def run_task(*_):
env_name = "Ant"
hidden_sizes = (32,32)
env = TheanoEnv(normalize(SwimmerEnv()))
policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=hidden_sizes)
backup_policy = GaussianMLPPolicy(env.spec, hidden_sizes=hidden_sizes)
mix_policy = GaussianMLPPolicy(env.spec, hidden_sizes=hidden_sizes)
pos_eps_policy = GaussianMLPPolicy(env.spec, hidden_sizes=hidden_sizes)
neg_eps_policy = GaussianMLPPolicy(env.spec, hidden_sizes=hidden_sizes)
baseline = ZeroBaseline(env_spec=env.spec)
algo = CAPG(
env=env,
policy=policy,
backup_policy=backup_policy,
mix_policy=mix_policy,
pos_eps_policy=pos_eps_policy,
neg_eps_policy=neg_eps_policy,
n_timestep=5e6,
learning_rate=0.01,
batch_size=5000,
minibatch_size=500,
n_sub_itr = 10,
baseline=baseline,
max_path_length=500,
discount=0.99,
decay_learing_rate=True,
log_dir='./logs/' + env_name,
)
algo.train()
def test_polopt_algo(self, algo_cls, env_cls, policy_cls):
print("Testing %s, %s, %s" %
(algo_cls.__name__, env_cls.__name__, policy_cls.__name__))
env = TheanoEnv(env_cls())
policy = policy_cls(env_spec=env.spec, )
baseline = ZeroBaseline(env_spec=env.spec)
algo = algo_cls(env=env,
policy=policy,
baseline=baseline,
**(algo_args.get(algo_cls, dict())))
algo.train()
assert not np.any(np.isnan(policy.get_param_values()))
def test_adaptive_std():
"""
Checks if the adaptive_std parameter works.
"""
env = TheanoEnv(CartpoleEnv())
policy = GaussianMLPPolicy(env_spec=env, adaptive_std=True)
baseline = ZeroBaseline(env_spec=env.spec)
algo = TRPO(env=env,
policy=policy,
baseline=baseline,
batch_size=100,
n_itr=1)
algo.train()
def test_trpo_relu_nan(self):
env = TheanoEnv(DummyEnv())
policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=(1, ))
baseline = ZeroBaseline(env_spec=env.spec)
algo = TRPO(env=env,
policy=policy,
baseline=baseline,
n_itr=1,
batch_size=1000,
max_path_length=100,
step_size=0.001)
algo.train()
assert not np.isnan(np.sum(policy.get_param_values()))
def test_trpo_deterministic_nan():
env = TheanoEnv(DummyEnv())
policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=(1, ))
policy._l_log_std.param.set_value([np.float32(np.log(1e-8))])
baseline = ZeroBaseline(env_spec=env.spec)
algo = TRPO(env=env,
policy=policy,
baseline=baseline,
n_itr=10,
batch_size=1000,
max_path_length=100,
step_size=0.01)
algo.train()
assert not np.isnan(np.sum(policy.get_param_values()))
def test_issue_3():
"""
As reported in https://github.com/garage/garage/issues/3, the adaptive_std
parameter was not functioning properly
"""
env = CartpoleEnv()
policy = GaussianMLPPolicy(env_spec=env, adaptive_std=True)
baseline = ZeroBaseline(env_spec=env.spec)
algo = TRPO(env=env,
policy=policy,
baseline=baseline,
batch_size=100,
n_itr=1)
algo.train()
for _ in range(10):
seed = np.random.randint(1, 10000)
env_name = "SGD_nn_CartPole"
hidden_sizes = (8, )
env = TheanoEnv(normalize(CartpoleEnv()))
policy = GaussianMLPPolicy(env_spec=env.spec,
hidden_sizes=hidden_sizes)
backup_policy = GaussianMLPPolicy(env.spec,
hidden_sizes=hidden_sizes)
mix_policy = GaussianMLPPolicy(env.spec, hidden_sizes=hidden_sizes)
pos_eps_policy = GaussianMLPPolicy(env.spec,
hidden_sizes=hidden_sizes)
neg_eps_policy = GaussianMLPPolicy(env.spec,
hidden_sizes=hidden_sizes)
baseline = ZeroBaseline(env_spec=env.spec)
algo = CAPG(
env=env,
policy=policy,
backup_policy=backup_policy,
mix_policy=mix_policy,
pos_eps_policy=pos_eps_policy,
neg_eps_policy=neg_eps_policy,
n_timestep=5e5,
learning_rate=learning_rate,
batch_size=batch_size,
minibatch_size=500,
n_sub_itr=0,
center_adv=True,
baseline=baseline,
from garage.baselines import LinearFeatureBaseline from garage.baselines import ZeroBaseline from garage.envs import normalize from garage.envs.box2d import CartpoleEnv from garage.envs.mujoco import SwimmerEnv from garage.theano.envs import TheanoEnv from garage.theano.policies import GaussianMLPPolicy # normalize() makes sure that the actions for the environment lies within the # range [-1, 1] (only works for environments with continuous actions) env = TheanoEnv(normalize(SwimmerEnv())) # Initialize a neural network policy with a single hidden layer of 8 hidden # units policy = GaussianMLPPolicy(env.spec, hidden_sizes=(32, 32)) # Initialize a linear baseline estimator using default hand-crafted features baseline = ZeroBaseline(env.spec) # We will collect 100 trajectories per iteration N = 10 # Each trajectory will have at most 100 time steps T = 500 # Number of iterations n_itr = 1000 # Set the discount factor for the problem discount = 0.995 # Learning rate for the gradient update learning_rate = 0.005 # Construct the computation graph # Create a Theano variable for storing the observations We could have
