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 test_rl2_ppo_ml10(self):
# pylint: disable=import-outside-toplevel
from metaworld.benchmarks import ML10
ML_train_envs = [
RL2Env(ML10.from_task(task_name))
for task_name in ML10.get_train_tasks().all_task_names
]
tasks = task_sampler.EnvPoolSampler(ML_train_envs)
tasks.grow_pool(self.meta_batch_size)
env_spec = ML_train_envs[0].spec
policy = GaussianGRUPolicy(env_spec=env_spec,
hidden_dim=64,
state_include_action=False,
name='policy')
baseline = LinearFeatureBaseline(env_spec=env_spec)
with LocalTFRunner(snapshot_config, sess=self.sess) as runner:
algo = RL2PPO(rl2_max_path_length=self.max_path_length,
meta_batch_size=self.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,
stop_entropy_gradient=True,
entropy_method='max',
policy_ent_coeff=0.02,
center_adv=False,
max_path_length=self.max_path_length *
self.episode_per_task)
runner.setup(
algo,
self.tasks.sample(self.meta_batch_size),
sampler_cls=LocalSampler,
n_workers=self.meta_batch_size,
worker_class=RL2Worker,
worker_args=dict(n_paths_per_trial=self.episode_per_task))
runner.train(n_epochs=1,
batch_size=self.episode_per_task *
self.max_path_length * self.meta_batch_size)
def test_env_pool_sampler():
# Import, construct environments here to avoid using up too much
# resources if this test isn't run.
# pylint: disable=import-outside-toplevel
from metaworld.benchmarks import ML10
train_tasks = ML10.get_train_tasks().all_task_names
ML10_train_envs = [
ML10.from_task(train_task) for train_task in train_tasks
]
tasks = task_sampler.EnvPoolSampler(ML10_train_envs)
assert tasks.n_tasks == 10
updates = tasks.sample(10)
for env in ML10_train_envs:
assert any(env is update() for update in updates)
with pytest.raises(ValueError):
tasks.sample(10, with_replacement=True)
with pytest.raises(ValueError):
tasks.sample(11)
tasks.grow_pool(20)
tasks.sample(20)
def rl2_ppo_ml10(ctxt, seed, max_path_length, meta_batch_size, n_epochs,
episode_per_task):
"""Train PPO with ML10 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.
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:
ML_train_envs = [
RL2Env(ML10.from_task(task_name))
for task_name in ML10.get_train_tasks().all_task_names
]
tasks = task_sampler.EnvPoolSampler(ML_train_envs)
tasks.grow_pool(meta_batch_size)
env_spec = ML_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)
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 torch_pearl_ml10(ctxt=None,
seed=1,
num_epochs=1000,
num_train_tasks=10,
num_test_tasks=5,
latent_size=7,
encoder_hidden_size=200,
net_size=300,
meta_batch_size=16,
num_steps_per_epoch=4000,
num_initial_steps=4000,
num_tasks_sample=15,
num_steps_prior=750,
num_extra_rl_steps_posterior=750,
batch_size=256,
embedding_batch_size=64,
embedding_mini_batch_size=64,
max_path_length=150,
reward_scale=10.,
use_gpu=False):
"""Train PEARL with ML10 environments.
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.
num_epochs (int): Number of training epochs.
num_train_tasks (int): Number of tasks for training.
num_test_tasks (int): Number of tasks for testing.
latent_size (int): Size of latent context vector.
encoder_hidden_size (int): Output dimension of dense layer of the
context encoder.
net_size (int): Output dimension of a dense layer of Q-function and
value function.
meta_batch_size (int): Meta batch size.
num_steps_per_epoch (int): Number of iterations per epoch.
num_initial_steps (int): Number of transitions obtained per task before
training.
num_tasks_sample (int): Number of random tasks to obtain data for each
iteration.
num_steps_prior (int): Number of transitions to obtain per task with
z ~ prior.
num_extra_rl_steps_posterior (int): Number of additional transitions
to obtain per task with z ~ posterior that are only used to train
the policy and NOT the encoder.
batch_size (int): Number of transitions in RL batch.
embedding_batch_size (int): Number of transitions in context batch.
embedding_mini_batch_size (int): Number of transitions in mini context
batch; should be same as embedding_batch_size for non-recurrent
encoder.
max_path_length (int): Maximum path length.
reward_scale (int): Reward scale.
use_gpu (bool): Whether or not to use GPU for training.
"""
set_seed(seed)
encoder_hidden_sizes = (encoder_hidden_size, encoder_hidden_size,
encoder_hidden_size)
# create multi-task environment and sample tasks
ML_train_envs = [
GarageEnv(normalize(ML10.from_task(task_name)))
for task_name in ML10.get_train_tasks().all_task_names
]
ML_test_envs = [
GarageEnv(normalize(ML10.from_task(task_name)))
for task_name in ML10.get_test_tasks().all_task_names
]
env_sampler = EnvPoolSampler(ML_train_envs)
env = env_sampler.sample(num_train_tasks)
test_env_sampler = EnvPoolSampler(ML_test_envs)
runner = LocalRunner(ctxt)
# instantiate networks
augmented_env = PEARL.augment_env_spec(env[0](), latent_size)
qf = ContinuousMLPQFunction(env_spec=augmented_env,
hidden_sizes=[net_size, net_size, net_size])
vf_env = PEARL.get_env_spec(env[0](), latent_size, 'vf')
vf = ContinuousMLPQFunction(env_spec=vf_env,
hidden_sizes=[net_size, net_size, net_size])
inner_policy = TanhGaussianMLPPolicy(
env_spec=augmented_env, hidden_sizes=[net_size, net_size, net_size])
pearl = PEARL(
env=env,
policy_class=ContextConditionedPolicy,
encoder_class=MLPEncoder,
inner_policy=inner_policy,
qf=qf,
vf=vf,
num_train_tasks=num_train_tasks,
num_test_tasks=num_test_tasks,
latent_dim=latent_size,
encoder_hidden_sizes=encoder_hidden_sizes,
test_env_sampler=test_env_sampler,
meta_batch_size=meta_batch_size,
num_steps_per_epoch=num_steps_per_epoch,
num_initial_steps=num_initial_steps,
num_tasks_sample=num_tasks_sample,
num_steps_prior=num_steps_prior,
num_extra_rl_steps_posterior=num_extra_rl_steps_posterior,
batch_size=batch_size,
embedding_batch_size=embedding_batch_size,
embedding_mini_batch_size=embedding_mini_batch_size,
max_path_length=max_path_length,
reward_scale=reward_scale,
)
tu.set_gpu_mode(use_gpu, gpu_id=0)
if use_gpu:
pearl.to()
runner.setup(algo=pearl,
env=env[0](),
sampler_cls=LocalSampler,
sampler_args=dict(max_path_length=max_path_length),
n_workers=1,
worker_class=PEARLWorker)
runner.train(n_epochs=num_epochs, batch_size=batch_size)