def _create_trajectories(self):
# Order of args for trajectory methods:
# observation, action, policy_info, reward, discount
ts0 = nest_utils.stack_nested_tensors([
trajectory.boundary((), (), (), 0., 1.),
trajectory.boundary((), (), (), 0., 1.)
])
ts1 = nest_utils.stack_nested_tensors([
trajectory.first((), (), (), 1., 1.),
trajectory.first((), (), (), 2., 1.)
])
ts2 = nest_utils.stack_nested_tensors([
trajectory.last((), (), (), 3., 1.),
trajectory.last((), (), (), 4., 1.)
])
ts3 = nest_utils.stack_nested_tensors([
trajectory.boundary((), (), (), 0., 1.),
trajectory.boundary((), (), (), 0., 1.)
])
ts4 = nest_utils.stack_nested_tensors([
trajectory.first((), (), (), 5., 1.),
trajectory.first((), (), (), 6., 1.)
])
ts5 = nest_utils.stack_nested_tensors([
trajectory.last((), (), (), 7., 1.),
trajectory.last((), (), (), 8., 1.)
])
return [ts0, ts1, ts2, ts3, ts4, ts5]
def _create_misaligned_trajectories(self):
def _concat_nested_tensors(nest1, nest2):
return tf.nest.map_structure(lambda t1, t2: tf.concat([t1, t2], axis=0),
nest1, nest2)
# Order of args for trajectory methods:
# observation, action, policy_info, reward, discount
ts1 = _concat_nested_tensors(
trajectory.first((), tf.constant([2]), (),
tf.constant([1.], dtype=tf.float32), [1.]),
trajectory.boundary((), tf.constant([1]), (),
tf.constant([0.], dtype=tf.float32), [1.]))
ts2 = _concat_nested_tensors(
trajectory.last((), tf.constant([1]), (),
tf.constant([3.], dtype=tf.float32), [1.]),
trajectory.first((), tf.constant([1]), (),
tf.constant([2.], dtype=tf.float32), [1.]))
ts3 = _concat_nested_tensors(
trajectory.boundary((), tf.constant([2]), (),
tf.constant([0.], dtype=tf.float32), [1.]),
trajectory.last((), tf.constant([1]), (),
tf.constant([4.], dtype=tf.float32), [1.]))
return [ts1, ts2, ts3]
def setUp(self):
super(BatchedPyMetricTest, self).setUp()
# Order of args for trajectory methods:
# observation, action, policy_info, reward, discount
ts0 = nest_utils.stack_nested_tensors([
trajectory.boundary((), (), (), 0., 1.),
trajectory.boundary((), (), (), 0., 1.)
])
ts1 = nest_utils.stack_nested_tensors([
trajectory.first((), (), (), 1., 1.),
trajectory.first((), (), (), 2., 1.)
])
ts2 = nest_utils.stack_nested_tensors([
trajectory.last((), (), (), 3., 1.),
trajectory.last((), (), (), 4., 1.)
])
ts3 = nest_utils.stack_nested_tensors([
trajectory.boundary((), (), (), 0., 1.),
trajectory.boundary((), (), (), 0., 1.)
])
ts4 = nest_utils.stack_nested_tensors([
trajectory.first((), (), (), 5., 1.),
trajectory.first((), (), (), 6., 1.)
])
ts5 = nest_utils.stack_nested_tensors([
trajectory.last((), (), (), 7., 1.),
trajectory.last((), (), (), 8., 1.)
])
self._ts = [ts0, ts1, ts2, ts3, ts4, ts5]
def testBatchSizeProvided(self, metric_class, expected_result):
metric = metric_class(batch_size=2)
metric(
nest_utils.stack_nested_arrays([
trajectory.boundary((), (), (), 0., 1.),
trajectory.boundary((), (), (), 0., 1.)
]))
metric(
nest_utils.stack_nested_arrays([
trajectory.first((), (), (), 1., 1.),
trajectory.first((), (), (), 1., 1.)
]))
metric(
nest_utils.stack_nested_arrays([
trajectory.mid((), (), (), 2., 1.),
trajectory.last((), (), (), 3., 0.)
]))
metric(
nest_utils.stack_nested_arrays([
trajectory.last((), (), (), 3., 0.),
trajectory.boundary((), (), (), 0., 1.)
]))
metric(
nest_utils.stack_nested_arrays([
trajectory.boundary((), (), (), 0., 1.),
trajectory.first((), (), (), 1., 1.)
]))
self.assertEqual(metric.result(), expected_result)
def setUp(self):
super(PyDriverTest, self).setUp()
f0 = np.array(0., dtype=np.float32)
f1 = np.array(1., dtype=np.float32)
# Order of args for trajectory methods:
# (observation, action, policy_info, reward, discount)
self._trajectories = [
trajectory.first(0, 1, 2, f1, f1),
trajectory.last(1, 2, 4, f1, f0),
trajectory.boundary(3, 1, 2, f0, f1),
trajectory.first(0, 1, 2, f1, f1),
trajectory.last(1, 2, 4, f1, f0),
trajectory.boundary(3, 1, 2, f0, f1),
trajectory.first(0, 1, 2, f1, f1),
]
def testAverageOneEpisode(self, metric_class, expected_result):
metric = metric_class()
metric(trajectory.boundary((), (), (), 0., 1.))
metric(trajectory.mid((), (), (), 1., 1.))
metric(trajectory.mid((), (), (), 2., 1.))
metric(trajectory.last((), (), (), 3., 0.))
self.assertEqual(expected_result, metric.result())
def testLastArrays(self):
observation = ()
action = ()
policy_info = ()
reward = np.array([1.0, 1.0, 2.0])
discount = np.array([1.0, 1.0, 1.0])
traj = trajectory.last(observation, action, policy_info, reward,
discount)
self.assertFalse(tf.is_tensor(traj.step_type))
self.assertAllEqual(traj.step_type, [ts.StepType.MID] * 3)
self.assertAllEqual(traj.next_step_type, [ts.StepType.LAST] * 3)
def testLastTensors(self):
observation = ()
action = ()
policy_info = ()
reward = tf.constant([1.0, 1.0, 2.0])
discount = tf.constant([1.0, 1.0, 1.0])
traj = trajectory.last(observation, action, policy_info, reward,
discount)
self.assertTrue(tf.is_tensor(traj.step_type))
traj_val = self.evaluate(traj)
self.assertAllEqual(traj_val.step_type, [ts.StepType.MID] * 3)
self.assertAllEqual(traj_val.next_step_type, [ts.StepType.LAST] * 3)
def testAverageTwoEpisode(self, metric_class, expected_result):
metric = metric_class()
metric(trajectory.boundary((), (), (), 0., 1.))
metric(trajectory.first((), (), (), 1., 1.))
metric(trajectory.mid((), (), (), 2., 1.))
metric(trajectory.last((), (), (), 3., 0.))
metric(trajectory.boundary((), (), (), 0., 1.))
# TODO(kbanoop): Add optional next_step_type arg to trajectory.first. Or
# implement trajectory.first_last().
metric(
trajectory.Trajectory(ts.StepType.FIRST, (), (), (),
ts.StepType.LAST, -6., 1.))
self.assertEqual(expected_result, metric.result())
def testAverageOneEpisodeWithReset(self, metric_class, expected_result):
metric = metric_class()
metric(trajectory.first((), (), (), 0., 1.))
metric(trajectory.mid((), (), (), 1., 1.))
metric(trajectory.mid((), (), (), 2., 1.))
# The episode is reset.
#
# This could happen when using the dynamic_episode_driver with
# parallel_py_environment. When the parallel episodes are of different
# lengths and num_episodes is reached, some episodes would be left in "MID".
# When the driver runs again, all environments are reset at the beginning
# of the tf.while_loop and the unfinished episodes would get "FIRST" without
# seeing "LAST".
metric(trajectory.first((), (), (), 3., 1.))
metric(trajectory.last((), (), (), 4., 1.))
self.assertEqual(expected_result, metric.result())
def testSaveRestore(self):
metrics = [
py_metrics.AverageReturnMetric(),
py_metrics.AverageEpisodeLengthMetric(),
py_metrics.EnvironmentSteps(),
py_metrics.NumberOfEpisodes()
]
for metric in metrics:
metric(trajectory.boundary((), (), (), 0., 1.))
metric(trajectory.mid((), (), (), 1., 1.))
metric(trajectory.mid((), (), (), 2., 1.))
metric(trajectory.last((), (), (), 3., 0.))
checkpoint = tf.train.Checkpoint(**{m.name: m for m in metrics})
prefix = self.get_temp_dir() + '/ckpt'
save_path = checkpoint.save(prefix)
for metric in metrics:
metric.reset()
self.assertEqual(0, metric.result())
checkpoint.restore(save_path).assert_consumed()
for metric in metrics:
self.assertGreater(metric.result(), 0)