def test_vector_dataset_with_continuous_action():
ref, _ = get_pendulum()
# save as csv
export_mdp_dataset_as_csv(ref, "test_data/test.csv")
# load from csv
dataset = import_csv_as_mdp_dataset("test_data/test.csv")
assert dataset.get_observation_shape() == ref.get_observation_shape()
assert dataset.get_action_size() == ref.get_action_size()
assert np.allclose(dataset.observations, ref.observations)
assert np.allclose(dataset.actions, ref.actions)
assert np.allclose(dataset.rewards, ref.rewards)
assert len(ref) == len(dataset)
assert not dataset.is_action_discrete()
def algo_pendulum_tester(algo, n_evaluations=100, n_episodes=500, n_trials=3):
# load dataset
dataset, env = get_pendulum()
upper_bound = env.action_space.high
# try multiple trials to reduce failures due to random seeds
trial_count = 0
for _ in range(n_trials):
# reset parameters
algo._impl = None
# train
algo.fit(
dataset.episodes[:n_episodes],
n_epochs=3,
logdir="test_data",
verbose=False,
tensorboard=False,
)
# evaluation loop
success_count = 0
evaluation_count = 0
while evaluation_count < n_evaluations:
observation = env.reset()
episode_rew = 0.0
while True:
action = algo.predict([observation])[0]
observation, reward, done, _ = env.step(upper_bound * action)
episode_rew += reward
if done:
break
evaluation_count += 1
if episode_rew >= -600:
success_count += 1
if success_count >= n_evaluations * 0.8:
break
trial_count += 1
if trial_count == n_trials:
assert False, "performance is not good enough: %d." % success_count
def test_get_pendulum(dataset_type):
get_pendulum(dataset_type=dataset_type)
from d3rlpy.datasets import get_pendulum from minerva.dataset import export_mdp_dataset_as_csv # prepare MDPDataset dataset, _ = get_pendulum() # save as CSV export_mdp_dataset_as_csv(dataset, 'pendulum.csv')
from d3rlpy.datasets import get_pendulum from minerva.dataset import export_mdp_dataset_as_csv # prepare MDPDataset dataset, _ = get_pendulum(dataset_type="random") # save as CSV export_mdp_dataset_as_csv(dataset, 'pendulum_random.csv')