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(self):
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()))
from garage.envs import normalize
from garage.envs.box2d import CartpoleEnv
from garage.theano.envs import TheanoEnv
from garage.theano.policies import GaussianMLPPolicy
from garage.sampler import parallel_sampler
# 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(CartpoleEnv()))
# Initialize a neural network policy with a single hidden layer of 8 hidden
# units
policy = GaussianMLPPolicy(env.spec, hidden_sizes=(8, ))
parallel_sampler.populate_task(env, policy)
parallel_sampler.initialize(10)
paths = parallel_sampler.sample_paths(policy.get_param_values(), 100)
# We will collect 100 trajectories per iteration
N = 100
# Each trajectory will have at most 100 time steps
T = 100
# Number of iterations
n_itr = 100
# Set the discount factor for the problem
discount = 0.99
# Learning rate for the gradient update
learning_rate = 0.01
# Construct the computation graph
# Create a Theano variable for storing the observations We could have simply
# written `observations_var = TT.matrix('observations')` instead for this
