def test_update_policy_and_value_function(self):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = ActorCritic('', observation_space, action_space)
# Set up.
self._setup_trajectory(sut)
sut._process_accumulated_trajectory(True)
sut._trainer = MagicMock()
sut._adjust_learning_rate = MagicMock()
# Call test method.
sut._update_networks()
# Verify value network behavior.
self.assertEqual(sut._trainer.train_minibatch.call_count, 1)
call_args = sut._trainer.train_minibatch.call_args
np.testing.assert_array_equal(
call_args[0][0][sut._input_variables],
[np.array([0.1], np.float32),
np.array([0.2], np.float32)])
np.testing.assert_array_almost_equal(
call_args[0][0][sut._value_network_output_variables],
[[2.9975], [3.05]])
np.testing.assert_array_equal(
call_args[0][0][sut._policy_network_output_variables],
[np.array([1, 0], np.float32),
np.array([0, 1], np.float32)])
np.testing.assert_array_almost_equal(
call_args[0][0][sut._policy_network_weight_variables],
[[0.9975], [2.05]])
# Verify data buffer size.
self.assertEqual(len(sut._input_buffer), 0)
def test_init_dqn_preprocessing(self,
mock_parameters,
mock_replay_memory):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.preprocessing = \
'cntk.contrib.deeprl.agent.shared.preprocessing.AtariPreprocessing'
mock_parameters.return_value.preprocessing_args = '()'
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
# Preprocessor with default arguments.
self.assertIsNotNone(sut._preprocessor)
self.assertEqual(sut._preprocessor.output_shape(), (4, 84, 84))
# Preprocessor with arguments passed as a tuple.
mock_parameters.return_value.preprocessing_args = '(3,)'
sut = QLearning('', observation_space, action_space)
self.assertEqual(sut._preprocessor.output_shape(), (3, 84, 84))
# Preprocessor with inappropriate arguments.
mock_parameters.return_value.preprocessing_args = '(3, 4)'
self.assertRaises(
TypeError, QLearning, '', observation_space, action_space)
# Undefined preprocessor.
mock_parameters.return_value.preprocessing = 'undefined'
self.assertRaises(
ValueError, QLearning, '', observation_space, action_space)
def test_init(self):
# Discrete observation space.
action_space = spaces.Discrete(2)
observation_space = spaces.Discrete(3)
sut = TabularQLearning('', observation_space, action_space)
self.assertEqual(sut._num_actions, 2)
self.assertEqual(sut._num_states, 3)
self.assertEqual(sut._shape_of_inputs, (3, ))
self.assertTrue(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
# Discretize observation space to default resolution.
observation_space = spaces.Box(0, 1, (2, ))
sut = TabularQLearning('', observation_space, action_space)
self.assertEqual(sut._num_states, 100)
self.assertEqual(sut._shape_of_inputs, (100, ))
self.assertTrue(sut._discrete_observation_space)
self.assertIsNotNone(sut._space_discretizer)
# Verify encoding of state
self.assertEqual(sut._discretize_state_if_necessary([0, 0]), 0)
self.assertEqual(sut._discretize_state_if_necessary([0.05, 0]), 0)
self.assertEqual(sut._discretize_state_if_necessary([0.95, 0]), 90)
self.assertEqual(sut._discretize_state_if_necessary([0, 0.05]), 0)
self.assertEqual(sut._discretize_state_if_necessary([0, 0.95]), 9)
self.assertEqual(sut._discretize_state_if_necessary([0.1, 0.2]), 12)
self.assertEqual(sut._discretize_state_if_necessary([1, 1]), 99)
# Unsupported observation space for tabular representation
observation_space = spaces.MultiBinary(10)
self.assertRaises(ValueError, TabularQLearning, '', observation_space,
action_space)
def test_init_shared_representation(self, mock_parameters):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.shared_representation = True
sut = ActorCritic('', observation_space, action_space)
self.assertEqual(sut._num_actions, 2)
self.assertIsNone(sut._num_states)
self.assertEqual(sut._shape_of_inputs, (1, ))
self.assertFalse(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
self.assertTrue(
set(sut._policy_network.parameters).issubset(
set(sut._value_network.parameters)))
diff = set(sut._value_network.parameters).difference(
set(sut._policy_network.parameters))
# one for W and one for b
self.assertEqual(len(diff), 2)
shapes = []
for item in diff:
shapes.append(item.shape)
self.assertEqual(set(shapes), {(2, 1), (1, )})
def test_update_q_dqn(self,
mock_parameters,
mock_replay_memory):
self._setup_parameters(mock_parameters.return_value)
self._setup_replay_memory(mock_replay_memory.return_value)
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
sut._q.eval = \
MagicMock(return_value=np.array([[[0.2, 0.1]]], np.float32))
sut._target_q.eval = \
MagicMock(return_value=np.array([[[0.3, 0.4]]], np.float32))
sut._trainer = MagicMock()
sut._update_q_periodically()
np.testing.assert_array_equal(
sut._trainer.train_minibatch.call_args[0][0][sut._input_variables],
[np.array([0.1], np.float32)])
# 10 (reward) + 0.9 (gamma) x 0.4 (max q_target) -> update action 0
np.testing.assert_array_equal(
sut._trainer.train_minibatch.call_args[0][0][sut._output_variables],
[np.array([10.36, 0.1], np.float32)])
def test_rollout_preprocess(self, mock_parameters):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.preprocessing = \
'cntk.contrib.deeprl.agent.shared.preprocessing.SlidingWindow'
mock_parameters.return_value.preprocessing_args = '(2, "float32")'
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = ActorCritic('', observation_space, action_space)
sut._choose_action = Mock(side_effect=[(0, ''), (1, ''), (1, '')])
sut.start(np.array([0.1], np.float32))
sut.step(0.1, np.array([0.2], np.float32))
sut.step(0.2, np.array([0.3], np.float32))
self.assertEqual(sut._trajectory_rewards, [0.1, 0.2])
self.assertEqual(sut._trajectory_actions, [0, 1, 1])
np.testing.assert_array_equal(sut._trajectory_states, [
np.array([[0], [0.1]], np.float32),
np.array([[0.1], [0.2]], np.float32),
np.array([[0.2], [0.3]], np.float32)
])
sut.end(0.3, np.array([0.4], np.float32))
self.assertEqual(sut._trajectory_rewards, [0.1, 0.2, 0.3])
self.assertEqual(sut._trajectory_actions, [0, 1, 1])
np.testing.assert_array_equal(sut._trajectory_states, [
np.array([[0], [0.1]], np.float32),
np.array([[0.1], [0.2]], np.float32),
np.array([[0.2], [0.3]], np.float32)
])
def test_init_unsupported_q(self, mock_parameters):
instance = mock_parameters.return_value
instance.q_representation = 'undefined'
instance.preprocessing = ''
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
self.assertRaises(
ValueError, QLearning, '', observation_space, action_space)
def test_update_q(self, mock_parameters, mock_replay_memory):
"""Test if _update_q_periodically() can finish successfully."""
self._setup_parameters(mock_parameters.return_value)
self._setup_replay_memory(mock_replay_memory.return_value)
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = QLearning('', observation_space, action_space)
sut._trainer.train_minibatch = MagicMock()
sut._choose_action = MagicMock(side_effect=[
(1, 'GREEDY'),
(0, 'GREEDY'),
(1, 'RANDOM'),
])
action, debug_info = sut.start(np.array([0.1], np.float32))
self.assertEqual(action, 1)
self.assertEqual(debug_info['action_behavior'], 'GREEDY')
self.assertEqual(sut.episode_count, 1)
self.assertEqual(sut.step_count, 0)
self.assertEqual(sut._epsilon, 0.1)
self.assertEqual(sut._trainer.parameter_learners[0].learning_rate(),
0.1)
self.assertEqual(sut._last_state, np.array([0.1], np.float32))
self.assertEqual(sut._last_action, 1)
action, debug_info = sut.step(1, np.array([0.2], np.float32))
self.assertEqual(action, 0)
self.assertEqual(debug_info['action_behavior'], 'GREEDY')
self.assertEqual(sut.episode_count, 1)
self.assertEqual(sut.step_count, 1)
self.assertEqual(sut._epsilon, 0.09)
# learning rate remains 0.1 as Q is not updated during this time step.
self.assertEqual(sut._trainer.parameter_learners[0].learning_rate(),
0.1)
self.assertEqual(sut._last_state, np.array([0.2], np.float32))
self.assertEqual(sut._last_action, 0)
action, debug_info = sut.step(2, np.array([0.3], np.float32))
self.assertEqual(action, 1)
self.assertEqual(debug_info['action_behavior'], 'RANDOM')
self.assertEqual(sut.episode_count, 1)
self.assertEqual(sut.step_count, 2)
self.assertEqual(sut._epsilon, 0.08)
self.assertEqual(sut._trainer.parameter_learners[0].learning_rate(),
0.08)
self.assertEqual(sut._last_state, np.array([0.3], np.float32))
self.assertEqual(sut._last_action, 1)
sut.end(3, np.array([0.4], np.float32))
self.assertEqual(sut.episode_count, 1)
self.assertEqual(sut.step_count, 3)
self.assertEqual(sut._epsilon, 0.08)
# learning rate remains 0.08 as Q is not updated during this time step.
self.assertEqual(sut._trainer.parameter_learners[0].learning_rate(),
0.08)
def test_init_dqn_huber_loss(self, mock_parameters, mock_model):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.use_error_clipping = True
mock_model.return_value = self._setup_test_model()
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
mock_model.assert_called_with((1,), 2, '[2]', huber_loss)
def test_scalar(self):
s = spaces.Box(0, 1, (2, ))
sut = BoxSpaceDiscretizer(s, 10)
self.assertEqual(sut.discretize([0, 0]), 0)
self.assertEqual(sut.discretize([0.05, 0]), 0)
self.assertEqual(sut.discretize([0.95, 0]), 90)
self.assertEqual(sut.discretize([0, 0.05]), 0)
self.assertEqual(sut.discretize([0, 0.95]), 9)
self.assertEqual(sut.discretize([0.1, 0.2]), 12)
self.assertEqual(sut.discretize([1, 1]), 99)
def test_init_dqn_prioritized_replay(self, mock_parameters,
mock_replay_memory):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.use_prioritized_replay = True
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = QLearning('', observation_space, action_space)
self.assertIsNotNone(sut._weight_variables)
mock_replay_memory.assert_called_with(100, True)
def test_init_box_observation_space(self):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = FakeAgentBaseClass(observation_space, action_space)
self.assertEqual(sut._num_actions, 2)
self.assertIsNone(sut._num_states)
self.assertEqual(sut._shape_of_inputs, (1, ))
self.assertFalse(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
def test_update_q_dqn_prioritized_replay(self,
mock_parameters,
mock_replay_memory):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.use_prioritized_replay = True
self._setup_prioritized_replay_memory(mock_replay_memory.return_value)
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
def new_q_value(self):
return np.array([[[0.2, 0.1]]], np.float32)
sut._q.eval = MagicMock(side_effect=new_q_value)
sut._target_q.eval = MagicMock(
return_value=np.array([[[0.3, 0.4]]], np.float32))
sut._trainer = MagicMock()
sut._update_q_periodically()
self.assertEqual(sut._trainer.train_minibatch.call_count, 1)
np.testing.assert_array_equal(
sut._trainer.train_minibatch.call_args[0][0][sut._input_variables],
[
np.array([0.1], np.float32),
np.array([0.3], np.float32),
np.array([0.1], np.float32)
])
np.testing.assert_array_equal(
sut._trainer.train_minibatch.call_args[0][0][sut._output_variables],
[
# 10 (reward) + 0.9 (gamma) x 0.4 (max q_target)
np.array([10.36, 0.1], np.float32),
# 11 (reward) + 0.9 (gamma) x 0.4 (max q_target)
np.array([0.2, 11.36], np.float32),
np.array([10.36, 0.1], np.float32)
])
np.testing.assert_almost_equal(
sut._trainer.train_minibatch.call_args[0][0][sut._weight_variables],
[
[0.16666667],
[0.66666667],
[0.16666667]
])
self.assertAlmostEqual(
sut._replay_memory.update_priority.call_args[0][0][3],
105.2676) # (10.16 + 0.1)^2
self.assertAlmostEqual(
sut._replay_memory.update_priority.call_args[0][0][4],
129.0496,
places=6) # (11.26 + 0.1) ^ 2
def test_array(self):
s = spaces.Box(0, 1, (2, 2))
sut = BoxSpaceDiscretizer(s, np.array([[2, 2], [2, 2]]))
self.assertEqual(sut.discretize([[0, 0], [0, 0]]), 0)
self.assertEqual(sut.discretize([[0.05, 0], [0, 0]]), 0)
self.assertEqual(sut.discretize([[0.95, 0], [0, 0]]), 8)
self.assertEqual(sut.discretize([[0, 0.05], [0, 0]]), 0)
self.assertEqual(sut.discretize([[0, 0.95], [0, 0]]), 4)
self.assertEqual(sut.discretize([[0, 0], [0.05, 0]]), 0)
self.assertEqual(sut.discretize([[0, 0], [0.95, 0]]), 2)
self.assertEqual(sut.discretize([[0, 0], [0, 0.05]]), 0)
self.assertEqual(sut.discretize([[0, 0], [0, 0.95]]), 1)
self.assertEqual(sut.discretize([[0.1, 0.6], [0.5, 0.2]]), 6)
self.assertEqual(sut.discretize([[1, 1], [1, 1]]), 15)
def test_process_accumulated_trajectory_keep_last(self):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = ActorCritic('', observation_space, action_space)
# Set up.
self._setup_trajectory(sut)
# Call test method.
sut._process_accumulated_trajectory(True)
# Verify results.
self.assertEqual(len(sut._trajectory_rewards), 0)
self.assertEqual(len(sut._trajectory_actions), 0)
self.assertEqual(sut._trajectory_states, [np.array([0.3], np.float32)])
def test_init_from_existing_model(self, mock_parameters):
action_space = spaces.Discrete(3)
observation_space = spaces.Box(np.array([-1.2, -0.07]),
np.array([0.6, 0.07]))
mock_parameters.return_value.policy_representation = 'nn'
mock_parameters.return_value.policy_network_hidden_layers = '[2]'
mock_parameters.return_value.initial_policy_network = \
'tests/data/initial_policy_network.dnn'
mock_parameters.return_value.preprocessing = ''
sut = ActorCritic('', observation_space, action_space)
self.assertEqual(sut._num_actions, 3)
self.assertIsNone(sut._num_states)
self.assertEqual(sut._shape_of_inputs, (2, ))
self.assertFalse(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
# Incompatible network structure.
mock_parameters.return_value.policy_network_hidden_layers = '[]'
self.assertRaises(Exception, ActorCritic, '', observation_space,
action_space)
# Incompatible action space.
mock_parameters.return_value.policy_network_hidden_layers = '[2]'
action_space = spaces.Discrete(2)
self.assertRaises(ValueError, ActorCritic, '', observation_space,
action_space)
# Incompatible observation space.
action_space = spaces.Discrete(3)
observation_space = spaces.Box(np.array([-1.2, -0.07, -1.0]),
np.array([0.6, 0.07, 1.0]))
self.assertRaises(ValueError, ActorCritic, '', observation_space,
action_space)
def test_rollout_with_update(self, mock_parameters):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.update_frequency = 2
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = ActorCritic('', observation_space, action_space)
sut._update_networks = MagicMock()
sut._choose_action = Mock(
side_effect=[(0, ''), (1, ''), (1, ''), (0, ''), (1, ''), (0, '')])
sut.start(np.array([0.1], np.float32))
sut.step(0.1, np.array([0.2], np.float32))
self.assertEqual(sut._trajectory_rewards, [0.1])
self.assertEqual(sut._trajectory_actions, [0, 1])
self.assertEqual(sut._trajectory_states, [0.1, 0.2])
self.assertEqual(sut._update_networks.call_count, 0)
sut.step(0.2, np.array([0.3], np.float32))
self.assertEqual(sut._trajectory_rewards, [])
self.assertEqual(sut._trajectory_actions, [1])
self.assertEqual(sut._trajectory_states, [0.3])
self.assertEqual(sut._update_networks.call_count, 1)
sut.step(0.3, np.array([0.4], np.float32))
self.assertEqual(sut._trajectory_rewards, [0.3])
self.assertEqual(sut._trajectory_actions, [1, 0])
self.assertEqual(sut._trajectory_states, [0.3, 0.4])
self.assertEqual(sut._update_networks.call_count, 1)
sut.start(np.array([0.5], np.float32))
self.assertEqual(sut._trajectory_rewards, [])
self.assertEqual(sut._trajectory_actions, [1])
self.assertEqual(sut._trajectory_states, [0.5])
self.assertEqual(sut._update_networks.call_count, 1)
sut.step(0.4, np.array([0.6], np.float32))
self.assertEqual(sut._trajectory_rewards, [])
self.assertEqual(sut._trajectory_actions, [0])
self.assertEqual(sut._trajectory_states, [0.6])
self.assertEqual(sut._update_networks.call_count, 2)
sut.end(0.5, np.array([0.7], np.float32))
self.assertEqual(sut._trajectory_rewards, [0.5])
self.assertEqual(sut._trajectory_actions, [0])
self.assertEqual(sut._trajectory_states, [0.6])
self.assertEqual(sut._update_networks.call_count, 2)
def test_init_unsupported_model(self, mock_parameters):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
self._setup_parameters(mock_parameters.return_value)
# Verify sut can be constructed.
sut = ActorCritic('', observation_space, action_space)
mock_parameters.return_value.policy_representation = 'undefined'
self.assertRaises(ValueError, ActorCritic, '', observation_space,
action_space)
mock_parameters.return_value.policy_representation = 'nn'
mock_parameters.return_value.value_function_representation = 'undefined'
self.assertRaises(ValueError, ActorCritic, '', observation_space,
action_space)
def test_replay_start_size(self, mock_parameters):
self._setup_parameters(mock_parameters.return_value)
# Set exploration rate to 0
mock_parameters.return_value.initial_epsilon = 0
mock_parameters.return_value.epsilon_decay_step_count = 100
mock_parameters.return_value.epsilon_minimum = 0
mock_parameters.return_value.replay_start_size = 3
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
sut._trainer = MagicMock()
sut._replay_memory = MagicMock()
_, debug = sut.start(np.array([0.1], np.float32))
self.assertEqual(sut.step_count, 0)
self.assertEqual(sut._trainer.train_minibatch.call_count, 0)
self.assertEqual(debug['action_behavior'], 'RANDOM')
_, debug = sut.step(0.1, np.array([0.2], np.float32))
self.assertEqual(sut.step_count, 1)
self.assertEqual(sut._trainer.train_minibatch.call_count, 0)
self.assertEqual(debug['action_behavior'], 'RANDOM')
sut.end(0.2, np.array([0.3], np.float32))
self.assertEqual(sut.step_count, 2)
self.assertEqual(sut._trainer.train_minibatch.call_count, 0)
_, debug = sut.start(np.array([0.4], np.float32))
self.assertEqual(sut.step_count, 2)
self.assertEqual(sut._trainer.train_minibatch.call_count, 0)
self.assertEqual(debug['action_behavior'], 'RANDOM')
a, debug = sut.step(0.3, np.array([0.5], np.float32))
self.assertEqual(sut.step_count, 3)
self.assertEqual(sut._trainer.train_minibatch.call_count, 0)
self.assertEqual(debug['action_behavior'], 'GREEDY')
a, debug = sut.start(np.array([0.6], np.float32))
self.assertEqual(sut.step_count, 3)
self.assertEqual(sut._trainer.train_minibatch.call_count, 0)
self.assertEqual(debug['action_behavior'], 'GREEDY')
a, debug = sut.step(0.4, np.array([0.7], np.float32))
self.assertEqual(sut.step_count, 4)
self.assertEqual(sut._trainer.train_minibatch.call_count, 1)
self.assertEqual(debug['action_behavior'], 'GREEDY')
def test_init_customized_q(self, mock_parameters, mock_model):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.q_representation = \
'cntk.contrib.deeprl.agent.shared.customized_models.conv_dqn'
mock_model.return_value = self._setup_test_model()
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
self.assertEqual(sut._num_actions, 2)
self.assertIsNone(sut._num_states)
self.assertEqual(sut._shape_of_inputs, (1,))
self.assertFalse(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
mock_model.assert_called_with((1,), 2, None)
def test_init_dueling_dqn(self, mock_parameters, mock_model):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.q_representation = 'dueling-dqn'
mock_parameters.return_value.hidden_layers = '[2, [2], [2]]'
mock_model.return_value = self._setup_test_model()
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
self.assertEqual(sut._num_actions, 2)
self.assertIsNone(sut._num_states)
self.assertEqual(sut._shape_of_inputs, (1,))
self.assertFalse(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
mock_model.assert_called_with((1,), 2, '[2, [2], [2]]', None)
def test_init_dqn(self, mock_parameters, mock_model, mock_replay_memory):
self._setup_parameters(mock_parameters.return_value)
mock_model.return_value = self._setup_test_model()
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = QLearning('', observation_space, action_space)
self.assertEqual(sut._num_actions, 2)
self.assertIsNone(sut._num_states)
self.assertEqual(sut._shape_of_inputs, (1, ))
self.assertFalse(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
self.assertFalse(hasattr(sut, 'weight_variables'))
self.assertIsNotNone(sut._trainer)
mock_model.assert_called_with((1, ), 2, '[2]', None)
mock_replay_memory.assert_called_with(100, False)
def test_populate_replay_memory(self, mock_parameters):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.preprocessing = \
'cntk.contrib.deeprl.agent.shared.preprocessing.SlidingWindow'
mock_parameters.return_value.preprocessing_args = '(2, )'
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = QLearning('', observation_space, action_space)
sut._compute_priority = Mock(side_effect=[1, 2, 3])
sut._choose_action = Mock(
side_effect=[(0, ''), (0, ''), (1, ''), (1, '')])
sut._replay_memory = MagicMock()
sut._update_q_periodically = MagicMock()
sut.start(np.array([0.1], np.float32))
sut.step(0.1, np.array([0.2], np.float32))
sut.step(0.2, np.array([0.3], np.float32))
sut.end(0.3, np.array([0.4], np.float32))
self.assertEqual(sut._replay_memory.store.call_count, 3)
call_args = sut._replay_memory.store.call_args_list[0]
np.testing.assert_array_equal(
call_args[0][0],
np.array([[0], [0.1]], np.float32))
self.assertEqual(call_args[0][1], 0)
self.assertEqual(call_args[0][2], 0.1)
np.testing.assert_array_equal(
call_args[0][3],
np.array([[0.1], [0.2]], np.float32))
self.assertEqual(call_args[0][4], 1)
call_args = sut._replay_memory.store.call_args_list[2]
np.testing.assert_array_equal(
call_args[0][0],
np.array([[0.2], [0.3]], np.float32))
self.assertEqual(call_args[0][1], 1)
self.assertEqual(call_args[0][2], 0.3)
self.assertIsNone(call_args[0][3])
self.assertEqual(call_args[0][4], 3)
def test_rollout(self):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1, ))
sut = ActorCritic('', observation_space, action_space)
sut._choose_action = Mock(side_effect=[(0, ''), (1, ''), (1, '')])
sut.start(np.array([0.1], np.float32))
sut.step(0.1, np.array([0.2], np.float32))
sut.step(0.2, np.array([0.3], np.float32))
self.assertEqual(sut._trajectory_rewards, [0.1, 0.2])
self.assertEqual(sut._trajectory_actions, [0, 1, 1])
self.assertEqual(sut._trajectory_states, [0.1, 0.2, 0.3])
sut.end(0.3, np.array([0.4], np.float32))
self.assertEqual(sut._trajectory_rewards, [0.1, 0.2, 0.3])
self.assertEqual(sut._trajectory_actions, [0, 1, 1])
self.assertEqual(sut._trajectory_states, [0.1, 0.2, 0.3])
def test_init_customized_model(self, mock_parameters, mock_model):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.policy_representation = \
'cntk.contrib.deeprl.agent.shared.customized_models.conv_dqn'
mock_parameters.return_value.value_function_representation = \
'cntk.contrib.deeprl.agent.shared.customized_models.conv_dqn'
mock_model.side_effect = self._setup_test_model
sut = ActorCritic('', observation_space, action_space)
self.assertEqual(mock_model.call_count, 2)
mock_model.assert_has_calls(
[
mock.call((1,), 2, cross_entropy_with_softmax,
use_placeholder_for_input=True),
mock.call((1,), 1, use_placeholder_for_input=True)
],
any_order=True)
def test_list(self):
s = spaces.Box(0, 1, (2, ))
sut = BoxSpaceDiscretizer(s, np.array([10, 2]))
self.assertEqual(sut.discretize([0, 0]), 0)
self.assertEqual(sut.discretize([0.05, 0]), 0)
self.assertEqual(sut.discretize([0.95, 0]), 18)
self.assertEqual(sut.discretize([0, 0.05]), 0)
self.assertEqual(sut.discretize([0, 0.95]), 1)
self.assertEqual(sut.discretize([0.1, 0.2]), 2)
self.assertEqual(sut.discretize([1, 1]), 19)
sut = BoxSpaceDiscretizer(s, np.array([10, 1]))
self.assertEqual(sut.discretize([0, 0]), 0)
self.assertEqual(sut.discretize([0.05, 0]), 0)
self.assertEqual(sut.discretize([0.95, 0]), 9)
self.assertEqual(sut.discretize([0, 0.05]), 0)
self.assertEqual(sut.discretize([0, 0.95]), 0)
self.assertEqual(sut.discretize([0.1, 0.2]), 1)
self.assertEqual(sut.discretize([1, 1]), 9)
def test_init_preprocess(self, mock_parameters, mock_model):
self._setup_parameters(mock_parameters.return_value)
mock_parameters.return_value.preprocessing = \
'cntk.contrib.deeprl.agent.shared.preprocessing.SlidingWindow'
mock_parameters.return_value.preprocessing_args = '(2, )'
mock_model.side_effect = self._setup_test_model
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = ActorCritic('', observation_space, action_space)
self.assertIsNotNone(sut._preprocessor)
self.assertEqual(sut._preprocessor.output_shape(), (2, 1))
self.assertEqual(mock_model.call_count, 2)
mock_model.assert_has_calls(
[
mock.call((2, 1), 2, '[2]', cross_entropy_with_softmax,
use_placeholder_for_input=True),
mock.call((2, 1), 1, '[2]', use_placeholder_for_input=True)
],
any_order=True)
def test_init(self, mock_model):
mock_model.side_effect = self._setup_test_model
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = ActorCritic('', observation_space, action_space)
self.assertEqual(sut._num_actions, 2)
self.assertIsNone(sut._num_states)
self.assertEqual(sut._shape_of_inputs, (1,))
self.assertFalse(sut._discrete_observation_space)
self.assertIsNone(sut._space_discretizer)
self.assertIsNone(sut._preprocessor)
self.assertEqual(mock_model.call_count, 2)
mock_model.assert_has_calls(
[
mock.call((1,), 2, '[10]', cross_entropy_with_softmax,
use_placeholder_for_input=True),
mock.call((1,), 1, '[10]', use_placeholder_for_input=True)
],
any_order=True)
def test_process_accumulated_trajectory(self):
action_space = spaces.Discrete(2)
observation_space = spaces.Box(0, 1, (1,))
sut = ActorCritic('', observation_space, action_space)
# Set up.
self._setup_trajectory(sut)
# Call test method.
sut._process_accumulated_trajectory(False)
# Verify results.
self.assertEqual(len(sut._trajectory_rewards), 0)
self.assertEqual(len(sut._trajectory_actions), 0)
self.assertEqual(len(sut._trajectory_states), 0)
np.testing.assert_array_equal(
sut._input_buffer,
[np.array([0.1], np.float32), np.array([0.2], np.float32)])
# For unknown reason, got [2.9974999999999996] instead of [2.9975] for
# the following testcase, therefore use assert_array_almost_equal.
np.testing.assert_array_almost_equal(
sut._value_network_output_buffer,
[
[2.9975], # 3.05 * 0.95 + 0.1
[3.05] # 3 (initial_r) * 0.95 + 0.2
])
np.testing.assert_array_equal(
sut._policy_network_output_buffer,
[
np.array([1, 0], np.float32),
np.array([0, 1], np.float32)
]
)
np.testing.assert_array_almost_equal(
sut._policy_network_weight_buffer,
[
[0.9975], # 2.9975 - 2
[2.05] # 3.05 - 1
])
def test_init_unsupported_action_space(self):
action_space = spaces.Box(0, 1, (1, ))
observation_space = spaces.Discrete(3)
self.assertRaises(ValueError, FakeAgentBaseClass, observation_space,
action_space)
