def test_batched_step(self):
"""
Make sure that the outputs of batched steps are
the correct shape.
"""
with self.graph.as_default():
env = TupleCartPole()
try:
obs_space = env.observation_space
finally:
env.close()
batch_size = 7
in_obses = [obs_space.sample() for _ in range(batch_size)]
in_states = self.model.start_state(batch_size)
outs = self.model.step(in_obses, in_states)
self.test_case.assertEqual(len(outs['actions']), batch_size)
self.test_case.assertEqual(_state_shape_uid(outs['states']),
_state_shape_uid(in_states))
if 'action_params' in outs:
param_shape = self.model.action_dist.param_shape
self.test_case.assertEqual(
np.array(outs['action_params']).shape,
(batch_size, ) + param_shape)
if 'values' in outs:
self.test_case.assertEqual(
np.array(outs['values']).shape, (batch_size, ))
def run_ac_test(maker):
"""
Run a test given a model constructor.
"""
env = TupleCartPole()
try:
spaces = gym_spaces(env)
finally:
env.close()
ModelTester(lambda sess: maker(sess, *spaces)).test_all()
def __init__(self, *args, **kwargs):
super(ACTest, self).__init__(*args, **kwargs)
self.session = tf.Session()
env = TupleCartPole()
try:
action_space = env.action_space
observation_space = env.observation_space
finally:
env.close()
self.action_dist = gym_space_distribution(action_space)
self.obs_vectorizer = gym_space_vectorizer(observation_space)
def run_ac_test(maker):
"""
Run a test given a model constructor.
"""
env = TupleCartPole()
try:
action_space = env.action_space
observation_space = env.observation_space
finally:
env.close()
action_dist = gym_space_distribution(action_space)
obs_vectorizer = gym_space_vectorizer(observation_space)
ModelTester(
lambda sess: maker(sess, action_dist, obs_vectorizer)).test_all()
def _test_batches_consistency(self, batch_size, trunc_start):
"""
Make sure that batches() produces the same outputs
that we got with step().
"""
env = TupleCartPole()
try:
roller = BasicRoller(env, self.model, min_episodes=7)
rollouts = roller.rollouts()
if trunc_start:
rollouts = self._truncate_first(rollouts)
num_batches = 10
for batch in self.model.batches(rollouts, batch_size=batch_size):
num_batches -= 1
if num_batches == 0:
break
self._test_batch(rollouts, batch)
finally:
env.close()