def test_ppo_optimizer_update_gail(gail_dummy_config, dummy_config): # noqa: F811
# Test evaluate
dummy_config.reward_signals = gail_dummy_config
config = ppo_dummy_config()
optimizer = create_test_poca_optimizer(
config, use_rnn=False, use_discrete=False, use_visual=False
)
# Test update
update_buffer = mb.simulate_rollout(
BUFFER_INIT_SAMPLES, optimizer.policy.behavior_spec
)
# Mock out reward signal eval
copy_buffer_fields(
update_buffer,
src_key=BufferKey.ENVIRONMENT_REWARDS,
dst_keys=[
BufferKey.ADVANTAGES,
RewardSignalUtil.returns_key("extrinsic"),
RewardSignalUtil.value_estimates_key("extrinsic"),
RewardSignalUtil.baseline_estimates_key("extrinsic"),
RewardSignalUtil.returns_key("gail"),
RewardSignalUtil.value_estimates_key("gail"),
RewardSignalUtil.baseline_estimates_key("gail"),
],
)
update_buffer[BufferKey.CONTINUOUS_LOG_PROBS] = np.ones_like(
update_buffer[BufferKey.CONTINUOUS_ACTION]
)
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences // optimizer.policy.sequence_length,
)
# Check if buffer size is too big
update_buffer = mb.simulate_rollout(3000, optimizer.policy.behavior_spec)
# Mock out reward signal eval
copy_buffer_fields(
update_buffer,
src_key=BufferKey.ENVIRONMENT_REWARDS,
dst_keys=[
BufferKey.ADVANTAGES,
RewardSignalUtil.returns_key("extrinsic"),
RewardSignalUtil.value_estimates_key("extrinsic"),
RewardSignalUtil.baseline_estimates_key("extrinsic"),
RewardSignalUtil.returns_key("gail"),
RewardSignalUtil.value_estimates_key("gail"),
RewardSignalUtil.baseline_estimates_key("gail"),
],
)
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences // optimizer.policy.sequence_length,
)
def test_ppo_optimizer_update_gail(gail_dummy_config,
dummy_config): # noqa: F811
# Test evaluate
tf.reset_default_graph()
dummy_config.reward_signals = gail_dummy_config
optimizer = _create_ppo_optimizer_ops_mock(
attr.evolve(ppo_dummy_config(), framework=FrameworkType.TENSORFLOW),
use_rnn=False,
use_discrete=False,
use_visual=False,
)
# Test update
behavior_spec = optimizer.policy.behavior_spec
update_buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
# NOTE: This is because TF outputs the log probs of all actions whereas PyTorch does not
n_agents = len(update_buffer["continuous_log_probs"])
update_buffer["continuous_log_probs"] = np.ones(
(n_agents, behavior_spec.action_spec.continuous_size),
dtype=np.float32)
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
# Check if buffer size is too big
update_buffer = mb.simulate_rollout(3000, optimizer.policy.behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
def test_subprocess_failing_step(num_envs):
def failing_step_env_factory(_worker_id, _config):
env = UnexpectedExceptionEnvironment(
["1D"], use_discrete=True, to_raise=CustomTestOnlyException
)
return env
env_manager = SubprocessEnvManager(failing_step_env_factory, RunOptions())
# Expect the exception raised to be routed back up to the top level.
with pytest.raises(CustomTestOnlyException):
check_environment_trains(
failing_step_env_factory(0, []),
{"1D": ppo_dummy_config()},
env_manager=env_manager,
success_threshold=None,
)
env_manager.close()
def test_ppo_optimizer_update_gail(gail_dummy_config,
dummy_config): # noqa: F811
# Test evaluate
dummy_config.reward_signals = gail_dummy_config
config = attr.evolve(ppo_dummy_config(), framework=FrameworkType.PYTORCH)
optimizer = create_test_ppo_optimizer(config,
use_rnn=False,
use_discrete=False,
use_visual=False)
# Test update
update_buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES,
optimizer.policy.behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
# Check if buffer size is too big
update_buffer = mb.simulate_rollout(3000, optimizer.policy.behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
# NOTE: In TensorFlow, the log_probs are saved as one for every discrete action, whereas
# in PyTorch it is saved as the total probability per branch. So we need to modify the
# log prob in the fake buffer here.
update_buffer["action_probs"] = np.ones_like(update_buffer["actions"])
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
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 test_ppo_optimizer_update_gail(gail_dummy_config,
dummy_config): # noqa: F811
# Test evaluate
dummy_config.reward_signals = gail_dummy_config
config = ppo_dummy_config()
optimizer = create_test_ppo_optimizer(config,
use_rnn=False,
use_discrete=False,
use_visual=False)
# Test update
update_buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES,
optimizer.policy.behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["continuous_log_probs"] = np.ones_like(
update_buffer["continuous_action"])
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
# Check if buffer size is too big
update_buffer = mb.simulate_rollout(3000, optimizer.policy.behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
def test_ppo_optimizer_update_gail(gail_dummy_config,
dummy_config): # noqa: F811
# Test evaluate
tf.reset_default_graph()
dummy_config.reward_signals = gail_dummy_config
optimizer = _create_ppo_optimizer_ops_mock(
attr.evolve(ppo_dummy_config(), framework=FrameworkType.TENSORFLOW),
use_rnn=False,
use_discrete=False,
use_visual=False,
)
# Test update
update_buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES,
optimizer.policy.behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
# Check if buffer size is too big
update_buffer = mb.simulate_rollout(3000, optimizer.policy.behavior_spec)
# Mock out reward signal eval
update_buffer["advantages"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_returns"] = update_buffer["environment_rewards"]
update_buffer["extrinsic_value_estimates"] = update_buffer[
"environment_rewards"]
update_buffer["gail_returns"] = update_buffer["environment_rewards"]
update_buffer["gail_value_estimates"] = update_buffer[
"environment_rewards"]
optimizer.update(
update_buffer,
num_sequences=update_buffer.num_experiences //
optimizer.policy.sequence_length,
)
def test_subprocess_env_endtoend(num_envs):
def simple_env_factory(worker_id, config):
env = SimpleEnvironment(["1D"], use_discrete=True)
return env
env_manager = SubprocessEnvManager(simple_env_factory,
EngineConfig.default_config(), num_envs)
# Run PPO using env_manager
check_environment_trains(
simple_env_factory(0, []),
{"1D": ppo_dummy_config()},
env_manager=env_manager,
success_threshold=None,
)
# Note we can't check the env's rewards directly (since they're in separate processes) so we
# check the StatsReporter's debug stat writer's last reward.
assert isinstance(StatsReporter.writers[0], DebugWriter)
assert all(val > 0.7
for val in StatsReporter.writers[0].get_last_rewards().values())
env_manager.close()
def test_trainer_increment_step(ppo_optimizer, mock_create_model_saver):
trainer_params = attr.evolve(
attr.evolve(ppo_dummy_config(), framework=FrameworkType.TENSORFLOW),
framework=FrameworkType.TENSORFLOW,
)
mock_optimizer = mock.Mock()
mock_optimizer.reward_signals = {}
ppo_optimizer.return_value = mock_optimizer
trainer = PPOTrainer("test_brain", 0, trainer_params, True, False, 0, "0")
policy_mock = mock.Mock(spec=TFPolicy)
policy_mock.get_current_step.return_value = 0
step_count = (
5 # 10 hacked because this function is no longer called through trainer
)
policy_mock.increment_step = mock.Mock(return_value=step_count)
behavior_id = BehaviorIdentifiers.from_name_behavior_id(trainer.brain_name)
trainer.add_policy(behavior_id, policy_mock)
trainer._increment_step(5, trainer.brain_name)
policy_mock.increment_step.assert_called_with(5)
assert trainer.step == step_count
GAILSettings,
RewardSignalType,
EncoderType,
FrameworkType,
)
from mlagents.trainers.environment_parameter_manager import EnvironmentParameterManager
from mlagents_envs.side_channel.environment_parameters_channel import (
EnvironmentParametersChannel, )
from mlagents_envs.communicator_objects.demonstration_meta_pb2 import (
DemonstrationMetaProto, )
from mlagents_envs.communicator_objects.brain_parameters_pb2 import BrainParametersProto
from mlagents_envs.communicator_objects.space_type_pb2 import discrete, continuous
from mlagents.trainers.tests.dummy_config import ppo_dummy_config, sac_dummy_config
PPO_TF_CONFIG = attr.evolve(ppo_dummy_config(),
framework=FrameworkType.TENSORFLOW)
SAC_TF_CONFIG = attr.evolve(sac_dummy_config(),
framework=FrameworkType.TENSORFLOW)
BRAIN_NAME = "1D"
# The reward processor is passed as an argument to _check_environment_trains.
# It is applied to the list of all final rewards for each brain individually.
# This is so that we can process all final rewards in different ways for different algorithms.
# Custom reward processors should be built within the test function and passed to _check_environment_trains
# Default is average over the last 5 final rewards
def default_reward_processor(rewards, last_n_rewards=5):
rewards_to_use = rewards[-last_n_rewards:]
# For debugging tests
import pytest
from mlagents.trainers.tests.simple_test_envs import (
SimpleEnvironment,
MemoryEnvironment,
)
from mlagents.trainers.settings import NetworkSettings, FrameworkType
from mlagents.trainers.tests.dummy_config import ppo_dummy_config, sac_dummy_config
from mlagents.trainers.tests.check_env_trains import check_environment_trains
BRAIN_NAME = "1D"
PPO_TORCH_CONFIG = attr.evolve(ppo_dummy_config(), framework=FrameworkType.PYTORCH)
SAC_TORCH_CONFIG = attr.evolve(sac_dummy_config(), framework=FrameworkType.PYTORCH)
@pytest.mark.parametrize("action_size", [(1, 1), (2, 2), (1, 2), (2, 1)])
def test_hybrid_ppo(action_size):
env = SimpleEnvironment([BRAIN_NAME], action_sizes=action_size, step_size=0.8)
new_network_settings = attr.evolve(PPO_TORCH_CONFIG.network_settings)
new_hyperparams = attr.evolve(
PPO_TORCH_CONFIG.hyperparameters, batch_size=64, buffer_size=1024
)
config = attr.evolve(
PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network_settings,
max_steps=10000,
def dummy_config():
return attr.evolve(ppo_dummy_config(), framework=FrameworkType.PYTORCH)
def dummy_config():
return attr.evolve(ppo_dummy_config(), framework=FrameworkType.TENSORFLOW)
def dummy_config():
return RunOptions(behaviors={"testbrain": ppo_dummy_config()})
def dummy_config():
# poca has the same hyperparameters as ppo for now
return ppo_dummy_config()
def dummy_config():
return ppo_dummy_config()
import attr
import pytest
from mlagents.trainers.tests.simple_test_envs import (
SimpleEnvironment,
MemoryEnvironment,
)
from mlagents.trainers.settings import NetworkSettings
from mlagents.trainers.tests.dummy_config import ppo_dummy_config, sac_dummy_config
from mlagents.trainers.tests.check_env_trains import check_environment_trains
BRAIN_NAME = "1D"
PPO_TORCH_CONFIG = ppo_dummy_config()
SAC_TORCH_CONFIG = sac_dummy_config()
@pytest.mark.parametrize("action_size", [(1, 1), (2, 2), (1, 2), (2, 1)])
def test_hybrid_ppo(action_size):
env = SimpleEnvironment([BRAIN_NAME],
action_sizes=action_size,
step_size=0.8)
new_network_settings = attr.evolve(PPO_TORCH_CONFIG.network_settings)
new_hyperparams = attr.evolve(PPO_TORCH_CONFIG.hyperparameters,
batch_size=64,
buffer_size=1024)
config = attr.evolve(
PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
assert rsig_result.unscaled_reward.shape == (BATCH_SIZE,)
def reward_signal_update(optimizer, reward_signal_name):
buffer = mb.simulate_rollout(BUFFER_INIT_SAMPLES, optimizer.policy.behavior_spec)
feed_dict = optimizer.reward_signals[reward_signal_name].prepare_update(
optimizer.policy, buffer.make_mini_batch(0, 10), 2
)
out = optimizer.policy._execute_model(
feed_dict, optimizer.reward_signals[reward_signal_name].update_dict
)
assert type(out) is dict
@pytest.mark.parametrize(
"trainer_config", [ppo_dummy_config(), sac_dummy_config()], ids=["ppo", "sac"]
)
def test_gail_cc(trainer_config, gail_dummy_config): # noqa: F811
trainer_config.behavioral_cloning = BehavioralCloningSettings(
demo_path=CONTINUOUS_DEMO_PATH
)
optimizer = create_optimizer_mock(
trainer_config, gail_dummy_config, False, False, False
)
reward_signal_eval(optimizer, "gail")
reward_signal_update(optimizer, "gail")
@pytest.mark.parametrize(
"trainer_config", [ppo_dummy_config(), sac_dummy_config()], ids=["ppo", "sac"]
)