def __init__(self, observable, physics, random_state,
strip_singleton_buffer_dim, pad_with_initial_value):
self.observable = observable
self.observation_callable = (observable.observation_callable(
physics, random_state))
self._bind_attribute_from_observable('update_interval',
DEFAULT_UPDATE_INTERVAL,
random_state)
self._bind_attribute_from_observable('delay', DEFAULT_DELAY,
random_state)
self._bind_attribute_from_observable('buffer_size',
DEFAULT_BUFFER_SIZE, random_state)
obs_spec = self.observable.array_spec
if obs_spec is None:
# We take an observation to determine the shape and dtype of the array.
# This occurs outside of an episode and doesn't affect environment
# behavior. At this point the physics state is not guaranteed to be valid,
# so we might get a `PhysicsError` if the observation callable calls
# `physics.forward`. We suppress such errors since they do not matter as
# far as the shape and dtype of the observation are concerned.
with physics.suppress_physics_errors():
obs_array = self.observation_callable()
obs_array = np.asarray(obs_array)
obs_spec = specs.Array(shape=obs_array.shape,
dtype=obs_array.dtype)
self.buffer = obs_buffer.Buffer(
buffer_size=self.buffer_size,
shape=obs_spec.shape,
dtype=obs_spec.dtype,
pad_with_initial_value=pad_with_initial_value,
strip_singleton_buffer_dim=strip_singleton_buffer_dim)
self.update_schedule = collections.deque()
def testStripSingletonDimension(self, shape):
buf = obs_buffer.Buffer(buffer_size=1,
shape=shape,
dtype=np.float,
strip_singleton_buffer_dim=True)
expected_value = np.full(shape, 42, dtype=np.float)
buf.insert(timestamp=0, delay=0, value=expected_value)
np.testing.assert_array_equal(buf.read(current_time=1), expected_value)
def testOutOfOrderArrival(self): buf = obs_buffer.Buffer(buffer_size=3, shape=(), dtype=float) buf.insert(timestamp=0, delay=4, value=1) buf.insert(timestamp=1, delay=2, value=2) buf.insert(timestamp=2, delay=3, value=3) np.testing.assert_array_equal(buf.read(current_time=2), [0., 0., 0.]) np.testing.assert_array_equal(buf.read(current_time=3), [0., 0., 2.]) np.testing.assert_array_equal(buf.read(current_time=4), [0., 2., 1.]) np.testing.assert_array_equal(buf.read(current_time=5), [2., 1., 3.]) np.testing.assert_array_equal(buf.read(current_time=6), [2., 1., 3.])
def testPlanTwoStepsAhead(self):
buf = obs_buffer.Buffer(
buffer_size=1, shape=(), dtype=np.float)
control_timestep = 5
observation_schedule = _generate_constant_schedule(
update_timestep=2, delay=3,
control_timestep=control_timestep, n_observed_steps=2)
buf.drop_unobserved_upcoming_items(
observation_schedule, read_interval=control_timestep)
self.assertEqual(observation_schedule, [(2, 3), (6, 3), (10, 3)])
def testPlanToSingleUndelayedObservation(self):
buf = obs_buffer.Buffer(
buffer_size=1, shape=(), dtype=np.float)
control_timestep = 20
observation_schedule = _generate_constant_schedule(
update_timestep=1, delay=0,
control_timestep=control_timestep, n_observed_steps=1)
buf.drop_unobserved_upcoming_items(
observation_schedule, read_interval=control_timestep)
self.assertEqual(observation_schedule, [(20, 0)])
def __init__(self, observable, physics, random_state,
strip_singleton_buffer_dim):
self.observable = observable
# We take an observation here to determine the shape and size.
# This occurs outside of an episode and doesn't affect environment behavior.
obs_value = np.array(
observable.observation_callable(physics, random_state)())
self.buffer = obs_buffer.Buffer(
buffer_size=(observable.buffer_size or DEFAULT_BUFFER_SIZE),
shape=obs_value.shape,
dtype=obs_value.dtype,
strip_singleton_buffer_dim=strip_singleton_buffer_dim)
self.update_schedule = collections.deque()
