def run_task(*_):
"""Wrap VPG training task in the run_task function."""
env = TfEnv(env_name='CartPole-v1')
policy = CategoricalMLPPolicy(
name="policy", env_spec=env.spec, hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=100,
discount=0.99,
optimizer_args=dict(tf_optimizer_args=dict(learning_rate=0.01, )))
algo.train()
def test_vpg_cartpole(self):
"""Test VPG with Cartpole environment."""
logger.reset()
env = TfEnv(normalize(CartpoleEnv()))
policy = GaussianMLPPolicy(name="policy",
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=10,
discount=0.99,
optimizer_args=dict(tf_optimizer_args=dict(learning_rate=0.01, )))
last_avg_ret = algo.train(sess=self.sess)
assert last_avg_ret > 160
def test_vpg_cartpole(self):
"""Test VPG with CartPole-v1 environment."""
env = TfEnv(env_name="CartPole-v1")
policy = CategoricalMLPPolicy(name="policy",
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=10000,
max_path_length=100,
n_itr=10,
discount=0.99,
optimizer_args=dict(tf_optimizer_args=dict(learning_rate=0.01, )))
last_avg_ret = algo.train(sess=self.sess)
assert last_avg_ret > 90
env.close()
Example using VPG with ISSampler, iterations alternate between live and
importance sampled iterations.
"""
import gym
from garage.contrib.alexbeloi.is_sampler import ISSampler
from garage.envs import normalize
from garage.np.baselines import LinearFeatureBaseline
from garage.tf.algos import VPG
from garage.tf.policies import GaussianMLPPolicy
env = normalize(gym.make('InvertedPendulum-v2'))
policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(
env=env,
policy=policy,
baseline=baseline,
batch_size=4000,
max_path_length=100,
n_itr=40,
discount=0.99,
step_size=0.01,
sampler_cls=ISSampler,
sampler_args=dict(n_backtrack=1),
)
algo.train()
from garage.envs.box2d import CartpoleEnv
from garage.envs.mujoco import SwimmerEnv
from garage.tf.algos import VPG
from garage.tf.envs import TfEnv
from garage.tf.policies import GaussianMLPPolicy
from garage.misc.instrument import run_experiment
env = TfEnv(normalize(SwimmerEnv()))
policy = GaussianMLPPolicy(name="policy",
env_spec=env.spec,
hidden_sizes=(32, 32))
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = VPG(env=env,
policy=policy,
baseline=baseline,
batch_size=5000,
max_path_length=500,
n_itr=40,
discount=0.995,
optimizer_args=dict(tf_optimizer_args=dict(learning_rate=1e-4, )))
run_experiment(algo.train(),
n_parallel=1,
snapshot_mode="last",
seed=1,
use_gpu=True,
use_tf=True)