def test_update_buffer_append():
trainer = create_rl_trainer()
mock_policy = mock.Mock()
trainer.add_policy("TestBrain", mock_policy)
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
time_horizon = 10
trajectory = mb.make_fake_trajectory(
length=time_horizon,
observation_specs=create_observation_specs_with_shapes([(1, )]),
max_step_complete=True,
action_spec=ActionSpec.create_discrete((2, )),
)
agentbuffer_trajectory = trajectory.to_agentbuffer()
assert trainer.update_buffer.num_experiences == 0
# Check that if we append, our update buffer gets longer.
# max_steps = 100
for i in range(10):
trainer._process_trajectory(trajectory)
trainer._append_to_update_buffer(agentbuffer_trajectory)
assert trainer.update_buffer.num_experiences == (i + 1) * time_horizon
# Check that if we append after stopping training, nothing happens.
# We process enough trajectories to hit max steps
trainer.set_is_policy_updating(False)
trainer._process_trajectory(trajectory)
trainer._append_to_update_buffer(agentbuffer_trajectory)
assert trainer.update_buffer.num_experiences == (i + 1) * time_horizon
def test_process_trajectory(dummy_config):
brain_params_team0 = BrainParameters(
brain_name="test_brain?team=0",
vector_observation_space_size=1,
camera_resolutions=[],
vector_action_space_size=[2],
vector_action_descriptions=[],
vector_action_space_type=0,
)
brain_name = BehaviorIdentifiers.from_name_behavior_id(
brain_params_team0.brain_name).brain_name
brain_params_team1 = BrainParameters(
brain_name="test_brain?team=1",
vector_observation_space_size=1,
camera_resolutions=[],
vector_action_space_size=[2],
vector_action_descriptions=[],
vector_action_space_type=0,
)
dummy_config["summary_path"] = "./summaries/test_trainer_summary"
dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
ppo_trainer = PPOTrainer(brain_name, 0, dummy_config, True, False, 0, "0",
False)
trainer = GhostTrainer(ppo_trainer, brain_name, 0, dummy_config, True, "0")
# first policy encountered becomes policy trained by wrapped PPO
policy = trainer.create_policy(brain_params_team0)
trainer.add_policy(brain_params_team0.brain_name, policy)
trajectory_queue0 = AgentManagerQueue(brain_params_team0.brain_name)
trainer.subscribe_trajectory_queue(trajectory_queue0)
# Ghost trainer should ignore this queue because off policy
policy = trainer.create_policy(brain_params_team1)
trainer.add_policy(brain_params_team1.brain_name, policy)
trajectory_queue1 = AgentManagerQueue(brain_params_team1.brain_name)
trainer.subscribe_trajectory_queue(trajectory_queue1)
time_horizon = 15
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
trajectory_queue0.put(trajectory)
trainer.advance()
# Check that trainer put trajectory in update buffer
assert trainer.trainer.update_buffer.num_experiences == 15
trajectory_queue1.put(trajectory)
trainer.advance()
# Check that ghost trainer ignored off policy queue
assert trainer.trainer.update_buffer.num_experiences == 15
# Check that it emptied the queue
assert trajectory_queue1.empty()
def test_summary_checkpoint(mock_add_checkpoint, mock_write_summary):
trainer = create_rl_trainer()
mock_policy = mock.Mock()
trainer.add_policy("TestBrain", mock_policy)
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
time_horizon = 10
summary_freq = trainer.trainer_settings.summary_freq
checkpoint_interval = trainer.trainer_settings.checkpoint_interval
trajectory = mb.make_fake_trajectory(
length=time_horizon,
observation_specs=create_observation_specs_with_shapes([(1, )]),
max_step_complete=True,
action_spec=ActionSpec.create_discrete((2, )),
)
# Check that we can turn off the trainer and that the buffer is cleared
num_trajectories = 5
for _ in range(0, num_trajectories):
trajectory_queue.put(trajectory)
trainer.advance()
# Check that there is stuff in the policy queue
policy_queue.get_nowait()
# Check that we have called write_summary the appropriate number of times
calls = [
mock.call(step) for step in range(summary_freq, num_trajectories *
time_horizon, summary_freq)
]
mock_write_summary.assert_has_calls(calls, any_order=True)
checkpoint_range = range(checkpoint_interval,
num_trajectories * time_horizon,
checkpoint_interval)
calls = [mock.call(trainer.brain_name, step) for step in checkpoint_range]
trainer.model_saver.save_checkpoint.assert_has_calls(calls, any_order=True)
export_ext = "onnx"
add_checkpoint_calls = [
mock.call(
trainer.brain_name,
ModelCheckpoint(
step,
f"{trainer.model_saver.model_path}{os.path.sep}{trainer.brain_name}-{step}.{export_ext}",
None,
mock.ANY,
[
f"{trainer.model_saver.model_path}{os.path.sep}{trainer.brain_name}-{step}.pt"
],
),
trainer.trainer_settings.keep_checkpoints,
) for step in checkpoint_range
]
mock_add_checkpoint.assert_has_calls(add_checkpoint_calls)
def test_summary_checkpoint(mock_add_checkpoint, mock_write_summary):
trainer = create_rl_trainer()
mock_policy = mock.Mock()
mock_policy.model_path = "mock_model_path"
trainer.add_policy("TestBrain", mock_policy)
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
time_horizon = 10
summary_freq = trainer.trainer_settings.summary_freq
checkpoint_interval = trainer.trainer_settings.checkpoint_interval
trajectory = mb.make_fake_trajectory(
length=time_horizon,
observation_shapes=[(1, )],
max_step_complete=True,
action_space=[2],
)
# Check that we can turn off the trainer and that the buffer is cleared
num_trajectories = 5
for _ in range(0, num_trajectories):
trajectory_queue.put(trajectory)
trainer.advance()
# Check that there is stuff in the policy queue
policy_queue.get_nowait()
# Check that we have called write_summary the appropriate number of times
calls = [
mock.call(step) for step in range(summary_freq, num_trajectories *
time_horizon, summary_freq)
]
mock_write_summary.assert_has_calls(calls, any_order=True)
checkpoint_range = range(checkpoint_interval,
num_trajectories * time_horizon,
checkpoint_interval)
calls = [
mock.call(f"{mock_policy.model_path}/{trainer.brain_name}-{step}",
mock.ANY) for step in checkpoint_range
]
mock_policy.checkpoint.assert_has_calls(calls, any_order=True)
add_checkpoint_calls = [
mock.call(
trainer.brain_name,
NNCheckpoint(
step,
f"{mock_policy.model_path}/{trainer.brain_name}-{step}.nn",
None,
mock.ANY,
),
trainer.trainer_settings.keep_checkpoints,
) for step in checkpoint_range
]
mock_add_checkpoint.assert_has_calls(add_checkpoint_calls)
def test_process_trajectory(dummy_config):
mock_specs = mb.setup_test_behavior_specs(True,
False,
vector_action_space=[2],
vector_obs_space=1)
behavior_id_team0 = "test_brain?team=0"
behavior_id_team1 = "test_brain?team=1"
brain_name = BehaviorIdentifiers.from_name_behavior_id(
behavior_id_team0).brain_name
ppo_trainer = PPOTrainer(brain_name, 0, dummy_config, True, False, 0, "0")
controller = GhostController(100)
trainer = GhostTrainer(ppo_trainer, brain_name, controller, 0,
dummy_config, True, "0")
# first policy encountered becomes policy trained by wrapped PPO
parsed_behavior_id0 = BehaviorIdentifiers.from_name_behavior_id(
behavior_id_team0)
policy = trainer.create_policy(parsed_behavior_id0, mock_specs)
trainer.add_policy(parsed_behavior_id0, policy)
trajectory_queue0 = AgentManagerQueue(behavior_id_team0)
trainer.subscribe_trajectory_queue(trajectory_queue0)
# Ghost trainer should ignore this queue because off policy
parsed_behavior_id1 = BehaviorIdentifiers.from_name_behavior_id(
behavior_id_team1)
policy = trainer.create_policy(parsed_behavior_id1, mock_specs)
trainer.add_policy(parsed_behavior_id1, policy)
trajectory_queue1 = AgentManagerQueue(behavior_id_team1)
trainer.subscribe_trajectory_queue(trajectory_queue1)
time_horizon = 15
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
observation_shapes=[(1, )],
action_space=[2],
)
trajectory_queue0.put(trajectory)
trainer.advance()
# Check that trainer put trajectory in update buffer
assert trainer.trainer.update_buffer.num_experiences == 15
trajectory_queue1.put(trajectory)
trainer.advance()
# Check that ghost trainer ignored off policy queue
assert trainer.trainer.update_buffer.num_experiences == 15
# Check that it emptied the queue
assert trajectory_queue1.empty()
def publish_policy_queue(self,
policy_queue: AgentManagerQueue[Policy]) -> None:
"""
Adds a policy queue for every member of the team to the list of queues to publish to when this Trainer
makes a policy update. Creates an internal policy queue for the wrapped
trainer to push to. The GhostTrainer pushes all policies to the env.
:param queue: Policy queue to publish to.
"""
super().publish_policy_queue(policy_queue)
parsed_behavior_id = self._name_to_parsed_behavior_id[
policy_queue.behavior_id]
try:
self._team_to_name_to_policy_queue[parsed_behavior_id.team_id][
parsed_behavior_id.brain_name] = policy_queue
except KeyError:
self._team_to_name_to_policy_queue[parsed_behavior_id.team_id] = {
parsed_behavior_id.brain_name: policy_queue
}
if parsed_behavior_id.team_id == self.wrapped_trainer_team:
# With a future multiagent trainer, this will be indexed by 'role'
internal_policy_queue: AgentManagerQueue[
Policy] = AgentManagerQueue(parsed_behavior_id.brain_name)
self._internal_policy_queues[
parsed_behavior_id.brain_name] = internal_policy_queue
self.trainer.publish_policy_queue(internal_policy_queue)
def test_advance(mocked_clear_update_buffer):
trainer = create_rl_trainer()
trajectory_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
time_horizon = 15
trajectory = mb.make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that get_step is correct
assert trainer.get_step == time_horizon
# Check that we can turn off the trainer and that the buffer is cleared
for _ in range(0, 10):
trajectory_queue.put(trajectory)
trainer.advance()
# Check that the buffer has been cleared
assert not trainer.should_still_train
assert mocked_clear_update_buffer.call_count > 0
def test_agent_manager_queue():
queue = AgentManagerQueue(behavior_id="testbehavior")
trajectory = mock.Mock(spec=Trajectory)
assert queue.empty()
queue.put(trajectory)
assert not queue.empty()
queue_traj = queue.get_nowait()
assert isinstance(queue_traj, Trajectory)
assert queue.empty()
def test_process_trajectory(dummy_config):
brain_params = BrainParameters(
brain_name="test_brain",
vector_observation_space_size=1,
camera_resolutions=[],
vector_action_space_size=[2],
vector_action_descriptions=[],
vector_action_space_type=0,
)
dummy_config["summary_path"] = "./summaries/test_trainer_summary"
dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
trainer = PPOTrainer(brain_params, 0, dummy_config, True, False, 0, "0")
policy = trainer.create_policy(brain_params)
trainer.add_policy(brain_params.brain_name, policy)
trajectory_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
time_horizon = 15
trajectory = make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that trainer put trajectory in update buffer
assert trainer.update_buffer.num_experiences == 15
# Check that GAE worked
assert (
"advantages" in trainer.update_buffer
and "discounted_returns" in trainer.update_buffer
)
# Check that the stats are being collected as episode isn't complete
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent > 0
# Add a terminal trajectory
trajectory = make_fake_trajectory(
length=time_horizon + 1,
max_step_complete=False,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that the stats are reset as episode is finished
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent == 0
assert trainer.stats_reporter.get_stats_summaries("Policy/Extrinsic Reward").num > 0
def test_advance(mocked_clear_update_buffer, mocked_save_model):
trainer = create_rl_trainer()
mock_policy = mock.Mock()
mock_policy.model_path = "mock_model_path"
trainer.add_policy("TestBrain", mock_policy)
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
time_horizon = 10
trajectory = mb.make_fake_trajectory(
length=time_horizon,
observation_shapes=[(1, )],
max_step_complete=True,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
policy_queue.get_nowait()
# Check that get_step is correct
assert trainer.get_step == time_horizon
# Check that we can turn off the trainer and that the buffer is cleared
for _ in range(0, 5):
trajectory_queue.put(trajectory)
trainer.advance()
# Check that there is stuff in the policy queue
policy_queue.get_nowait()
# Check that if the policy doesn't update, we don't push it to the queue
trainer.set_is_policy_updating(False)
for _ in range(0, 10):
trajectory_queue.put(trajectory)
trainer.advance()
# Check that there nothing in the policy queue
with pytest.raises(AgentManagerQueue.Empty):
policy_queue.get_nowait()
# Check that the buffer has been cleared
assert not trainer.should_still_train
assert mocked_clear_update_buffer.call_count > 0
assert mocked_save_model.call_count == 0
def test_process_trajectory(dummy_config):
behavior_spec = mb.setup_test_behavior_specs(
True,
False,
vector_action_space=DISCRETE_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
)
mock_brain_name = "MockBrain"
behavior_id = BehaviorIdentifiers.from_name_behavior_id(mock_brain_name)
trainer = PPOTrainer("test_brain", 0, dummy_config, True, False, 0, "0")
policy = trainer.create_policy(behavior_id, behavior_spec)
trainer.add_policy(behavior_id, policy)
trajectory_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
time_horizon = 15
trajectory = make_fake_trajectory(
length=time_horizon,
observation_shapes=behavior_spec.observation_shapes,
max_step_complete=True,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that trainer put trajectory in update buffer
assert trainer.update_buffer.num_experiences == 15
# Check that GAE worked
assert (
"advantages" in trainer.update_buffer
and "discounted_returns" in trainer.update_buffer
)
# Check that the stats are being collected as episode isn't complete
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent > 0
# Add a terminal trajectory
trajectory = make_fake_trajectory(
length=time_horizon + 1,
max_step_complete=False,
observation_shapes=behavior_spec.observation_shapes,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that the stats are reset as episode is finished
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent == 0
assert trainer.stats_reporter.get_stats_summaries("Policy/Extrinsic Reward").num > 0
def test_poca_end_episode():
name_behavior_id = "test_trainer"
trainer = POCATrainer(
name_behavior_id,
10,
TrainerSettings(max_steps=100, checkpoint_interval=10,
summary_freq=20),
True,
False,
0,
"mock_model_path",
)
behavior_spec = BehaviorSpec(create_observation_specs_with_shapes([(1, )]),
ActionSpec.create_discrete((2, )))
parsed_behavior_id = BehaviorIdentifiers.from_name_behavior_id(
name_behavior_id)
mock_policy = trainer.create_policy(parsed_behavior_id, behavior_spec)
trainer.add_policy(parsed_behavior_id, mock_policy)
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
time_horizon = 10
trajectory = mb.make_fake_trajectory(
length=time_horizon,
observation_specs=behavior_spec.observation_specs,
max_step_complete=False,
action_spec=behavior_spec.action_spec,
num_other_agents_in_group=2,
group_reward=1.0,
is_terminal=False,
)
trajectory_queue.put(trajectory)
trainer.advance()
# Test that some trajectoories have been injested
for reward in trainer.collected_group_rewards.values():
assert reward == 10
# Test end episode
trainer.end_episode()
assert len(trainer.collected_group_rewards.keys()) == 0
def test_process_trajectory(dummy_config):
brain_params = make_brain_parameters(discrete_action=False,
visual_inputs=0,
vec_obs_size=6)
dummy_config["summary_path"] = "./summaries/test_trainer_summary"
dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
trainer = SACTrainer(brain_params, 0, dummy_config, True, False, 0, "0")
policy = trainer.create_policy(brain_params.brain_name, brain_params)
trainer.add_policy(brain_params.brain_name, policy)
trajectory_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trajectory = make_fake_trajectory(
length=15,
max_step_complete=True,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that trainer put trajectory in update buffer
assert trainer.update_buffer.num_experiences == 15
# Check that the stats are being collected as episode isn't complete
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent > 0
# Add a terminal trajectory
trajectory = make_fake_trajectory(
length=15,
max_step_complete=False,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that the stats are reset as episode is finished
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent == 0
assert trainer.stats_reporter.get_stats_summaries(
"Policy/Extrinsic Reward").num > 0
# Assert we're not just using the default values
assert (trainer.stats_reporter.get_stats_summaries(
"Policy/Extrinsic Reward").mean > 0)
def test_summary_checkpoint(mock_write_summary, mock_save_model):
trainer = create_rl_trainer()
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
time_horizon = 10
summary_freq = trainer.trainer_settings.summary_freq
checkpoint_interval = trainer.trainer_settings.checkpoint_interval
trajectory = mb.make_fake_trajectory(
length=time_horizon,
max_step_complete=True,
vec_obs_size=1,
num_vis_obs=0,
action_space=[2],
)
# Check that we can turn off the trainer and that the buffer is cleared
num_trajectories = 5
for _ in range(0, num_trajectories):
trajectory_queue.put(trajectory)
trainer.advance()
# Check that there is stuff in the policy queue
policy_queue.get_nowait()
# Check that we have called write_summary the appropriate number of times
calls = [
mock.call(step) for step in range(summary_freq, num_trajectories *
time_horizon, summary_freq)
]
mock_write_summary.assert_has_calls(calls, any_order=True)
calls = [
mock.call(trainer.brain_name)
for step in range(checkpoint_interval, num_trajectories *
time_horizon, checkpoint_interval)
]
mock_save_model.assert_has_calls(calls, any_order=True)
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.
"""
if trajectory_queue.behavior_id == self.learning_behavior_name:
super().subscribe_trajectory_queue(trajectory_queue)
internal_trajectory_queue: AgentManagerQueue[
Trajectory] = AgentManagerQueue(trajectory_queue.behavior_id)
self.internal_trajectory_queues.append(internal_trajectory_queue)
self.trainer.subscribe_trajectory_queue(internal_trajectory_queue)
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.
"""
super().publish_policy_queue(policy_queue)
if policy_queue.behavior_id == self.learning_behavior_name:
internal_policy_queue: AgentManagerQueue[Policy] = AgentManagerQueue(
policy_queue.behavior_id
)
self.internal_policy_queues.append(internal_policy_queue)
self.learning_policy_queues[policy_queue.behavior_id] = policy_queue
self.trainer.publish_policy_queue(internal_policy_queue)
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 subscribe_trajectory_queue(
self, trajectory_queue: AgentManagerQueue[Trajectory]) -> None:
"""
Adds a trajectory queue for every member of the team to the list of queues for the trainer
to ingest Trajectories from. Creates an internal trajectory queue to push trajectories from
the learning team. The wrapped trainer subscribes to this queue.
:param queue: Trajectory queue to publish to.
"""
super().subscribe_trajectory_queue(trajectory_queue)
parsed_behavior_id = self._name_to_parsed_behavior_id[
trajectory_queue.behavior_id]
if parsed_behavior_id.team_id == self.wrapped_trainer_team:
# With a future multiagent trainer, this will be indexed by 'role'
internal_trajectory_queue: AgentManagerQueue[
Trajectory] = AgentManagerQueue(parsed_behavior_id.brain_name)
self._internal_trajectory_queues[
parsed_behavior_id.brain_name] = internal_trajectory_queue
self.trainer.subscribe_trajectory_queue(internal_trajectory_queue)
def test_publish_queue(dummy_config):
mock_specs = mb.setup_test_behavior_specs(
True, False, vector_action_space=[1], vector_obs_space=8
)
behavior_id_team0 = "test_brain?team=0"
behavior_id_team1 = "test_brain?team=1"
parsed_behavior_id0 = BehaviorIdentifiers.from_name_behavior_id(behavior_id_team0)
brain_name = parsed_behavior_id0.brain_name
ppo_trainer = PPOTrainer(brain_name, 0, dummy_config, True, False, 0, "0")
controller = GhostController(100)
trainer = GhostTrainer(
ppo_trainer, brain_name, controller, 0, dummy_config, True, "0"
)
# First policy encountered becomes policy trained by wrapped PPO
# This queue should remain empty after swap snapshot
policy = trainer.create_policy(parsed_behavior_id0, mock_specs)
trainer.add_policy(parsed_behavior_id0, policy)
policy_queue0 = AgentManagerQueue(behavior_id_team0)
trainer.publish_policy_queue(policy_queue0)
# Ghost trainer should use this queue for ghost policy swap
parsed_behavior_id1 = BehaviorIdentifiers.from_name_behavior_id(behavior_id_team1)
policy = trainer.create_policy(parsed_behavior_id1, mock_specs)
trainer.add_policy(parsed_behavior_id1, policy)
policy_queue1 = AgentManagerQueue(behavior_id_team1)
trainer.publish_policy_queue(policy_queue1)
# check ghost trainer swap pushes to ghost queue and not trainer
assert policy_queue0.empty() and policy_queue1.empty()
trainer._swap_snapshots()
assert policy_queue0.empty() and not policy_queue1.empty()
# clear
policy_queue1.get_nowait()
mock_specs = mb.setup_test_behavior_specs(
False,
False,
vector_action_space=VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
)
buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, mock_specs)
# Mock out reward signal eval
buffer["extrinsic_rewards"] = buffer["environment_rewards"]
buffer["extrinsic_returns"] = buffer["environment_rewards"]
buffer["extrinsic_value_estimates"] = buffer["environment_rewards"]
buffer["curiosity_rewards"] = buffer["environment_rewards"]
buffer["curiosity_returns"] = buffer["environment_rewards"]
buffer["curiosity_value_estimates"] = buffer["environment_rewards"]
buffer["advantages"] = buffer["environment_rewards"]
trainer.trainer.update_buffer = buffer
# when ghost trainer advance and wrapped trainer buffers full
# the wrapped trainer pushes updated policy to correct queue
assert policy_queue0.empty() and policy_queue1.empty()
trainer.advance()
assert not policy_queue0.empty() and policy_queue1.empty()
def test_publish_queue(dummy_config):
mock_specs = mb.setup_test_behavior_specs(True,
False,
vector_action_space=[1],
vector_obs_space=8)
behavior_id_team0 = "test_brain?team=0"
behavior_id_team1 = "test_brain?team=1"
parsed_behavior_id0 = BehaviorIdentifiers.from_name_behavior_id(
behavior_id_team0)
brain_name = parsed_behavior_id0.brain_name
ppo_trainer = PPOTrainer(brain_name, 0, dummy_config, True, False, 0, "0")
controller = GhostController(100)
trainer = GhostTrainer(ppo_trainer, brain_name, controller, 0,
dummy_config, True, "0")
# First policy encountered becomes policy trained by wrapped PPO
# This queue should remain empty after swap snapshot
policy = trainer.create_policy(parsed_behavior_id0, mock_specs)
trainer.add_policy(parsed_behavior_id0, policy)
policy_queue0 = AgentManagerQueue(behavior_id_team0)
trainer.publish_policy_queue(policy_queue0)
# Ghost trainer should use this queue for ghost policy swap
parsed_behavior_id1 = BehaviorIdentifiers.from_name_behavior_id(
behavior_id_team1)
policy = trainer.create_policy(parsed_behavior_id1, mock_specs)
trainer.add_policy(parsed_behavior_id1, policy)
policy_queue1 = AgentManagerQueue(behavior_id_team1)
trainer.publish_policy_queue(policy_queue1)
# check ghost trainer swap pushes to ghost queue and not trainer
assert policy_queue0.empty() and policy_queue1.empty()
trainer._swap_snapshots()
assert policy_queue0.empty() and not policy_queue1.empty()
# clear
policy_queue1.get_nowait()
buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, mock_specs)
# Mock out reward signal eval
copy_buffer_fields(
buffer,
src_key=BufferKey.ENVIRONMENT_REWARDS,
dst_keys=[
BufferKey.ADVANTAGES,
RewardSignalUtil.rewards_key("extrinsic"),
RewardSignalUtil.returns_key("extrinsic"),
RewardSignalUtil.value_estimates_key("extrinsic"),
RewardSignalUtil.rewards_key("curiosity"),
RewardSignalUtil.returns_key("curiosity"),
RewardSignalUtil.value_estimates_key("curiosity"),
],
)
trainer.trainer.update_buffer = buffer
# when ghost trainer advance and wrapped trainer buffers full
# the wrapped trainer pushes updated policy to correct queue
assert policy_queue0.empty() and policy_queue1.empty()
trainer.advance()
assert not policy_queue0.empty() and policy_queue1.empty()
def test_advance(dummy_config):
brain_params = make_brain_parameters(discrete_action=False,
visual_inputs=0,
vec_obs_size=6)
dummy_config["output_path"] = "./results/test_trainer_models/TestModel"
dummy_config["steps_per_update"] = 20
trainer = SACTrainer(brain_params, 0, dummy_config, True, False, 0, "0")
policy = trainer.create_policy(brain_params.brain_name, brain_params)
trainer.add_policy(brain_params.brain_name, policy)
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
trajectory = make_fake_trajectory(
length=15,
max_step_complete=True,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
is_discrete=False,
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that trainer put trajectory in update buffer
assert trainer.update_buffer.num_experiences == 15
# Check that the stats are being collected as episode isn't complete
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent > 0
# Add a terminal trajectory
trajectory = make_fake_trajectory(
length=6,
max_step_complete=False,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
is_discrete=False,
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that the stats are reset as episode is finished
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent == 0
assert trainer.stats_reporter.get_stats_summaries(
"Policy/Extrinsic Reward").num > 0
# Assert we're not just using the default values
assert (trainer.stats_reporter.get_stats_summaries(
"Policy/Extrinsic Reward").mean > 0)
# Make sure there is a policy on the queue
policy_queue.get_nowait()
# Add another trajectory. Since this is less than 20 steps total (enough for)
# two updates, there should NOT be a policy on the queue.
trajectory = make_fake_trajectory(
length=5,
max_step_complete=False,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
is_discrete=False,
)
trajectory_queue.put(trajectory)
trainer.advance()
with pytest.raises(AgentManagerQueue.Empty):
policy_queue.get_nowait()
def test_publish_queue(dummy_config):
brain_params_team0 = BrainParameters(
brain_name="test_brain?team=0",
vector_observation_space_size=8,
camera_resolutions=[],
vector_action_space_size=[1],
vector_action_descriptions=[],
vector_action_space_type=0,
)
parsed_behavior_id0 = BehaviorIdentifiers.from_name_behavior_id(
brain_params_team0.brain_name
)
brain_name = parsed_behavior_id0.brain_name
brain_params_team1 = BrainParameters(
brain_name="test_brain?team=1",
vector_observation_space_size=8,
camera_resolutions=[],
vector_action_space_size=[1],
vector_action_descriptions=[],
vector_action_space_type=0,
)
dummy_config["summary_path"] = "./summaries/test_trainer_summary"
dummy_config["model_path"] = "./models/test_trainer_models/TestModel"
ppo_trainer = PPOTrainer(brain_name, 0, dummy_config, True, False, 0, "0")
controller = GhostController(100)
trainer = GhostTrainer(
ppo_trainer, brain_name, controller, 0, dummy_config, True, "0"
)
# First policy encountered becomes policy trained by wrapped PPO
# This queue should remain empty after swap snapshot
policy = trainer.create_policy(parsed_behavior_id0, brain_params_team0)
trainer.add_policy(parsed_behavior_id0, policy)
policy_queue0 = AgentManagerQueue(brain_params_team0.brain_name)
trainer.publish_policy_queue(policy_queue0)
# Ghost trainer should use this queue for ghost policy swap
parsed_behavior_id1 = BehaviorIdentifiers.from_name_behavior_id(
brain_params_team1.brain_name
)
policy = trainer.create_policy(parsed_behavior_id1, brain_params_team1)
trainer.add_policy(parsed_behavior_id1, policy)
policy_queue1 = AgentManagerQueue(brain_params_team1.brain_name)
trainer.publish_policy_queue(policy_queue1)
# check ghost trainer swap pushes to ghost queue and not trainer
assert policy_queue0.empty() and policy_queue1.empty()
trainer._swap_snapshots()
assert policy_queue0.empty() and not policy_queue1.empty()
# clear
policy_queue1.get_nowait()
mock_brain = mb.setup_mock_brain(
False,
False,
vector_action_space=VECTOR_ACTION_SPACE,
vector_obs_space=VECTOR_OBS_SPACE,
discrete_action_space=DISCRETE_ACTION_SPACE,
)
buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, mock_brain)
# Mock out reward signal eval
buffer["extrinsic_rewards"] = buffer["environment_rewards"]
buffer["extrinsic_returns"] = buffer["environment_rewards"]
buffer["extrinsic_value_estimates"] = buffer["environment_rewards"]
buffer["curiosity_rewards"] = buffer["environment_rewards"]
buffer["curiosity_returns"] = buffer["environment_rewards"]
buffer["curiosity_value_estimates"] = buffer["environment_rewards"]
buffer["advantages"] = buffer["environment_rewards"]
trainer.trainer.update_buffer = buffer
# when ghost trainer advance and wrapped trainer buffers full
# the wrapped trainer pushes updated policy to correct queue
assert policy_queue0.empty() and policy_queue1.empty()
trainer.advance()
assert not policy_queue0.empty() and policy_queue1.empty()
def test_advance(dummy_config):
brain_params = make_brain_parameters(
discrete_action=False, visual_inputs=0, vec_obs_size=6
)
dummy_config.hyperparameters.steps_per_update = 20
dummy_config.hyperparameters.reward_signal_steps_per_update = 20
dummy_config.hyperparameters.buffer_init_steps = 0
trainer = SACTrainer(brain_params, 0, dummy_config, True, False, 0, "0")
policy = trainer.create_policy(brain_params.brain_name, brain_params)
trainer.add_policy(brain_params.brain_name, policy)
trajectory_queue = AgentManagerQueue("testbrain")
policy_queue = AgentManagerQueue("testbrain")
trainer.subscribe_trajectory_queue(trajectory_queue)
trainer.publish_policy_queue(policy_queue)
trajectory = make_fake_trajectory(
length=15,
max_step_complete=True,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
is_discrete=False,
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that trainer put trajectory in update buffer
assert trainer.update_buffer.num_experiences == 15
# Check that the stats are being collected as episode isn't complete
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent > 0
# Add a terminal trajectory
trajectory = make_fake_trajectory(
length=6,
max_step_complete=False,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
is_discrete=False,
)
trajectory_queue.put(trajectory)
trainer.advance()
# Check that the stats are reset as episode is finished
for reward in trainer.collected_rewards.values():
for agent in reward.values():
assert agent == 0
assert trainer.stats_reporter.get_stats_summaries("Policy/Extrinsic Reward").num > 0
# Assert we're not just using the default values
assert (
trainer.stats_reporter.get_stats_summaries("Policy/Extrinsic Reward").mean > 0
)
# Make sure there is a policy on the queue
policy_queue.get_nowait()
# Add another trajectory. Since this is less than 20 steps total (enough for)
# two updates, there should NOT be a policy on the queue.
trajectory = make_fake_trajectory(
length=5,
max_step_complete=False,
vec_obs_size=6,
num_vis_obs=0,
action_space=[2],
is_discrete=False,
)
trajectory_queue.put(trajectory)
trainer.advance()
with pytest.raises(AgentManagerQueue.Empty):
policy_queue.get_nowait()
# Call add_policy and check that we update the correct number of times.
# This is to emulate a load from checkpoint.
policy = trainer.create_policy(brain_params.brain_name, brain_params)
policy.get_current_step = lambda: 200
trainer.add_policy(brain_params.brain_name, policy)
trainer.optimizer.update = mock.Mock()
trainer.optimizer.update_reward_signals = mock.Mock()
trainer.optimizer.update_reward_signals.return_value = {}
trainer.optimizer.update.return_value = {}
trajectory_queue.put(trajectory)
trainer.advance()
# Make sure we did exactly 1 update
assert trainer.optimizer.update.call_count == 1
assert trainer.optimizer.update_reward_signals.call_count == 1
