test add_property
statsreporter1.add_property("key", "this is a text")
mock_writer.add_property.assert_called_once_with("category1", "key",
"this is a text")
def test_stat_reporter_text():
# Test add_writer
mock_writer = mock.Mock()
StatsReporter.writers.clear()
StatsReporter.add_writer(mock_writer)
assert len(StatsReporter.writers) == 1
statsreporter1 = StatsReporter("category1")
# Test write_text
step = 10
statsreporter1.write_text("this is a text", step)
mock_writer.write_text.assert_called_once_with("category1", "this is a text", step)
def test_stat_reporter_add_summary_write():
# Test add_writer
StatsReporter.writers.clear()
mock_writer1 = mock.Mock()
mock_writer2 = mock.Mock()
StatsReporter.add_writer(mock_writer1)
StatsReporter.add_writer(mock_writer2)
assert len(StatsReporter.writers) == 2
# Test add_stats and summaries
statsreporter1 = StatsReporter("category1")
statsreporter2 = StatsReporter("category2")
for i in range(10):
statsreporter1.add_stat("key1", float(i))
statsreporter2.add_stat("key2", float(i))
statssummary1 = statsreporter1.get_stats_summaries("key1")
statssummary2 = statsreporter2.get_stats_summaries("key2")
assert statssummary1.num == 10
assert statssummary2.num == 10
assert statssummary1.mean == 4.5
assert statssummary2.mean == 4.5
assert statssummary1.std == pytest.approx(2.9, abs=0.1)
assert statssummary2.std == pytest.approx(2.9, abs=0.1)
# Test write_stats
step = 10
statsreporter1.write_stats(step)
mock_writer1.write_stats.assert_called_once_with(
"category1", {"key1": statssummary1}, step
)
mock_writer2.write_stats.assert_called_once_with(
"category1", {"key1": statssummary1}, step
)
def run_training(run_seed: int, options: RunOptions) -> None:
"""
Launches training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
model_path = f"./models/{options.run_id}"
summaries_dir = "./summaries"
port = options.base_port
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
summaries_dir,
required_fields=[
"Environment/Cumulative Reward", "Environment/Episode Length"
],
)
tb_writer = TensorboardWriter(summaries_dir)
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
if options.env_path is None:
port = 5004 # This is the in Editor Training Port
env_factory = create_environment_factory(options.env_path,
options.no_graphics, run_seed,
port, options.env_args,
options.env_id, options.n_steps)
env_manager = SubprocessEnvManager(env_factory=env_factory,
n_env=options.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum_config, env_manager, options.lesson)
sampler_manager, resampling_interval = create_sampler_manager(
options.sampler_config, run_seed)
trainer_factory = TrainerFactory(
options.trainer_config,
summaries_dir,
options.run_id,
model_path,
options.keep_checkpoints,
options.train_model,
options.load_model,
run_seed,
maybe_meta_curriculum,
options.multi_gpu,
)
# Create controller and begin training.
tc = TrainerController(trainer_factory=trainer_factory,
model_path=model_path,
summaries_dir=summaries_dir,
run_id=options.run_id,
save_freq=options.save_freq,
meta_curriculum=maybe_meta_curriculum,
train=options.train_model,
training_seed=run_seed,
sampler_manager=sampler_manager,
resampling_interval=resampling_interval,
n_steps=options.n_steps)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
def run_training(run_seed: int, options: RunOptions) -> None:
"""
Launches training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
with hierarchical_timer("run_training.setup"):
model_path = f"./models/{options.run_id}"
maybe_init_path = (
f"./models/{options.initialize_from}" if options.initialize_from else None
)
summaries_dir = "./summaries"
port = options.base_port
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
summaries_dir,
required_fields=[
"Environment/Cumulative Reward",
"Environment/Episode Length",
],
)
handle_existing_directories(
model_path, summaries_dir, options.resume, options.force, maybe_init_path
)
tb_writer = TensorboardWriter(summaries_dir, clear_past_data=not options.resume)
gauge_write = GaugeWriter()
console_writer = ConsoleWriter()
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
StatsReporter.add_writer(gauge_write)
StatsReporter.add_writer(console_writer)
if options.env_path is None:
port = UnityEnvironment.DEFAULT_EDITOR_PORT
env_factory = create_environment_factory(
options.env_path, options.no_graphics, run_seed, port, options.env_args
)
engine_config = EngineConfig(
width=options.width,
height=options.height,
quality_level=options.quality_level,
time_scale=options.time_scale,
target_frame_rate=options.target_frame_rate,
capture_frame_rate=options.capture_frame_rate,
)
env_manager = SubprocessEnvManager(env_factory, engine_config, options.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum_config, env_manager, options.lesson
)
sampler_manager, resampling_interval = create_sampler_manager(
options.sampler_config, run_seed
)
trainer_factory = TrainerFactory(
options.trainer_config,
summaries_dir,
options.run_id,
model_path,
options.keep_checkpoints,
not options.inference,
options.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
options.multi_gpu,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
model_path,
summaries_dir,
options.run_id,
options.save_freq,
maybe_meta_curriculum,
not options.inference,
run_seed,
sampler_manager,
resampling_interval,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_timing_tree(summaries_dir, options.run_id)
def run_training(run_seed: int, options: RunOptions) -> None:
"""
Launches training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
# Recognize and use docker volume if one is passed as an argument
if not options.docker_target_name:
model_path = f"./models/{options.run_id}"
summaries_dir = "./summaries"
else:
model_path = f"/{options.docker_target_name}/models/{options.run_id}"
summaries_dir = f"/{options.docker_target_name}/summaries"
port = options.base_port
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
summaries_dir,
required_fields=[
"Environment/Cumulative Reward", "Environment/Episode Length"
],
)
tb_writer = TensorboardWriter(summaries_dir)
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
if options.env_path is None:
port = UnityEnvironment.DEFAULT_EDITOR_PORT
env_factory = create_environment_factory(
options.env_path,
options.docker_target_name,
options.no_graphics,
run_seed,
port,
options.env_args,
)
engine_config = EngineConfig(
options.width,
options.height,
options.quality_level,
options.time_scale,
options.target_frame_rate,
)
env_manager = SubprocessEnvManager(env_factory, engine_config,
options.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum_config, env_manager, options.lesson)
sampler_manager, resampling_interval = create_sampler_manager(
options.sampler_config, run_seed)
trainer_factory = TrainerFactory(
options.trainer_config,
summaries_dir,
options.run_id,
model_path,
options.keep_checkpoints,
options.train_model,
options.load_model,
run_seed,
maybe_meta_curriculum,
options.multi_gpu,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
model_path,
summaries_dir,
options.run_id,
options.save_freq,
maybe_meta_curriculum,
options.train_model,
run_seed,
sampler_manager,
resampling_interval,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
class Trainer(abc.ABC):
"""This class is the base class for the mlagents_envs.trainers"""
def __init__(
self,
brain_name: str,
trainer_parameters: dict,
training: bool,
run_id: str,
reward_buff_cap: int = 1,
):
"""
Responsible for collecting experiences and training a neural network model.
:BrainParameters brain: Brain to be trained.
:dict trainer_parameters: The parameters for the trainer (dictionary).
:bool training: Whether the trainer is set for training.
:str run_id: The identifier of the current run
:int reward_buff_cap:
"""
self.param_keys: List[str] = []
self.brain_name = brain_name
self.run_id = run_id
self.trainer_parameters = trainer_parameters
self.summary_path = trainer_parameters["summary_path"]
self.stats_reporter = StatsReporter(self.summary_path)
self.cumulative_returns_since_policy_update: List[float] = []
self.is_training = training
self._reward_buffer: Deque[float] = deque(maxlen=reward_buff_cap)
self.policy_queues: List[AgentManagerQueue[Policy]] = []
self.trajectory_queues: List[AgentManagerQueue[Trajectory]] = []
self.step: int = 0
self.training_start_time = time.time()
self.summary_freq = self.trainer_parameters["summary_freq"]
self.next_summary_step = self.summary_freq
def _check_param_keys(self):
for k in self.param_keys:
if k not in self.trainer_parameters:
raise UnityTrainerException(
"The hyper-parameter {0} could not be found for the {1} trainer of "
"brain {2}.".format(k, self.__class__, self.brain_name)
)
def write_tensorboard_text(self, key: str, input_dict: Dict[str, Any]) -> None:
"""
Saves text to Tensorboard.
Note: Only works on tensorflow r1.2 or above.
:param key: The name of the text.
:param input_dict: A dictionary that will be displayed in a table on Tensorboard.
"""
try:
with tf.Session(config=tf_utils.generate_session_config()) as sess:
s_op = tf.summary.text(
key,
tf.convert_to_tensor(
([[str(x), str(input_dict[x])] for x in input_dict])
),
)
s = sess.run(s_op)
self.stats_reporter.write_text(s, self.get_step)
except Exception:
LOGGER.info("Could not write text summary for Tensorboard.")
pass
def _dict_to_str(self, param_dict: Dict[str, Any], num_tabs: int) -> str:
"""
Takes a parameter dictionary and converts it to a human-readable string.
Recurses if there are multiple levels of dict. Used to print out hyperaparameters.
param: param_dict: A Dictionary of key, value parameters.
return: A string version of this dictionary.
"""
if not isinstance(param_dict, dict):
return str(param_dict)
else:
append_newline = "\n" if num_tabs > 0 else ""
return append_newline + "\n".join(
[
"\t"
+ " " * num_tabs
+ "{0}:\t{1}".format(
x, self._dict_to_str(param_dict[x], num_tabs + 1)
)
for x in param_dict
]
)
def __str__(self) -> str:
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) -> Dict[str, Any]:
"""
Returns the trainer parameters of the trainer.
"""
return self.trainer_parameters
@property
def get_max_steps(self) -> int:
"""
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 int(float(self.trainer_parameters["max_steps"]))
@property
def get_step(self) -> int:
"""
Returns the number of steps the trainer has performed
:return: the step count of the trainer
"""
return self.step
@property
def should_still_train(self) -> bool:
"""
Returns whether or not the trainer should train. A Trainer could
stop training if it wasn't training to begin with, or if max_steps
is reached.
"""
return self.is_training and self.get_step <= self.get_max_steps
@property
def reward_buffer(self) -> Deque[float]:
"""
Returns the reward buffer. The reward buffer contains the cumulative
rewards of the most recent episodes completed by agents using this
trainer.
:return: the reward buffer.
"""
return self._reward_buffer
def _increment_step(self, n_steps: int, name_behavior_id: str) -> None:
"""
Increment the step count of the trainer
:param n_steps: number of steps to increment the step count by
"""
self.step += n_steps
self.next_summary_step = self._get_next_summary_step()
p = self.get_policy(name_behavior_id)
if p:
p.increment_step(n_steps)
def _get_next_summary_step(self) -> int:
"""
Get the next step count that should result in a summary write.
"""
return self.step + (self.summary_freq - self.step % self.summary_freq)
def save_model(self, name_behavior_id: str) -> None:
"""
Saves the model
"""
self.get_policy(name_behavior_id).save_model(self.get_step)
def export_model(self, name_behavior_id: str) -> None:
"""
Exports the model
"""
policy = self.get_policy(name_behavior_id)
settings = SerializationSettings(policy.model_path, policy.brain.brain_name)
export_policy_model(settings, policy.graph, policy.sess)
def _write_summary(self, step: int) -> None:
"""
Saves training statistics to Tensorboard.
"""
is_training = "Training." if self.should_still_train else "Not Training."
stats_summary = self.stats_reporter.get_stats_summaries(
"Environment/Cumulative Reward"
)
if stats_summary.num > 0:
LOGGER.info(
" {}: {}: Step: {}. "
"Time Elapsed: {:0.3f} s "
"Mean "
"Reward: {:0.3f}"
". Std of Reward: {:0.3f}. {}".format(
self.run_id,
self.brain_name,
step,
time.time() - self.training_start_time,
stats_summary.mean,
stats_summary.std,
is_training,
)
)
set_gauge(f"{self.brain_name}.mean_reward", stats_summary.mean)
else:
LOGGER.info(
" {}: {}: Step: {}. No episode was completed since last summary. {}".format(
self.run_id, self.brain_name, step, is_training
)
)
self.stats_reporter.write_stats(int(step))
@abc.abstractmethod
def _process_trajectory(self, trajectory: Trajectory) -> None:
"""
Takes a trajectory and processes it, putting it into the update buffer.
:param trajectory: The Trajectory tuple containing the steps to be processed.
"""
self._maybe_write_summary(self.get_step + len(trajectory.steps))
self._increment_step(len(trajectory.steps), trajectory.behavior_id)
def _maybe_write_summary(self, step_after_process: int) -> None:
"""
If processing the trajectory will make the step exceed the next summary write,
write the summary. This logic ensures summaries are written on the update step and not in between.
:param step_after_process: the step count after processing the next trajectory.
"""
if step_after_process >= self.next_summary_step and self.get_step != 0:
self._write_summary(self.next_summary_step)
@abc.abstractmethod
def end_episode(self):
"""
A signal that the Episode has ended. The buffer must be reset.
Get only called when the academy resets.
"""
pass
@abc.abstractmethod
def create_policy(self, brain_parameters: BrainParameters) -> TFPolicy:
"""
Creates policy
"""
pass
@abc.abstractmethod
def add_policy(self, name_behavior_id: str, policy: TFPolicy) -> None:
"""
Adds policy to trainer
"""
pass
@abc.abstractmethod
def get_policy(self, name_behavior_id: str) -> TFPolicy:
"""
Gets policy from trainer
"""
pass
@abc.abstractmethod
def _is_ready_update(self):
"""
Returns whether or not the trainer has enough elements to run update model
:return: A boolean corresponding to wether or not update_model() can be run
"""
return False
@abc.abstractmethod
def _update_policy(self):
"""
Uses demonstration_buffer to update model.
"""
pass
def advance(self) -> None:
"""
Steps the trainer, taking in trajectories and updates if ready.
"""
with hierarchical_timer("process_trajectory"):
for traj_queue in self.trajectory_queues:
# We grab at most the maximum length of the queue.
# This ensures that even if the queue is being filled faster than it is
# being emptied, the trajectories in the queue are on-policy.
for _ in range(traj_queue.maxlen):
try:
t = traj_queue.get_nowait()
self._process_trajectory(t)
except AgentManagerQueue.Empty:
break
if self.should_still_train:
if self._is_ready_update():
with hierarchical_timer("_update_policy"):
self._update_policy()
for q in self.policy_queues:
# Get policies that correspond to the policy queue in question
q.put(self.get_policy(q.behavior_id))
def publish_policy_queue(self, policy_queue: AgentManagerQueue[Policy]) -> None:
"""
Adds a policy queue to the list of queues to publish to when this Trainer
makes a policy update
:param queue: Policy queue to publish to.
"""
self.policy_queues.append(policy_queue)
def subscribe_trajectory_queue(
self, trajectory_queue: AgentManagerQueue[Trajectory]
) -> None:
"""
Adds a trajectory queue to the list of queues for the trainer to ingest Trajectories from.
:param queue: Trajectory queue to publish to.
"""
self.trajectory_queues.append(trajectory_queue)
def run_training(run_seed: int, options: RunOptions) -> None:
"""
Launches training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
with hierarchical_timer("run_training.setup"):
checkpoint_settings = options.checkpoint_settings
env_settings = options.env_settings
engine_settings = options.engine_settings
base_path = "results"
write_path = os.path.join(base_path, checkpoint_settings.run_id)
maybe_init_path = (
os.path.join(base_path, checkpoint_settings.initialize_from)
if checkpoint_settings.initialize_from is not None
else None
)
run_logs_dir = os.path.join(write_path, "run_logs")
port: Optional[int] = env_settings.base_port
# Check if directory exists
validate_existing_directories(
write_path,
checkpoint_settings.resume,
checkpoint_settings.force,
maybe_init_path,
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Load any needed states
if checkpoint_settings.resume:
GlobalTrainingStatus.load_state(
os.path.join(run_logs_dir, "training_status.json")
)
# Configure Tensorboard Writers and StatsReporter
tb_writer = TensorboardWriter(
write_path, clear_past_data=not checkpoint_settings.resume
)
gauge_write = GaugeWriter()
console_writer = ConsoleWriter()
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(gauge_write)
StatsReporter.add_writer(console_writer)
if env_settings.env_path is None:
port = None
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(run_logs_dir), # Unity environment requires absolute path
)
engine_config = EngineConfig(
width=engine_settings.width,
height=engine_settings.height,
quality_level=engine_settings.quality_level,
time_scale=engine_settings.time_scale,
target_frame_rate=engine_settings.target_frame_rate,
capture_frame_rate=engine_settings.capture_frame_rate,
)
env_manager = SubprocessEnvManager(
env_factory, engine_config, env_settings.num_envs
)
env_parameter_manager = EnvironmentParameterManager(
options.environment_parameters, run_seed, restore=checkpoint_settings.resume
)
trainer_factory = TrainerFactory(
trainer_config=options.behaviors,
output_path=write_path,
train_model=not checkpoint_settings.inference,
load_model=checkpoint_settings.resume,
seed=run_seed,
param_manager=env_parameter_manager,
init_path=maybe_init_path,
multi_gpu=False,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
env_parameter_manager,
not checkpoint_settings.inference,
run_seed,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
def run_training(run_seed: int, options: RunOptions) -> None:
"""
Launches training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
options.checkpoint_settings.run_id = "test8"
with hierarchical_timer("run_training.setup"):
checkpoint_settings = options.checkpoint_settings
env_settings = options.env_settings
engine_settings = options.engine_settings
base_path = "results"
write_path = os.path.join(base_path, checkpoint_settings.run_id)
maybe_init_path = (os.path.join(base_path,
checkpoint_settings.initialize_from)
if checkpoint_settings.initialize_from else None)
run_logs_dir = os.path.join(write_path, "run_logs")
port: Optional[int] = env_settings.base_port
# Check if directory exists
handle_existing_directories(
write_path,
checkpoint_settings.resume,
checkpoint_settings.force,
maybe_init_path,
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Load any needed states
if checkpoint_settings.resume:
GlobalTrainingStatus.load_state(
os.path.join(run_logs_dir, "training_status.json"))
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
write_path,
required_fields=[
"Environment/Cumulative Reward",
"Environment/Episode Length",
],
)
tb_writer = TensorboardWriter(
write_path, clear_past_data=not checkpoint_settings.resume)
gauge_write = GaugeWriter()
console_writer = ConsoleWriter()
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
StatsReporter.add_writer(gauge_write)
StatsReporter.add_writer(console_writer)
engine_config = EngineConfig(
width=engine_settings.width,
height=engine_settings.height,
quality_level=engine_settings.quality_level,
time_scale=engine_settings.time_scale,
target_frame_rate=engine_settings.target_frame_rate,
capture_frame_rate=engine_settings.capture_frame_rate,
)
if env_settings.env_path is None:
port = None
# Begin training
env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mFindTarget_new/RLProject.exe"
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(
run_logs_dir), # Unity environment requires absolute path
)
env_manager = SubprocessEnvManager(env_factory, engine_config,
env_settings.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum, env_manager, restore=checkpoint_settings.resume)
sampler_manager, resampling_interval = create_sampler_manager(
options.parameter_randomization, run_seed)
max_steps = options.behaviors['Brain'].max_steps
options.behaviors['Brain'].max_steps = 10
trainer_factory = TrainerFactory(options,
write_path,
not checkpoint_settings.inference,
checkpoint_settings.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
False,
total_steps=0)
trainer_factory.trainer_config[
'Brain'].hyperparameters.learning_rate_schedule = ScheduleType.CONSTANT
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
maybe_meta_curriculum,
not checkpoint_settings.inference,
run_seed,
sampler_manager,
resampling_interval,
)
try:
# Get inital weights
tc.init_weights(env_manager)
inital_weights = deepcopy(tc.weights)
finally:
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
options.behaviors['Brain'].max_steps = max_steps
step = 0
counter = 0
max_meta_updates = 200
while counter < max_meta_updates:
sample = np.random.random_sample()
if (sample > 1):
print("Performing Meta-learning on Carry Object stage")
env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mCarryObject_new/RLProject.exe"
else:
print("Performing Meta-learning on Find Target stage")
env_settings.env_path = "C:/Users/Sebastian/Desktop/RLUnity/Training/mFindTarget_new/RLProject.exe"
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(
run_logs_dir), # Unity environment requires absolute path
)
env_manager = SubprocessEnvManager(env_factory, engine_config,
env_settings.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum,
env_manager,
restore=checkpoint_settings.resume)
sampler_manager, resampling_interval = create_sampler_manager(
options.parameter_randomization, run_seed)
trainer_factory = TrainerFactory(options,
write_path,
not checkpoint_settings.inference,
checkpoint_settings.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
False,
total_steps=step)
trainer_factory.trainer_config[
'Brain'].hyperparameters.learning_rate_schedule = ScheduleType.CONSTANT
trainer_factory.trainer_config[
'Brain'].hyperparameters.learning_rate = 0.0005 * (
1 - counter / max_meta_updates)
trainer_factory.trainer_config[
'Brain'].hyperparameters.beta = 0.005 * (
1 - counter / max_meta_updates)
trainer_factory.trainer_config[
'Brain'].hyperparameters.epsilon = 0.2 * (
1 - counter / max_meta_updates)
print("Current lr: {}\nCurrent beta: {}\nCurrent epsilon: {}".format(
trainer_factory.trainer_config['Brain'].hyperparameters.
learning_rate,
trainer_factory.trainer_config['Brain'].hyperparameters.beta,
trainer_factory.trainer_config['Brain'].hyperparameters.epsilon))
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
maybe_meta_curriculum,
not checkpoint_settings.inference,
run_seed,
sampler_manager,
resampling_interval,
)
try:
# Get inital weights
print("Start learning at step: " + str(step) + " meta_step: " +
str(counter))
print("Inital weights: " + str(inital_weights[8]))
weights_after_train = tc.start_learning(env_manager,
inital_weights)
print(tc.trainers['Brain'].optimizer)
# weights_after_train = tc.weights
# print("Trained weights: " + str(weights_after_train[8]))
step += options.behaviors['Brain'].max_steps
print("meta step:" + str(step))
# print(weights_after_train)
# equal = []
# for i, weight in enumerate(tc.weights):
# equal.append(np.array_equal(inital_weights[i], weights_after_train[i]))
# print(all(equal))
finally:
print(len(weights_after_train), len(inital_weights))
for i, weight in enumerate(weights_after_train):
inital_weights[i] = weights_after_train[i]
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
counter += 1
class Trainer(object):
"""This class is the base class for the mlagents_envs.trainers"""
def __init__(
self,
brain: BrainParameters,
trainer_parameters: dict,
training: bool,
run_id: str,
reward_buff_cap: int = 1,
):
"""
Responsible for collecting experiences and training a neural network model.
:BrainParameters brain: Brain to be trained.
:dict trainer_parameters: The parameters for the trainer (dictionary).
:bool training: Whether the trainer is set for training.
:str run_id: The identifier of the current run
:int reward_buff_cap:
"""
self.param_keys: List[str] = []
self.brain_name = brain.brain_name
self.run_id = run_id
self.trainer_parameters = trainer_parameters
self.summary_path = trainer_parameters["summary_path"]
self.stats_reporter = StatsReporter(self.summary_path)
self.cumulative_returns_since_policy_update: List[float] = []
self.is_training = training
self._reward_buffer: Deque[float] = deque(maxlen=reward_buff_cap)
self.policy: TFPolicy = None # type: ignore # this will always get set
self.step: int = 0
def check_param_keys(self):
for k in self.param_keys:
if k not in self.trainer_parameters:
raise UnityTrainerException(
"The hyper-parameter {0} could not be found for the {1} trainer of "
"brain {2}.".format(k, self.__class__, self.brain_name))
def write_tensorboard_text(self, key: str, input_dict: Dict[str,
Any]) -> None:
"""
Saves text to Tensorboard.
Note: Only works on tensorflow r1.2 or above.
:param key: The name of the text.
:param input_dict: A dictionary that will be displayed in a table on Tensorboard.
"""
try:
with tf.Session() as sess:
s_op = tf.summary.text(
key,
tf.convert_to_tensor(
([[str(x), str(input_dict[x])] for x in input_dict])),
)
s = sess.run(s_op)
self.stats_reporter.write_text(s, self.get_step)
except Exception:
LOGGER.info("Could not write text summary for Tensorboard.")
pass
def dict_to_str(self, param_dict: Dict[str, Any], num_tabs: int) -> str:
"""
Takes a parameter dictionary and converts it to a human-readable string.
Recurses if there are multiple levels of dict. Used to print out hyperaparameters.
param: param_dict: A Dictionary of key, value parameters.
return: A string version of this dictionary.
"""
if not isinstance(param_dict, dict):
return str(param_dict)
else:
append_newline = "\n" if num_tabs > 0 else ""
return append_newline + "\n".join([
"\t" + " " * num_tabs + "{0}:\t{1}".format(
x, self.dict_to_str(param_dict[x], num_tabs + 1))
for x in param_dict
])
def __str__(self) -> str:
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) -> Dict[str, Any]:
"""
Returns the trainer parameters of the trainer.
"""
return self.trainer_parameters
@property
def get_max_steps(self) -> float:
"""
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) -> int:
"""
Returns the number of steps the trainer has performed
:return: the step count of the trainer
"""
return self.step
@property
def reward_buffer(self) -> Deque[float]:
"""
Returns the reward buffer. The reward buffer contains the cumulative
rewards of the most recent episodes completed by agents using this
trainer.
:return: the reward buffer.
"""
return self._reward_buffer
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 save_model(self) -> None:
"""
Saves the model
"""
self.policy.save_model(self.get_step)
def export_model(self) -> None:
"""
Exports the model
"""
self.policy.export_model()
def write_summary(self, global_step: int,
delta_train_start: float) -> None:
"""
Saves training statistics to Tensorboard.
:param delta_train_start: Time elapsed since training started.
:param global_step: The number of steps the simulation has been going for
"""
if (global_step % self.trainer_parameters["summary_freq"] == 0
and global_step != 0):
is_training = ("Training." if self.is_training
and self.get_step <= self.get_max_steps else
"Not Training.")
step = min(self.get_step, self.get_max_steps)
stats_summary = self.stats_reporter.get_stats_summaries(
"Environment/Cumulative Reward")
if stats_summary.num > 0:
LOGGER.info(" {}: {}: Step: {}. "
"Time Elapsed: {:0.3f} s "
"Mean "
"Reward: {:0.3f}"
". Std of Reward: {:0.3f}. {}".format(
self.run_id,
self.brain_name,
step,
delta_train_start,
stats_summary.mean,
stats_summary.std,
is_training,
))
set_gauge(f"{self.brain_name}.mean_reward", stats_summary.mean)
else:
LOGGER.info(
" {}: {}: Step: {}. No episode was completed since last summary. {}"
.format(self.run_id, self.brain_name, step, is_training))
self.stats_reporter.write_stats(int(step))
def process_trajectory(self, trajectory: Trajectory) -> None:
"""
Takes a trajectory and processes it, putting it into the update buffer.
Processing involves calculating value and advantage targets for model updating step.
:param trajectory: The Trajectory tuple containing the steps to be processed.
"""
raise UnityTrainerException(
"The process_experiences method was not implemented.")
def end_episode(self):
"""
A signal that the Episode has ended. The buffer must be reset.
Get only called when the academy resets.
"""
raise UnityTrainerException(
"The end_episode method was not implemented.")
def is_ready_update(self):
"""
Returns whether or not the trainer has enough elements to run update model
:return: A boolean corresponding to wether or not update_model() can be run
"""
raise UnityTrainerException(
"The is_ready_update method was not implemented.")
def update_policy(self):
"""
Uses demonstration_buffer to update model.
"""
raise UnityTrainerException(
"The update_model method was not implemented.")
def advance(self) -> None:
pass
def run_training(sub_id: int, run_seed: int, options: CommandLineOptions,
process_queue: Queue) -> None:
"""
Launches training session.
:param process_queue: Queue used to send signal back to main.
:param sub_id: Unique id for training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
# Docker Parameters
trainer_config_path = options.trainer_config_path
curriculum_folder = options.curriculum_folder
# Recognize and use docker volume if one is passed as an argument
if not options.docker_target_name:
model_path = "./models/{run_id}-{sub_id}".format(run_id=options.run_id,
sub_id=sub_id)
summaries_dir = "./summaries"
else:
trainer_config_path = "/{docker_target_name}/{trainer_config_path}".format(
docker_target_name=options.docker_target_name,
trainer_config_path=trainer_config_path,
)
if curriculum_folder is not None:
curriculum_folder = "/{docker_target_name}/{curriculum_folder}".format(
docker_target_name=options.docker_target_name,
curriculum_folder=curriculum_folder,
)
model_path = "/{docker_target_name}/models/{run_id}-{sub_id}".format(
docker_target_name=options.docker_target_name,
run_id=options.run_id,
sub_id=sub_id,
)
summaries_dir = "/{docker_target_name}/summaries".format(
docker_target_name=options.docker_target_name)
trainer_config = load_config(trainer_config_path)
port = options.base_port + (sub_id * options.num_envs)
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
summaries_dir,
required_fields=[
"Environment/Cumulative Reward", "Environment/Episode Length"
],
)
tb_writer = TensorboardWriter(summaries_dir)
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
if options.env_path is None:
port = 5004 # This is the in Editor Training Port
env_factory = create_environment_factory(
options.env_path,
options.docker_target_name,
options.no_graphics,
run_seed,
port,
options.env_args,
)
engine_config = EngineConfig(
options.width,
options.height,
options.quality_level,
options.time_scale,
options.target_frame_rate,
)
env_manager = SubprocessEnvManager(env_factory, engine_config,
options.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(curriculum_folder,
env_manager,
options.lesson)
sampler_manager, resampling_interval = create_sampler_manager(
options.sampler_file_path, run_seed)
trainer_factory = TrainerFactory(
trainer_config,
summaries_dir,
options.run_id,
model_path,
options.keep_checkpoints,
options.train_model,
options.load_model,
run_seed,
maybe_meta_curriculum,
options.multi_gpu,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
model_path,
summaries_dir,
options.run_id + "-" + str(sub_id),
options.save_freq,
maybe_meta_curriculum,
options.train_model,
run_seed,
sampler_manager,
resampling_interval,
)
# Signal that environment has been launched.
process_queue.put(True)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
def run_training(run_seed: int, options: RunOptions, num_areas: int) -> None:
"""
Launches training session.
:param run_seed: Random seed used for training.
:param num_areas: Number of training areas to instantiate
:param options: parsed command line arguments
"""
with hierarchical_timer("run_training.setup"):
torch_utils.set_torch_config(options.torch_settings)
checkpoint_settings = options.checkpoint_settings
env_settings = options.env_settings
engine_settings = options.engine_settings
run_logs_dir = checkpoint_settings.run_logs_dir
port: Optional[int] = env_settings.base_port
# Check if directory exists
validate_existing_directories(
checkpoint_settings.write_path,
checkpoint_settings.resume,
checkpoint_settings.force,
checkpoint_settings.maybe_init_path,
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Load any needed states in case of resume
if checkpoint_settings.resume:
GlobalTrainingStatus.load_state(
os.path.join(run_logs_dir, "training_status.json")
)
# In case of initialization, set full init_path for all behaviors
elif checkpoint_settings.maybe_init_path is not None:
setup_init_path(options.behaviors, checkpoint_settings.maybe_init_path)
# Configure Tensorboard Writers and StatsReporter
stats_writers = register_stats_writer_plugins(options)
for sw in stats_writers:
StatsReporter.add_writer(sw)
if env_settings.env_path is None:
port = None
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
num_areas,
port,
env_settings.env_args,
os.path.abspath(run_logs_dir), # Unity environment requires absolute path
)
env_manager = SubprocessEnvManager(env_factory, options, env_settings.num_envs)
env_parameter_manager = EnvironmentParameterManager(
options.environment_parameters, run_seed, restore=checkpoint_settings.resume
)
trainer_factory = TrainerFactory(
trainer_config=options.behaviors,
output_path=checkpoint_settings.write_path,
train_model=not checkpoint_settings.inference,
load_model=checkpoint_settings.resume,
seed=run_seed,
param_manager=env_parameter_manager,
init_path=checkpoint_settings.maybe_init_path,
multi_gpu=False,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
checkpoint_settings.write_path,
checkpoint_settings.run_id,
env_parameter_manager,
not checkpoint_settings.inference,
run_seed,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_run_options(checkpoint_settings.write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
def run_training_aai(run_seed: int, options: RunOptionsAAI) -> None:
"""
Launches training session.
:param run_seed: Random seed used for training.
:param options: training parameters
"""
with hierarchical_timer("run_training.setup"):
# Recognize and use docker volume if one is passed as an argument
# if not options.docker_target_name:
model_path = f"./models/{options.run_id}"
summaries_dir = "./summaries"
# else:
# model_path = f"/{options.docker_target_name}/models/{options.run_id}"
# summaries_dir = f"/{options.docker_target_name}/summaries"
port = options.base_port
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
summaries_dir,
required_fields=[
"Environment/Cumulative Reward",
"Environment/Episode Length",
],
)
tb_writer = TensorboardWriter(summaries_dir)
gauge_write = GaugeWriter()
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
StatsReporter.add_writer(gauge_write)
if options.env_path is None:
port = AnimalAIEnvironment.DEFAULT_EDITOR_PORT
env_factory = create_environment_factory_aai(
options.env_path,
# options.docker_target_name,
run_seed,
port,
options.n_arenas_per_env,
options.arena_config,
options.resolution,
)
if options.train_model:
engine_config = EngineConfig(
options.width,
options.height,
AnimalAIEnvironment.QUALITY_LEVEL.train,
AnimalAIEnvironment.TIMESCALE.train,
AnimalAIEnvironment.TARGET_FRAME_RATE.train,
)
else:
engine_config = EngineConfig(
AnimalAIEnvironment.WINDOW_WIDTH.play,
AnimalAIEnvironment.WINDOW_HEIGHT.play,
AnimalAIEnvironment.QUALITY_LEVEL.play,
AnimalAIEnvironment.TIMESCALE.play,
AnimalAIEnvironment.TARGET_FRAME_RATE.play,
)
env_manager = SubprocessEnvManagerAAI(env_factory, engine_config,
options.num_envs)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum_config, env_manager, options.lesson)
trainer_factory = TrainerFactory(
options.trainer_config,
summaries_dir,
options.run_id,
model_path,
options.keep_checkpoints,
options.train_model,
options.load_model,
run_seed,
maybe_meta_curriculum,
# options.multi_gpu,
)
# Create controller and begin training.
tc = TrainerControllerAAI(
trainer_factory,
model_path,
summaries_dir,
options.run_id,
options.save_freq,
maybe_meta_curriculum,
options.train_model,
run_seed,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_timing_tree(summaries_dir, options.run_id)
def run_training(run_seed: int, options: RunOptions) -> None:
"""
Launches training session.
:param options: parsed command line arguments
:param run_seed: Random seed used for training.
:param run_options: Command line arguments for training.
"""
with hierarchical_timer("run_training.setup"):
checkpoint_settings = options.checkpoint_settings
env_settings = options.env_settings
engine_settings = options.engine_settings
base_path = "results"
write_path = os.path.join(base_path, checkpoint_settings.run_id)
maybe_init_path = (
os.path.join(base_path, checkpoint_settings.initialize_from)
if checkpoint_settings.initialize_from
else None
)
run_logs_dir = os.path.join(write_path, "run_logs")
port: Optional[int] = env_settings.base_port
# Check if directory exists
handle_existing_directories(
write_path,
checkpoint_settings.resume,
checkpoint_settings.force,
maybe_init_path,
)
# Make run logs directory
os.makedirs(run_logs_dir, exist_ok=True)
# Load any needed states
if checkpoint_settings.resume:
GlobalTrainingStatus.load_state(
os.path.join(run_logs_dir, "training_status.json")
)
# Configure CSV, Tensorboard Writers and StatsReporter
# We assume reward and episode length are needed in the CSV.
csv_writer = CSVWriter(
write_path,
required_fields=[
"Environment/Cumulative Reward",
"Environment/Episode Length",
],
)
tb_writer = TensorboardWriter(
write_path, clear_past_data=not checkpoint_settings.resume
)
gauge_write = GaugeWriter()
console_writer = ConsoleWriter()
StatsReporter.add_writer(tb_writer)
StatsReporter.add_writer(csv_writer)
StatsReporter.add_writer(gauge_write)
StatsReporter.add_writer(console_writer)
if env_settings.env_path is None:
port = None
env_factory = create_environment_factory(
env_settings.env_path,
engine_settings.no_graphics,
run_seed,
port,
env_settings.env_args,
os.path.abspath(run_logs_dir), # Unity environment requires absolute path
)
engine_config = EngineConfig(
width=engine_settings.width,
height=engine_settings.height,
quality_level=engine_settings.quality_level,
time_scale=engine_settings.time_scale,
target_frame_rate=engine_settings.target_frame_rate,
capture_frame_rate=engine_settings.capture_frame_rate,
)
env_manager = SubprocessEnvManager(
env_factory, engine_config, env_settings.num_envs
)
maybe_meta_curriculum = try_create_meta_curriculum(
options.curriculum, env_manager, restore=checkpoint_settings.resume
)
maybe_add_samplers(options.parameter_randomization, env_manager, run_seed)
trainer_factory = TrainerFactory(
options.behaviors,
write_path,
not checkpoint_settings.inference,
checkpoint_settings.resume,
run_seed,
maybe_init_path,
maybe_meta_curriculum,
False,
)
# Create controller and begin training.
tc = TrainerController(
trainer_factory,
write_path,
checkpoint_settings.run_id,
maybe_meta_curriculum,
not checkpoint_settings.inference,
run_seed,
)
# Begin training
try:
tc.start_learning(env_manager)
finally:
env_manager.close()
write_run_options(write_path, options)
write_timing_tree(run_logs_dir)
write_training_status(run_logs_dir)
