def rl2_ppo_metaworld_ml1_push(ctxt, seed, max_path_length, meta_batch_size,
n_epochs, episode_per_task):
"""Train PPO with ML1 environment.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
max_path_length (int): Maximum length of a single rollout.
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 LocalTFRunner(snapshot_config=ctxt) as runner:
tasks = task_sampler.SetTaskSampler(lambda: RL2Env(
env=mwb.ML1.get_train_tasks('push-v1')))
env_spec = RL2Env(env=mwb.ML1.get_train_tasks('push-v1')).spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
algo = RL2PPO(rl2_max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
task_sampler=tasks,
env_spec=env_spec,
policy=policy,
baseline=baseline,
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,
max_path_length=max_path_length * episode_per_task)
runner.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker,
worker_args=dict(n_paths_per_trial=episode_per_task))
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)
def setup_method(self):
super().setup_method()
self.max_path_length = 100
self.meta_batch_size = 10
self.episode_per_task = 4
self.tasks = task_sampler.SetTaskSampler(
lambda: RL2Env(env=normalize(HalfCheetahDirEnv())))
self.env_spec = RL2Env(env=normalize(HalfCheetahDirEnv())).spec
self.policy = GaussianGRUPolicy(env_spec=self.env_spec,
hidden_dim=64,
state_include_action=False)
self.baseline = LinearFeatureBaseline(env_spec=self.env_spec)
def rl2_trpo_halfcheetah(ctxt, seed, max_path_length, meta_batch_size,
n_epochs, episode_per_task):
"""Train TRPO with HalfCheetah environment.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
max_path_length (int): Maximum length of a single rollout.
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 LocalTFRunner(snapshot_config=ctxt) as runner:
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,
worker_args=dict(n_paths_per_trial=episode_per_task))
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)
def test_observation_dimension(self):
env = PointEnv()
wrapped_env = RL2Env(PointEnv())
assert wrapped_env.spec.observation_space.shape[0] == (
env.observation_space.shape[0] + env.action_space.shape[0] + 2)
obs = env.reset()
obs2 = wrapped_env.reset()
assert obs.shape[0] + env.action_space.shape[0] + 2 == obs2.shape[0]
obs, _, _, _ = env.step(env.action_space.sample())
obs2, _, _, _ = wrapped_env.step(env.action_space.sample())
assert obs.shape[0] + env.action_space.shape[0] + 2 == obs2.shape[0]
def rl2_ppo_metaworld_ml10_meta_test(ctxt, seed, max_path_length,
meta_batch_size, n_epochs,
episode_per_task):
"""Train PPO with ML10 environment with meta-test.
Args:
ctxt (metarl.experiment.ExperimentContext): The experiment
configuration used by LocalRunner to create the snapshotter.
seed (int): Used to seed the random number generator to produce
determinism.
max_path_length (int): Maximum length of a single rollout.
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 LocalTFRunner(snapshot_config=ctxt) as runner:
ml10_train_envs = [
RL2Env(mwb.ML10.from_task(task_name))
for task_name in mwb.ML10.get_train_tasks().all_task_names
]
tasks = task_sampler.EnvPoolSampler(ml10_train_envs)
tasks.grow_pool(meta_batch_size)
ml10_test_envs = [
RL2Env(mwb.ML10.from_task(task_name))
for task_name in mwb.ML10.get_test_tasks().all_task_names
]
test_tasks = task_sampler.EnvPoolSampler(ml10_test_envs)
env_spec = ml10_train_envs[0].spec
policy = GaussianGRUPolicy(name='policy',
hidden_dim=64,
env_spec=env_spec,
state_include_action=False)
baseline = LinearFeatureBaseline(env_spec=env_spec)
meta_evaluator = MetaEvaluator(test_task_sampler=test_tasks,
n_exploration_traj=10,
n_test_rollouts=10,
max_path_length=max_path_length,
n_test_tasks=5)
algo = RL2PPO(rl2_max_path_length=max_path_length,
meta_batch_size=meta_batch_size,
task_sampler=tasks,
env_spec=env_spec,
policy=policy,
baseline=baseline,
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,
max_path_length=max_path_length * episode_per_task,
meta_evaluator=meta_evaluator,
n_epochs_per_eval=10)
runner.setup(algo,
tasks.sample(meta_batch_size),
sampler_cls=LocalSampler,
n_workers=meta_batch_size,
worker_class=RL2Worker,
worker_args=dict(n_paths_per_trial=episode_per_task))
runner.train(n_epochs=n_epochs,
batch_size=episode_per_task * max_path_length *
meta_batch_size)