class TestPPO:
def setup_method(self):
self.env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.baseline = LinearFeatureBaseline(env_spec=self.env.spec)
def teardown_method(self):
self.env.close()
def test_ppo_pendulum(self):
"""Test DDPG with Pendulum environment."""
deterministic.set_seed(0)
runner = LocalRunner(snapshot_config)
algo = PPO(env_spec=self.env.spec,
policy=self.policy,
baseline=self.baseline,
optimizer=torch.optim.Adam,
max_path_length=100,
discount=0.99,
gae_lambda=0.97,
lr_clip_range=2e-1,
policy_lr=3e-4)
runner.setup(algo, self.env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 0
class TestTRPO:
"""Test class for TRPO."""
def setup_method(self):
"""Setup method which is called before every test."""
self.env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.value_function = LinearFeatureBaseline(env_spec=self.env.spec)
def teardown_method(self):
"""Teardown method which is called after every test."""
self.env.close()
@pytest.mark.mujoco
def test_trpo_pendulum(self):
"""Test TRPO with Pendulum environment."""
deterministic.set_seed(0)
runner = LocalRunner(snapshot_config)
algo = TRPO(env_spec=self.env.spec,
policy=self.policy,
value_function=self.value_function,
max_path_length=100,
discount=0.99,
gae_lambda=0.98)
runner.setup(algo, self.env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 50
def test_maml_trpo_pendulum():
"""Test PPO with Pendulum environment."""
env = GarageEnv(normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
rollouts_per_task = 5
max_path_length = 100
runner = LocalRunner(snapshot_config)
algo = MAMLTRPO(env=env,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
meta_batch_size=5,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=5,
batch_size=rollouts_per_task * max_path_length)
assert last_avg_ret > -5
env.close()
def maml_trpo(ctxt, seed, epochs, rollouts_per_task, meta_batch_size):
"""Set up environment and algorithm and run the task.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
epochs (int): Number of training epochs.
rollouts_per_task (int): Number of rollouts per epoch per task
for training.
meta_batch_size (int): Number of tasks sampled per batch.
"""
set_seed(seed)
env = GarageEnv(
normalize(ML10.get_train_tasks(), expected_action_scale=10.))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(100, 100),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
value_function = GaussianMLPValueFunction(env_spec=env.spec,
hidden_sizes=(32, 32),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
max_path_length = 100
test_task_names = ML10.get_test_tasks().all_task_names
test_tasks = [
GarageEnv(normalize(ML10.from_task(task), expected_action_scale=10.))
for task in test_task_names
]
test_sampler = EnvPoolSampler(test_tasks)
meta_evaluator = MetaEvaluator(test_task_sampler=test_sampler,
max_path_length=max_path_length,
n_test_tasks=len(test_task_names))
runner = LocalRunner(ctxt)
algo = MAMLTRPO(env=env,
policy=policy,
value_function=value_function,
max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1,
meta_evaluator=meta_evaluator)
runner.setup(algo, env)
runner.train(n_epochs=epochs,
batch_size=rollouts_per_task * max_path_length)
def setup_method(self):
self.env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.baseline = LinearFeatureBaseline(env_spec=self.env.spec)
def setup_method(self):
"""Setup method which is called before every test."""
self.env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.value_function = LinearFeatureBaseline(env_spec=self.env.spec)
def setup_method(self):
"""Setup method which is called before every test."""
self.env = GarageEnv(
normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.baseline = LinearFeatureBaseline(env_spec=self.env.spec)
class TestMAMLVPG:
"""Test class for MAML-VPG."""
def setup_method(self):
"""Setup method which is called before every test."""
self.env = GarageEnv(
normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.value_function = LinearFeatureBaseline(env_spec=self.env.spec)
def teardown_method(self):
"""Teardown method which is called after every test."""
self.env.close()
def test_ppo_pendulum(self):
"""Test PPO with Pendulum environment."""
deterministic.set_seed(0)
rollouts_per_task = 5
max_path_length = 100
task_sampler = SetTaskSampler(lambda: GarageEnv(
normalize(HalfCheetahDirEnv(), expected_action_scale=10.)))
meta_evaluator = MetaEvaluator(test_task_sampler=task_sampler,
max_path_length=max_path_length,
n_test_tasks=1,
n_test_rollouts=10)
runner = LocalRunner(snapshot_config)
algo = MAMLVPG(env=self.env,
policy=self.policy,
value_function=self.value_function,
max_path_length=max_path_length,
meta_batch_size=5,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1,
meta_evaluator=meta_evaluator)
runner.setup(algo, self.env)
last_avg_ret = runner.train(n_epochs=10,
batch_size=rollouts_per_task *
max_path_length)
assert last_avg_ret > -5
def maml_ppo(ctxt, seed, epochs, rollouts_per_task, meta_batch_size):
"""Set up environment and algorithm and run the task.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
epochs (int): Number of training epochs.
rollouts_per_task (int): Number of rollouts per epoch per task
for training.
meta_batch_size (int): Number of tasks sampled per batch.
"""
set_seed(seed)
env = GarageEnv(normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
max_path_length = 100
task_sampler = SetTaskSampler(lambda: GarageEnv(
normalize(HalfCheetahDirEnv(), expected_action_scale=10.)))
meta_evaluator = MetaEvaluator(test_task_sampler=task_sampler,
max_path_length=max_path_length,
n_test_tasks=1,
n_test_rollouts=10)
runner = LocalRunner(ctxt)
algo = MAMLPPO(env=env,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1,
meta_evaluator=meta_evaluator)
runner.setup(algo, env)
runner.train(n_epochs=epochs,
batch_size=rollouts_per_task * max_path_length)
def test_maml_trpo_dummy_named_env():
"""Test with dummy environment that has env_name."""
env = GarageEnv(
normalize(DummyMultiTaskBoxEnv(), expected_action_scale=10.))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
rollouts_per_task = 2
max_path_length = 100
runner = LocalRunner(snapshot_config)
algo = MAMLTRPO(env=env,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
meta_batch_size=5,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1)
runner.setup(algo, env)
runner.train(n_epochs=2, batch_size=rollouts_per_task * max_path_length)
def test_ppo_pendulum(self):
"""Test PPO with Pendulum environment."""
deterministic.set_seed(0)
rollouts_per_task = 5
max_path_length = 100
task_sampler = SetTaskSampler(lambda: GarageEnv(
normalize(HalfCheetahDirEnv(), expected_action_scale=10.)))
meta_evaluator = MetaEvaluator(test_task_sampler=task_sampler,
max_path_length=max_path_length,
n_test_tasks=1,
n_test_rollouts=10)
runner = LocalRunner(snapshot_config)
algo = MAMLVPG(env=self.env,
policy=self.policy,
value_function=self.value_function,
max_path_length=max_path_length,
meta_batch_size=5,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1,
meta_evaluator=meta_evaluator)
runner.setup(algo, self.env)
last_avg_ret = runner.train(n_epochs=10,
batch_size=rollouts_per_task *
max_path_length)
assert last_avg_ret > -5
def setup_method(self):
"""Setup method which is called before every test."""
self._env = GarageEnv(gym.make('InvertedDoublePendulum-v2'))
self._runner = LocalRunner(snapshot_config)
self._policy = GaussianMLPPolicy(env_spec=self._env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
self._params = {
'env_spec': self._env.spec,
'policy': self._policy,
'baseline': LinearFeatureBaseline(env_spec=self._env.spec),
'max_path_length': 100,
'discount': 0.99,
}
def setup_method(self):
self._env = GarageEnv(gym.make('InvertedDoublePendulum-v2'))
self._runner = LocalRunner(snapshot_config)
policy = GaussianMLPPolicy(env_spec=self._env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
self._params = {
'env_spec': self._env.spec,
'policy': policy,
'optimizer': torch.optim.Adam,
'baseline': LinearFeatureBaseline(env_spec=self._env.spec),
'max_path_length': 100,
'discount': 0.99,
'policy_lr': 1e-2
}
def torch_sac_half_cheetah(ctxt=None, seed=1):
"""Set up environment and algorithm and run the task.
Args:
ctxt (garage.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
"""
deterministic.set_seed(seed)
runner = LocalRunner(snapshot_config=ctxt)
env = GarageEnv(normalize(gym.make('HalfCheetah-v2')))
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[256, 256],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=1)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=1000,
max_path_length=500,
use_automatic_entropy_tuning=True,
replay_buffer=replay_buffer,
min_buffer_size=1e4,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=256,
reward_scale=1.,
steps_per_epoch=1)
if torch.cuda.is_available():
tu.set_gpu_mode(True)
else:
tu.set_gpu_mode(False)
sac.to()
runner.setup(algo=sac, env=env, sampler_cls=LocalSampler)
runner.train(n_epochs=1000, batch_size=1000)
def test_ddpg_pendulum(self):
"""
Test DDPG with Pendulum environment.
This environment has a [-3, 3] action_space bound.
"""
runner = LocalRunner()
env = GarageEnv(normalize(gym.make('InvertedPendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
algo = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
target_update_tau=1e-2,
policy_lr=1e-4,
qf_lr=1e-3,
discount=0.9)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10,
n_epoch_cycles=20,
batch_size=100)
assert last_avg_ret > 10
env.close()
def setup_method(self):
"""Setup method which is called before every test."""
self.env = GarageEnv(
normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.value_function = LinearFeatureBaseline(env_spec=self.env.spec)
self.algo = MAMLPPO(env=self.env,
policy=self.policy,
value_function=self.value_function,
max_path_length=100,
meta_batch_size=5,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1)
class TestLocalRunner:
"""Test class for LocalRunner."""
def setup_method(self):
"""Setup method which is called before every test."""
self.env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.baseline = LinearFeatureBaseline(env_spec=self.env.spec)
def teardown_method(self):
"""Teardown method which is called after every test."""
self.env.close()
@pytest.mark.mujoco
def test_set_plot(self):
deterministic.set_seed(0)
runner = LocalRunner(snapshot_config)
algo = PPO(env_spec=self.env.spec,
policy=self.policy,
baseline=self.baseline,
max_path_length=100,
discount=0.99,
gae_lambda=0.97,
lr_clip_range=2e-1)
runner.setup(algo, self.env)
runner.train(n_epochs=1, batch_size=100, plot=True)
assert isinstance(
runner._plotter,
Plotter), ('self.plotter in LocalRunner should be set to Plotter.')
def test_ddpg_double_pendulum(self):
"""Test DDPG with Pendulum environment."""
deterministic.set_seed(0)
runner = LocalRunner(snapshot_config)
env = GarageEnv(gym.make('InvertedDoublePendulum-v2'))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
algo = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
target_update_tau=1e-2,
discount=0.9)
runner.setup(algo, env)
last_avg_ret = runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 45
env.close()
def test_sac_inverted_pendulum():
"""Test Sac performance on inverted pendulum."""
# pylint: disable=unexpected-keyword-arg
env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
deterministic.set_seed(0)
policy = TanhGaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=[32, 32],
hidden_nonlinearity=torch.nn.ReLU,
output_nonlinearity=None,
min_std=np.exp(-20.),
max_std=np.exp(2.),
)
qf1 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[32, 32],
hidden_nonlinearity=F.relu)
qf2 = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[32, 32],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=1)
runner = LocalRunner(snapshot_config=snapshot_config)
sac = SAC(env_spec=env.spec,
policy=policy,
qf1=qf1,
qf2=qf2,
gradient_steps_per_itr=100,
max_path_length=100,
use_automatic_entropy_tuning=True,
replay_buffer=replay_buffer,
min_buffer_size=1e3,
target_update_tau=5e-3,
discount=0.99,
buffer_batch_size=64,
reward_scale=1.,
steps_per_epoch=2)
runner.setup(sac, env, sampler_cls=LocalSampler)
if torch.cuda.is_available():
tu.set_gpu_mode(True)
else:
tu.set_gpu_mode(False)
sac.to()
ret = runner.train(n_epochs=12, batch_size=200, plot=False)
assert ret > 85
def run_task(snapshot_config, *_):
"""Set up environment and algorithm and run the task.
Args:
snapshot_config (garage.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
If None, it will create one with default settings.
_ : Unused parameters
"""
runner = LocalRunner(snapshot_config)
env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
policy_optimizer = (torch.optim.Adagrad, {'lr': 1e-4, 'lr_decay': 0.99})
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
steps_per_epoch=20,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
target_update_tau=1e-2,
discount=0.9,
policy_optimizer=policy_optimizer,
qf_optimizer=torch.optim.Adam)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, batch_size=100)
def run_task(snapshot_config, *_):
"""Set up environment and algorithm and run the task.
Args:
snapshot_config (garage.experiment.SnapshotConfig): The snapshot
configuration used by LocalRunner to create the snapshotter.
If None, it will create one with default settings.
_ : Unused parameters
"""
env = GarageEnv(normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
policy = GaussianMLPPolicy(
env_spec=env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
baseline = LinearFeatureBaseline(env_spec=env.spec)
rollouts_per_task = 40
max_path_length = 100
runner = LocalRunner(snapshot_config)
algo = MAMLTRPO(env=env,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
meta_batch_size=20,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1)
runner.setup(algo, env)
runner.train(n_epochs=300, batch_size=rollouts_per_task * max_path_length)
def run_task(snapshot_config, *_):
"""Set up environment and algorithm and run the task."""
runner = LocalRunner(snapshot_config)
env = GarageEnv(normalize(gym.make('InvertedDoublePendulum-v2')))
action_noise = OUStrategy(env.spec, sigma=0.2)
policy = DeterministicMLPPolicy(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu,
output_nonlinearity=torch.tanh)
qf = ContinuousMLPQFunction(env_spec=env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=F.relu)
replay_buffer = SimpleReplayBuffer(env_spec=env.spec,
size_in_transitions=int(1e6),
time_horizon=100)
ddpg = DDPG(env_spec=env.spec,
policy=policy,
qf=qf,
replay_buffer=replay_buffer,
n_train_steps=50,
min_buffer_size=int(1e4),
exploration_strategy=action_noise,
target_update_tau=1e-2,
policy_lr=1e-4,
qf_lr=1e-3,
discount=0.9,
optimizer=torch.optim.Adam)
runner.setup(algo=ddpg, env=env)
runner.train(n_epochs=500, n_epoch_cycles=20, batch_size=100)
class TestVPG:
@classmethod
def setup_class(cls):
deterministic.set_seed(0)
def setup_method(self):
self._env = GarageEnv(gym.make('InvertedDoublePendulum-v2'))
self._runner = LocalRunner(snapshot_config)
policy = GaussianMLPPolicy(env_spec=self._env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
self._params = {
'env_spec': self._env.spec,
'policy': policy,
'optimizer': torch.optim.Adam,
'baseline': LinearFeatureBaseline(env_spec=self._env.spec),
'max_path_length': 100,
'discount': 0.99,
'policy_lr': 1e-2
}
def teardown_method(self):
self._env.close()
def test_vpg_no_entropy(self):
"""Test VPG with no_entropy."""
self._params['positive_adv'] = True
self._params['use_softplus_entropy'] = True
algo = VPG(**self._params)
self._runner.setup(algo, self._env)
last_avg_ret = self._runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 0
def test_vpg_max(self):
"""Test VPG with maximum entropy."""
self._params['center_adv'] = False
self._params['stop_entropy_gradient'] = True
self._params['entropy_method'] = 'max'
algo = VPG(**self._params)
self._runner.setup(algo, self._env)
last_avg_ret = self._runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 0
def test_vpg_regularized(self):
"""Test VPG with entropy_regularized."""
self._params['entropy_method'] = 'regularized'
algo = VPG(**self._params)
self._runner.setup(algo, self._env)
last_avg_ret = self._runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 30
@pytest.mark.parametrize('algo_param, error, msg', INVALID_ENTROPY_CONFIG)
def test_invalid_entropy_config(self, algo_param, error, msg):
self._params.update(algo_param)
with pytest.raises(error, match=msg):
VPG(**self._params)
def run_task(snapshot_config, *_,
env_params,
algo='TRPO',
algo_params={},
epochs=1000,
batch_size=4000,
policy_hidden_sizes=(32,32),
embed_state=False,
model_dir='../models/reacher_limited/train_6',
augment_embedded_state=False):
"""Run task."""
embed_config_file = os.path.join(model_dir, 'config.yaml')
ckpt_path = os.path.join(model_dir, 'model_latest.ckpt')
with LocalTFRunner(snapshot_config=snapshot_config) as runner:
if embed_state:
reacher_env = ReacherEmbeddedEnv(
embed_config_file,
ckpt_path,
augment_embedded_state=augment_embedded_state,
**env_params)
else:
reacher_env = ReacherEnv(**env_params)
env = GarageEnv(reacher_env)
policy = GaussianMLPPolicy(
name='policy',
env_spec=env.spec,
hidden_sizes=policy_hidden_sizes)
# hidden_sizes=(32, 32))
#************** TRPO ***************
if algo == 'TRPO':
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
**algo_params)
# max_path_length=100,
# discount=0.99,
# max_kl_step=0.01)
#**************** PPO *********************
elif algo == 'PPO':
baseline = GaussianMLPBaseline(
env_spec=env.spec,
regressor_args=dict(
hidden_sizes=(32, 32),
use_trust_region=True,
),
)
# NOTE: make sure when setting entropy_method to 'max', set
# center_adv to False and turn off policy gradient. See
# tf.algos.NPO for detailed documentation.
algo = PPO(
env_spec=env.spec,
policy=policy,
baseline=baseline,
# max_path_length=100,
# discount=0.99,
gae_lambda=0.95,
lr_clip_range=0.2,
optimizer_args=dict(
batch_size=32,
max_epochs=10,),
stop_entropy_gradient=True,
entropy_method='max',
policy_ent_coeff=0.02,
center_adv=False,
**algo_params)
#**************** Other? **********************
else:
print("ERROR: requested unrecognized algorithm: ", algo)
raise NotImplementedError
runner.setup(algo, env)
runner.train(n_epochs=epochs, batch_size=batch_size)
class TestVPG:
"""Test class for VPG."""
@classmethod
def setup_class(cls):
"""Setup method which is called once before all tests in this class."""
deterministic.set_seed(0)
def setup_method(self):
"""Setup method which is called before every test."""
self._env = GarageEnv(gym.make('InvertedDoublePendulum-v2'))
self._runner = LocalRunner(snapshot_config)
self._policy = GaussianMLPPolicy(env_spec=self._env.spec,
hidden_sizes=[64, 64],
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None)
self._params = {
'env_spec': self._env.spec,
'policy': self._policy,
'value_function':
GaussianMLPValueFunction(env_spec=self._env.spec),
'max_path_length': 100,
'discount': 0.99,
}
def teardown_method(self):
"""Teardown method which is called after every test."""
self._env.close()
@pytest.mark.mujoco
def test_vpg_no_entropy(self):
"""Test VPG with no_entropy."""
self._params['positive_adv'] = True
self._params['use_softplus_entropy'] = True
algo = VPG(**self._params)
self._runner.setup(algo, self._env)
last_avg_ret = self._runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 0
@pytest.mark.mujoco
def test_vpg_max(self):
"""Test VPG with maximum entropy."""
self._params['center_adv'] = False
self._params['stop_entropy_gradient'] = True
self._params['entropy_method'] = 'max'
algo = VPG(**self._params)
self._runner.setup(algo, self._env)
last_avg_ret = self._runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 0
@pytest.mark.mujoco
def test_vpg_regularized(self):
"""Test VPG with entropy_regularized."""
self._params['entropy_method'] = 'regularized'
algo = VPG(**self._params)
self._runner.setup(algo, self._env)
last_avg_ret = self._runner.train(n_epochs=10, batch_size=100)
assert last_avg_ret > 0
@pytest.mark.mujoco
@pytest.mark.parametrize('algo_param, error, msg', INVALID_ENTROPY_CONFIG)
def test_invalid_entropy_config(self, algo_param, error, msg):
"""Test VPG with invalid entropy config."""
self._params.update(algo_param)
with pytest.raises(error, match=msg):
VPG(**self._params)
class TestMAML:
"""Test class for MAML."""
def setup_method(self):
"""Setup method which is called before every test."""
self.env = GarageEnv(
normalize(HalfCheetahDirEnv(), expected_action_scale=10.))
self.policy = GaussianMLPPolicy(
env_spec=self.env.spec,
hidden_sizes=(64, 64),
hidden_nonlinearity=torch.tanh,
output_nonlinearity=None,
)
self.value_function = GaussianMLPValueFunction(env_spec=self.env.spec,
hidden_sizes=(32, 32))
self.algo = MAMLPPO(env=self.env,
policy=self.policy,
value_function=self.value_function,
max_path_length=100,
meta_batch_size=5,
discount=0.99,
gae_lambda=1.,
inner_lr=0.1,
num_grad_updates=1)
def teardown_method(self):
"""Teardown method which is called after every test."""
self.env.close()
@staticmethod
def _set_params(v, m):
"""Set the parameters of a module to a value."""
if isinstance(m, torch.nn.Linear):
m.weight.data.fill_(v)
m.bias.data.fill_(v)
@staticmethod
def _test_params(v, m):
"""Test if all parameters of a module equal to a value."""
if isinstance(m, torch.nn.Linear):
assert torch.all(torch.eq(m.weight.data, v))
assert torch.all(torch.eq(m.bias.data, v))
def test_get_exploration_policy(self):
"""Test if an independent copy of policy is returned."""
self.policy.apply(partial(self._set_params, 0.1))
adapt_policy = self.algo.get_exploration_policy()
adapt_policy.apply(partial(self._set_params, 0.2))
# Old policy should remain untouched
self.policy.apply(partial(self._test_params, 0.1))
adapt_policy.apply(partial(self._test_params, 0.2))
def test_adapt_policy(self):
"""Test if policy can adapt to samples."""
worker = WorkerFactory(seed=100, max_path_length=100)
sampler = LocalSampler.from_worker_factory(worker, self.policy,
self.env)
self.policy.apply(partial(self._set_params, 0.1))
adapt_policy = self.algo.get_exploration_policy()
trajs = sampler.obtain_samples(0, 100, adapt_policy)
self.algo.adapt_policy(adapt_policy, trajs)
# Old policy should remain untouched
self.policy.apply(partial(self._test_params, 0.1))
# Adapted policy should not be identical to old policy
for v1, v2 in zip(adapt_policy.parameters(), self.policy.parameters()):
if v1.data.ne(v2.data).sum() > 0:
break
else:
pytest.fail('Parameters of adapted policy should not be '
'identical to the old policy.')