def test_get_obs(self):
"""Check the functionality of the _get_obs() method.
This method is tested for three cases:
1. when the context is None
2. for 1-D observations and contexts
3. for 2-D observations and contexts
"""
policy = ActorCriticPolicy(**self.policy_params)
# test case 1
obs = np.array([0, 1, 2])
context = None
expected = obs
np.testing.assert_almost_equal(policy._get_obs(obs, context), expected)
# test case 2
obs = np.array([0, 1, 2])
context = np.array([3, 4])
expected = np.array([0, 1, 2, 3, 4])
np.testing.assert_almost_equal(policy._get_obs(obs, context), expected)
# test case 3
obs = np.array([[0, 1, 2]])
context = np.array([[3, 4]])
expected = np.array([[0, 1, 2, 3, 4]])
np.testing.assert_almost_equal(policy._get_obs(obs, context, axis=1),
expected)
def test_setup_target_updates(self):
"""Check the functionality of the _setup_target_updates() method.
This test validates both the init and soft update procedures generated
by the tested method.
"""
policy = ActorCriticPolicy(**self.policy_params)
_ = tf.Variable(initial_value=[[1, 1, 1, 1]],
dtype=tf.float32,
name="0")
val1 = tf.Variable(initial_value=[[0, 0, 0, 0]],
dtype=tf.float32,
name="1")
# Initialize the variables of the policy.
policy.sess.run(tf.compat.v1.global_variables_initializer())
init, soft = policy._setup_target_updates("0", "1", None, 0.1, 0)
# test soft update
policy.sess.run(soft)
expected = np.array([[0.1, 0.1, 0.1, 0.1]])
np.testing.assert_almost_equal(policy.sess.run(val1), expected)
# test init update
policy.sess.run(init)
expected = np.array([[1, 1, 1, 1]])
np.testing.assert_almost_equal(policy.sess.run(val1), expected)
def test_init(self):
"""Validate that the variables are initialized properly."""
policy = ActorCriticPolicy(**self.policy_params)
# Check that the abstract class has all the required attributes.
self.assertEqual(policy.sess, self.policy_params['sess'])
self.assertEqual(policy.ac_space, self.policy_params['ac_space'])
self.assertEqual(policy.ob_space, self.policy_params['ob_space'])
self.assertEqual(policy.co_space, self.policy_params['co_space'])
self.assertEqual(policy.buffer_size, self.policy_params['buffer_size'])
self.assertEqual(policy.batch_size, self.policy_params['batch_size'])
self.assertEqual(policy.actor_lr, self.policy_params['actor_lr'])
self.assertEqual(policy.critic_lr, self.policy_params['critic_lr'])
self.assertEqual(policy.verbose, self.policy_params['verbose'])
self.assertEqual(policy.tau, self.policy_params['tau'])
self.assertEqual(policy.gamma, self.policy_params['gamma'])
self.assertEqual(policy.use_huber, self.policy_params['use_huber'])
# Check that the abstract class has all the required methods.
self.assertRaises(NotImplementedError, policy.initialize)
self.assertRaises(NotImplementedError, policy.update,
update_actor=None)
self.assertRaises(NotImplementedError, policy.get_action,
obs=None, context=None, apply_noise=None,
random_actions=None)
self.assertRaises(NotImplementedError, policy.store_transition,
obs0=None, context0=None, action=None, reward=None,
obs1=None, context1=None, done=None,
is_final_step=None, evaluate=False)
self.assertRaises(NotImplementedError, policy.get_td_map)
def test_get_ob_dim(self):
"""Check the functionality of the _get_ob_dim() method.
This method is tested for two cases:
1. when the context is None
2. when the context is not None
"""
policy = ActorCriticPolicy(**self.policy_params)
# test case 1
ob_space = Box(0, 1, shape=(2, ), dtype=np.float32)
co_space = None
self.assertTupleEqual(policy._get_ob_dim(ob_space, co_space), (2, ))
# test case 2
ob_space = Box(0, 1, shape=(2, ), dtype=np.float32)
co_space = Box(0, 1, shape=(3, ), dtype=np.float32)
self.assertTupleEqual(policy._get_ob_dim(ob_space, co_space), (5, ))
def test_remove_fingerprint(self):
"""Check the functionality of the _remove_fingerprint() method.
This method is tested for two cases:
1. for an additional_dim of zero
2. for an additional_dim greater than zero
"""
policy = ActorCriticPolicy(**self.policy_params)
# test case 1
val = tf.constant(value=[[1, 2, 3, 4]], dtype=tf.float32)
new_val = policy._remove_fingerprint(val, 4, 2, 0)
expected = np.array([[1, 2, 0, 0]])
np.testing.assert_almost_equal(policy.sess.run(new_val), expected)
# test case 2
val = tf.constant(value=[[1, 2, 3, 4]], dtype=tf.float32)
new_val = policy._remove_fingerprint(val, 3, 2, 1)
expected = np.array([[1, 0, 0, 4]])
np.testing.assert_almost_equal(policy.sess.run(new_val), expected)