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_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_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 = 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)
