def basic_trainer_controller():
trainer_factory_mock = MagicMock()
trainer_factory_mock.ghost_controller = GhostController()
return TrainerController(
trainer_factory=trainer_factory_mock,
output_path="test_model_path",
run_id="test_run_id",
param_manager=EnvironmentParameterManager(),
train=True,
training_seed=99,
)
def test_curriculum_raises_all_completion_criteria_conversion():
with pytest.warns(TrainerConfigWarning):
run_options = RunOptions.from_dict(
yaml.safe_load(test_bad_curriculum_all_competion_criteria_config_yaml)
)
param_manager = EnvironmentParameterManager(
run_options.environment_parameters, 1337, False
)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (True, True)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (True, True)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (False, False)
assert param_manager.get_current_lesson_number() == {"param_1": 2}
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_handles_no_config_provided():
"""
Make sure the trainer setup handles no configs provided at all.
"""
brain_name = "testbrain"
no_default_config = RunOptions().behaviors
trainer_factory = trainer_util.TrainerFactory(
trainer_config=no_default_config,
output_path="output_path",
train_model=True,
load_model=False,
seed=42,
param_manager=EnvironmentParameterManager(),
)
trainer_factory.generate(brain_name)
def test_sac_trainer_update_normalization(sac_config):
behavior_id_team0 = "test_brain?team=0"
brain_name = BehaviorIdentifiers.from_name_behavior_id(
behavior_id_team0).brain_name
mock_specs = mb.setup_test_behavior_specs(True,
False,
vector_action_space=[2],
vector_obs_space=1)
base_config = sac_config.behaviors
output_path = "results_dir"
train_model = True
load_model = False
seed = 42
trainer_factory = TrainerFactory(
trainer_config=base_config,
output_path=output_path,
train_model=train_model,
load_model=load_model,
seed=seed,
param_manager=EnvironmentParameterManager(),
)
sac_trainer = trainer_factory.generate(brain_name)
parsed_behavior_id0 = BehaviorIdentifiers.from_name_behavior_id(
behavior_id_team0)
policy = sac_trainer.create_policy(parsed_behavior_id0, mock_specs)
sac_trainer.add_policy(parsed_behavior_id0, policy)
trajectory_queue0 = AgentManagerQueue(behavior_id_team0)
sac_trainer.subscribe_trajectory_queue(trajectory_queue0)
time_horizon = 15
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
observation_specs=create_observation_specs_with_shapes([(1, )]),
action_spec=mock_specs.action_spec,
)
trajectory_queue0.put(trajectory)
# mocking out update_normalization in both the policy and critic
with patch(
"mlagents.trainers.torch.networks.ValueNetwork.update_normalization"
) as optimizer_update_normalization_mock, patch(
"mlagents.trainers.policy.torch_policy.TorchPolicy.update_normalization"
) as policy_update_normalization_mock:
sac_trainer.advance()
optimizer_update_normalization_mock.assert_called_once()
policy_update_normalization_mock.assert_called_once()
def test_handles_no_config_provided():
"""
Make sure the trainer setup handles no configs provided at all.
"""
brain_name = "testbrain"
no_default_config = RunOptions().behaviors
# Pretend this was created without a YAML file
no_default_config.set_config_specified(False)
trainer_factory = TrainerFactory(
trainer_config=no_default_config,
output_path="output_path",
train_model=True,
load_model=False,
seed=42,
param_manager=EnvironmentParameterManager(),
)
trainer_factory.generate(brain_name)
def test_initialize_ppo_trainer(BehaviorSpecMock, dummy_config):
brain_name = "testbrain"
training_behaviors = {"testbrain": BehaviorSpecMock()}
output_path = "results_dir"
train_model = True
load_model = False
seed = 11
expected_reward_buff_cap = 1
base_config = dummy_config.behaviors
expected_config = ppo_dummy_config()
def mock_constructor(
self,
brain,
reward_buff_cap,
trainer_settings,
training,
load,
seed,
artifact_path,
):
assert brain == brain_name
assert trainer_settings == expected_config
assert reward_buff_cap == expected_reward_buff_cap
assert training == train_model
assert load == load_model
assert seed == seed
assert artifact_path == os.path.join(output_path, brain_name)
with patch.object(PPOTrainer, "__init__", mock_constructor):
trainer_factory = trainer_util.TrainerFactory(
trainer_config=base_config,
output_path=output_path,
train_model=train_model,
load_model=load_model,
seed=seed,
param_manager=EnvironmentParameterManager(),
)
trainers = {}
for brain_name in training_behaviors.keys():
trainers[brain_name] = trainer_factory.generate(brain_name)
assert "testbrain" in trainers
assert isinstance(trainers["testbrain"], PPOTrainer)
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 _initialize_trainer(
trainer_settings: TrainerSettings,
brain_name: str,
output_path: str,
train_model: bool,
load_model: bool,
ghost_controller: GhostController,
seed: int,
param_manager: EnvironmentParameterManager,
init_path: str = None,
multi_gpu: bool = False,
) -> Trainer:
"""
Initializes a trainer given a provided trainer configuration and brain parameters, as well as
some general training session options.
:param trainer_settings: Original trainer configuration loaded from YAML
:param brain_name: Name of the brain to be associated with trainer
:param output_path: Path to save the model and summary statistics
:param keep_checkpoints: How many model checkpoints to keep
:param train_model: Whether to train the model (vs. run inference)
:param load_model: Whether to load the model or randomly initialize
:param ghost_controller: The object that coordinates ghost trainers
:param seed: The random seed to use
:param param_manager: EnvironmentParameterManager, used to determine a reward buffer length for PPOTrainer
:param init_path: Path from which to load model, if different from model_path.
:return:
"""
trainer_artifact_path = os.path.join(output_path, brain_name)
if init_path is not None:
trainer_settings.init_path = os.path.join(init_path, brain_name)
min_lesson_length = param_manager.get_minimum_reward_buffer_size(
brain_name)
trainer: Trainer = None # type: ignore # will be set to one of these, or raise
trainer_type = trainer_settings.trainer_type
if trainer_type == TrainerType.PPO:
trainer = PPOTrainer(
brain_name,
min_lesson_length,
trainer_settings,
train_model,
load_model,
seed,
trainer_artifact_path,
)
elif trainer_type == TrainerType.SAC:
trainer = SACTrainer(
brain_name,
min_lesson_length,
trainer_settings,
train_model,
load_model,
seed,
trainer_artifact_path,
)
else:
raise TrainerConfigError(
f'The trainer config contains an unknown trainer type "{trainer_type}" for brain {brain_name}'
)
if trainer_settings.self_play is not None:
trainer = GhostTrainer(
trainer,
brain_name,
ghost_controller,
min_lesson_length,
trainer_settings,
train_model,
trainer_artifact_path,
)
return trainer
def test_create_manager():
run_options = RunOptions.from_dict(yaml.safe_load(test_everything_config_yaml))
param_manager = EnvironmentParameterManager(
run_options.environment_parameters, 1337, False
)
assert param_manager.get_minimum_reward_buffer_size("fake_behavior") == 100
assert param_manager.get_current_lesson_number() == {
"param_1": 0,
"param_2": 0,
"param_3": 0,
}
assert param_manager.get_current_samplers() == {
"param_1": ConstantSettings(seed=1337, value=1),
"param_2": GaussianSettings(seed=1337 + 3, mean=4, st_dev=5),
"param_3": ConstantSettings(seed=1337 + 3 + 1, value=20),
}
# Not enough episodes completed
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 99},
) == (False, False)
# Not enough episodes reward
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1] * 101},
) == (False, False)
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 500},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [1000] * 101},
) == (True, True)
assert param_manager.get_current_lesson_number() == {
"param_1": 1,
"param_2": 0,
"param_3": 0,
}
param_manager_2 = EnvironmentParameterManager(
run_options.environment_parameters, 1337, restore=True
)
# The use of global status should make it so that the lesson numbers are maintained
assert param_manager_2.get_current_lesson_number() == {
"param_1": 1,
"param_2": 0,
"param_3": 0,
}
# No reset required
assert param_manager.update_lessons(
trainer_steps={"fake_behavior": 700},
trainer_max_steps={"fake_behavior": 1000},
trainer_reward_buffer={"fake_behavior": [0] * 101},
) == (True, False)
assert param_manager.get_current_samplers() == {
"param_1": UniformSettings(seed=1337 + 2, min_value=1, max_value=3),
"param_2": GaussianSettings(seed=1337 + 3, mean=4, st_dev=5),
"param_3": ConstantSettings(seed=1337 + 3 + 1, value=20),
}
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 test_curriculum_no_behavior():
with pytest.raises(TypeError):
run_options = RunOptions.from_dict(
yaml.safe_load(test_curriculum_no_behavior_yaml))
EnvironmentParameterManager(run_options.environment_parameters, 1337,
False)
