def test_serialize_deserialize(self):
policy_1 = FeedforwardGaussianPolicy(
input_shapes=self.env.observation_shape,
output_shape=self.env.action_space.shape,
action_range=(
self.env.action_space.low,
self.env.action_space.high,
),
hidden_layer_sizes=self.hidden_layer_sizes,
observation_keys=self.env.observation_keys)
path = sampler_utils.rollout(self.env,
policy_1,
path_length=10,
break_on_terminal=False)
observations = path['observations']
weights_1 = policy_1.get_weights()
actions_1 = policy_1.actions(observations)
log_pis_1 = policy_1.log_probs(observations, actions_1)
config = policies.serialize(policy_1)
policy_2 = policies.deserialize(config)
policy_2.set_weights(policy_1.get_weights())
weights_2 = policy_2.get_weights()
log_pis_2 = policy_2.log_probs(observations, actions_1)
for weight_1, weight_2 in zip(weights_1, weights_2):
np.testing.assert_array_equal(weight_1, weight_2)
np.testing.assert_array_equal(log_pis_1, log_pis_2)
np.testing.assert_equal(actions_1.shape,
policy_2.actions(observations).shape)
class GaussianPolicyTest(tf.test.TestCase):
def setUp(self):
self.env = get_environment('gym', 'Swimmer', 'v3', {})
self.hidden_layer_sizes = (16, 16)
self.policy = FeedforwardGaussianPolicy(
input_shapes=self.env.observation_shape,
output_shape=self.env.action_space.shape,
action_range=(
self.env.action_space.low,
self.env.action_space.high,
),
hidden_layer_sizes=self.hidden_layer_sizes,
observation_keys=self.env.observation_keys)
def test_actions_and_log_pis_symbolic(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = {}
for key in observation1_np.keys():
observations_np[key] = np.stack(
(observation1_np[key],
observation2_np[key])).astype(np.float32)
observations_np = flatten_input_structure(observations_np)
observations_tf = [
tf.constant(x, dtype=tf.float32) for x in observations_np
]
actions = self.policy.actions(observations_tf)
log_pis = self.policy.log_pis(observations_tf, actions)
self.assertEqual(actions.shape, (2, *self.env.action_space.shape))
self.assertEqual(log_pis.shape, (2, 1))
self.evaluate(tf.compat.v1.global_variables_initializer())
actions_np = self.evaluate(actions)
log_pis_np = self.evaluate(log_pis)
self.assertEqual(actions_np.shape, (2, *self.env.action_space.shape))
self.assertEqual(log_pis_np.shape, (2, 1))
def test_actions_and_log_pis_numeric(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = {}
for key in observation1_np.keys():
observations_np[key] = np.stack(
(observation1_np[key],
observation2_np[key])).astype(np.float32)
observations_np = flatten_input_structure(observations_np)
actions_np = self.policy.actions_np(observations_np)
log_pis_np = self.policy.log_pis_np(observations_np, actions_np)
self.assertEqual(actions_np.shape, (2, *self.env.action_space.shape))
self.assertEqual(log_pis_np.shape, (2, 1))
def test_env_step_with_actions(self):
observation_np = self.env.reset()
observations_np = flatten_input_structure(
{key: value[None, :]
for key, value in observation_np.items()})
action = self.policy.actions_np(observations_np)[0, ...]
self.env.step(action)
def test_trainable_variables(self):
self.assertEqual(tuple(self.policy.trainable_variables),
tuple(self.policy.actions_model.trainable_variables))
self.assertEqual(tuple(self.policy.trainable_variables),
tuple(self.policy.log_pis_model.trainable_variables))
self.assertEqual(
tuple(self.policy.trainable_variables),
tuple(self.policy.deterministic_actions_model.trainable_variables))
self.assertEqual(len(self.policy.trainable_variables),
2 * (len(self.hidden_layer_sizes) + 1))
def test_get_diagnostics(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = {}
for key in observation1_np.keys():
observations_np[key] = np.stack(
(observation1_np[key],
observation2_np[key])).astype(np.float32)
observations_np = flatten_input_structure(observations_np)
diagnostics = self.policy.get_diagnostics(observations_np)
self.assertTrue(isinstance(diagnostics, OrderedDict))
self.assertEqual(
tuple(diagnostics.keys()),
('shifts-mean', 'shifts-std', 'log_scale_diags-mean',
'log_scale_diags-std', '-log-pis-mean', '-log-pis-std',
'raw-actions-mean', 'raw-actions-std', 'actions-mean',
'actions-std', 'actions-min', 'actions-max'))
for value in diagnostics.values():
self.assertTrue(np.isscalar(value))
def test_serialize_deserialize(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = {}
for key in observation1_np.keys():
observations_np[key] = np.stack(
(observation1_np[key],
observation2_np[key])).astype(np.float32)
observations_np = flatten_input_structure(observations_np)
weights = self.policy.get_weights()
actions_np = self.policy.actions_np(observations_np)
log_pis_np = self.policy.log_pis_np(observations_np, actions_np)
serialized = pickle.dumps(self.policy)
deserialized = pickle.loads(serialized)
weights_2 = deserialized.get_weights()
log_pis_np_2 = deserialized.log_pis_np(observations_np, actions_np)
for weight, weight_2 in zip(weights, weights_2):
np.testing.assert_array_equal(weight, weight_2)
np.testing.assert_array_equal(log_pis_np, log_pis_np_2)
np.testing.assert_equal(actions_np.shape,
deserialized.actions_np(observations_np).shape)
def test_latent_smoothing(self):
observation_np = self.env.reset()
smoothed_policy = FeedforwardGaussianPolicy(
input_shapes=self.env.observation_shape,
output_shape=self.env.action_space.shape,
action_range=(
self.env.action_space.low,
self.env.action_space.high,
),
hidden_layer_sizes=self.hidden_layer_sizes,
smoothing_coefficient=0.5,
observation_keys=self.env.observation_keys)
np.testing.assert_equal(smoothed_policy._smoothing_x, 0.0)
self.assertEqual(smoothed_policy._smoothing_alpha, 0.5)
self.assertEqual(
smoothed_policy._smoothing_beta,
np.sqrt((1.0 - np.power(smoothed_policy._smoothing_alpha, 2.0))) /
(1.0 - smoothed_policy._smoothing_alpha))
smoothing_x_previous = smoothed_policy._smoothing_x
for i in range(5):
observations_np = flatten_input_structure(
{key: value[None, :]
for key, value in observation_np.items()})
action_np = smoothed_policy.actions_np(observations_np)[0]
observation_np = self.env.step(action_np)[0]
self.assertFalse(
np.all(
np.equal(smoothing_x_previous,
smoothed_policy._smoothing_x)))
smoothing_x_previous = smoothed_policy._smoothing_x
smoothed_policy.reset()
np.testing.assert_equal(smoothed_policy._smoothing_x, 0.0)
class GaussianPolicyTest(tf.test.TestCase):
def setUp(self):
self.env = gym.envs.make('Swimmer-v2')
self.hidden_layer_sizes = (128, 128)
self.policy = FeedforwardGaussianPolicy(
input_shapes=(self.env.observation_space.shape, ),
output_shape=self.env.action_space.shape,
hidden_layer_sizes=self.hidden_layer_sizes)
def test_actions_and_log_pis_symbolic(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = np.stack((observation1_np, observation2_np))
observations_tf = tf.constant(observations_np, dtype=tf.float32)
actions = self.policy.actions([observations_tf])
log_pis = self.policy.log_pis([observations_tf], actions)
self.assertEqual(actions.shape, (2, *self.env.action_space.shape))
self.assertEqual(log_pis.shape, (2, 1))
self.evaluate(tf.global_variables_initializer())
actions_np = self.evaluate(actions)
log_pis_np = self.evaluate(log_pis)
self.assertEqual(actions_np.shape, (2, *self.env.action_space.shape))
self.assertEqual(log_pis_np.shape, (2, 1))
def test_actions_and_log_pis_numeric(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = np.stack((observation1_np, observation2_np))
actions_np = self.policy.actions_np([observations_np])
log_pis_np = self.policy.log_pis_np([observations_np], actions_np)
self.assertEqual(actions_np.shape, (2, *self.env.action_space.shape))
self.assertEqual(log_pis_np.shape, (2, 1))
def test_env_step_with_actions(self):
observation1_np = self.env.reset()
action = self.policy.actions_np(observation1_np[None])[0, ...]
self.env.step(action)
def test_trainable_variables(self):
self.assertEqual(len(self.policy.trainable_variables),
2 * (len(self.hidden_layer_sizes) + 1))
def test_get_diagnostics(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = np.stack((observation1_np, observation2_np))
diagnostics = self.policy.get_diagnostics([observations_np])
self.assertTrue(isinstance(diagnostics, OrderedDict))
self.assertEqual(tuple(diagnostics.keys()),
('shifts-mean', 'shifts-std', 'log_scale_diags-mean',
'log_scale_diags-std', '-log-pis-mean',
'-log-pis-std', 'raw-actions-mean',
'raw-actions-std', 'actions-mean', 'actions-std'))
for value in diagnostics.values():
self.assertTrue(np.isscalar(value))
def test_serialize_deserialize(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = np.stack(
(observation1_np, observation2_np)).astype(np.float32)
weights = self.policy.get_weights()
actions_np = self.policy.actions_np([observations_np])
log_pis_np = self.policy.log_pis_np([observations_np], actions_np)
serialized = pickle.dumps(self.policy)
deserialized = pickle.loads(serialized)
weights_2 = deserialized.get_weights()
log_pis_np_2 = deserialized.log_pis_np([observations_np], actions_np)
for weight, weight_2 in zip(weights, weights_2):
np.testing.assert_array_equal(weight, weight_2)
np.testing.assert_array_equal(log_pis_np, log_pis_np_2)
np.testing.assert_equal(
actions_np.shape,
deserialized.actions_np([observations_np]).shape)
def test_latent_smoothing(self):
observation_np = self.env.reset()
smoothed_policy = FeedforwardGaussianPolicy(
input_shapes=(self.env.observation_space.shape, ),
output_shape=self.env.action_space.shape,
hidden_layer_sizes=self.hidden_layer_sizes,
smoothing_coefficient=0.5)
np.testing.assert_equal(smoothed_policy._smoothing_x, 0.0)
self.assertEqual(smoothed_policy._smoothing_alpha, 0.5)
self.assertEqual(
smoothed_policy._smoothing_beta,
np.sqrt((1.0 - np.power(smoothed_policy._smoothing_alpha, 2.0))) /
(1.0 - smoothed_policy._smoothing_alpha))
smoothing_x_previous = smoothed_policy._smoothing_x
for i in range(5):
action_np = smoothed_policy.actions_np([observation_np[None, :]
])[0]
observation_np = self.env.step(action_np)[0]
self.assertFalse(
np.all(
np.equal(smoothing_x_previous,
smoothed_policy._smoothing_x)))
smoothing_x_previous = smoothed_policy._smoothing_x
smoothed_policy.reset()
np.testing.assert_equal(smoothed_policy._smoothing_x, 0.0)