def test_rl2_trpo_pendulum(self):
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
algo = RL2TRPO(
rl2_max_path_length=self.max_path_length,
meta_batch_size=self.meta_batch_size,
task_sampler=self.tasks,
env_spec=self.env_spec,
policy=self.policy,
baseline=self.baseline,
max_path_length=self.max_path_length * self.episode_per_task,
discount=0.99,
max_kl_step=0.01,
optimizer=ConjugateGradientOptimizer,
optimizer_args=dict(hvp_approach=FiniteDifferenceHvp(
base_eps=1e-5)))
runner.setup(algo,
self.tasks.sample(self.meta_batch_size),
sampler_cls=LocalSampler,
n_workers=self.meta_batch_size,
worker_class=RL2Worker)
last_avg_ret = runner.train(n_epochs=1,
batch_size=self.episode_per_task *
self.max_path_length *
self.meta_batch_size)
assert last_avg_ret > -40
def rl2_trpo_halfcheetah(ctxt, seed, max_episode_length, meta_batch_size,
n_epochs, episode_per_task):
"""Train TRPO with HalfCheetah environment.
Args:
ctxt (ExperimentContext): The experiment configuration used by
:class:`~Trainer` to create the :class:`~Snapshotter`.
seed (int): Used to seed the random number generator to produce
determinism.
max_episode_length (int): Maximum length of a single episode.
meta_batch_size (int): Meta batch size.
n_epochs (int): Total number of epochs for training.
episode_per_task (int): Number of training episode per task.
"""
set_seed(seed)
with TFTrainer(snapshot_config=ctxt) as trainer:
tasks = task_sampler.SetTaskSampler(
HalfCheetahVelEnv,
wrapper=lambda env, _: RL2Env(
GymEnv(env, max_episode_length=max_episode_length)))
env_spec = RL2Env(
GymEnv(HalfCheetahVelEnv(),
max_episode_length=max_episode_length)).spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
algo = RL2TRPO(meta_batch_size=meta_batch_size,
task_sampler=tasks,
env_spec=env_spec,
policy=policy,
baseline=baseline,
episodes_per_trial=episode_per_task,
discount=0.99,
max_kl_step=0.01,
optimizer=ConjugateGradientOptimizer,
optimizer_args=dict(hvp_approach=FiniteDifferenceHVP(
base_eps=1e-5)))
trainer.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker,
worker_args=dict(n_episodes_per_trial=episode_per_task))
trainer.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_episode_length *
meta_batch_size)
def test_rl2_trpo_pendulum_invalid_kl_constraint(self):
with LocalTFRunner(snapshot_config, sess=self.sess):
with pytest.raises(ValueError):
RL2TRPO(meta_batch_size=self.meta_batch_size,
task_sampler=self.tasks,
env_spec=self.env_spec,
policy=self.policy,
baseline=self.baseline,
kl_constraint='xyz',
episodes_per_trial=self.episode_per_task,
discount=0.99,
max_kl_step=0.01)
def test_ppo_pendulum_default_optimizer2(self):
with LocalTFRunner(snapshot_config, sess=self.sess):
algo = RL2TRPO(meta_batch_size=self.meta_batch_size,
task_sampler=self.tasks,
env_spec=self.env_spec,
policy=self.policy,
baseline=self.baseline,
kl_constraint='soft',
episodes_per_trial=self.episode_per_task,
discount=0.99,
max_kl_step=0.01)
assert isinstance(algo._inner_algo._optimizer,
PenaltyLbfgsOptimizer)
def test_rl2_trpo_pendulum_default_optimizer(self):
with TFTrainer(snapshot_config, sess=self.sess):
algo = RL2TRPO(meta_batch_size=self.meta_batch_size,
task_sampler=self.tasks,
env_spec=self.env_spec,
policy=self.policy,
baseline=self.baseline,
kl_constraint='hard',
episodes_per_trial=self.episode_per_task,
discount=0.99,
max_kl_step=0.01)
assert isinstance(algo._inner_algo._optimizer,
ConjugateGradientOptimizer)
def rl2_trpo_halfcheetah(ctxt=None, seed=1):
"""Train TRPO with HalfCheetah environment.
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.
"""
set_seed(seed)
with LocalTFRunner(snapshot_config=ctxt) as runner:
max_path_length = 100
meta_batch_size = 10
n_epochs = 50
episode_per_task = 4
tasks = task_sampler.SetTaskSampler(lambda: RL2Env(
env=HalfCheetahVelEnv()))
env_spec = RL2Env(env=HalfCheetahVelEnv()).spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
algo = RL2TRPO(rl2_max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
task_sampler=tasks,
env_spec=env_spec,
policy=policy,
baseline=baseline,
max_path_length=max_path_length * episode_per_task,
discount=0.99,
max_kl_step=0.01,
optimizer=ConjugateGradientOptimizer,
optimizer_args=dict(hvp_approach=FiniteDifferenceHvp(
base_eps=1e-5)))
runner.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker)
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)
def test_rl2_trpo_pendulum(self):
with TFTrainer(snapshot_config, sess=self.sess) as trainer:
algo = RL2TRPO(
meta_batch_size=self.meta_batch_size,
task_sampler=self.tasks,
env_spec=self.env_spec,
policy=self.policy,
baseline=self.baseline,
sampler=self.sampler,
episodes_per_trial=self.episode_per_task,
discount=0.99,
max_kl_step=0.01,
optimizer=ConjugateGradientOptimizer,
optimizer_args=dict(hvp_approach=FiniteDifferenceHVP(
base_eps=1e-5)))
trainer.setup(algo, self.tasks.sample(self.meta_batch_size))
last_avg_ret = trainer.train(n_epochs=1,
batch_size=self.episode_per_task *
self.max_episode_length *
self.meta_batch_size)
assert last_avg_ret > -40