def algo_tester(algo,
observation_shape,
imitator=False,
action_size=2,
state_value=False):
# dummy impl object
impl = DummyImpl(observation_shape, action_size)
base_tester(algo, impl, observation_shape, action_size)
algo._impl = impl
# check save policy
impl.save_policy = Mock()
algo.save_policy("policy.pt", False)
impl.save_policy.assert_called_with("policy.pt", False)
# check predict
x = np.random.random((2, 3)).tolist()
ref_y = np.random.random((2, action_size)).tolist()
impl.predict_best_action = Mock(return_value=ref_y)
y = algo.predict(x)
assert y == ref_y
impl.predict_best_action.assert_called_with(x)
# check predict_value
if not imitator and not state_value:
action = np.random.random((2, action_size)).tolist()
ref_value = np.random.random((2, 3)).tolist()
impl.predict_value = Mock(return_value=ref_value)
value = algo.predict_value(x, action)
assert value == ref_value
impl.predict_value.assert_called_with(x, action, False)
# check sample_action
impl.sample_action = Mock(return_value=ref_y)
try:
y = algo.sample_action(x)
assert y == ref_y
impl.sample_action.assert_called_with(x)
except NotImplementedError:
pass
algo._impl = None
def ope_tester(ope, observation_shape, action_size=2):
# dummy impl object
impl = DummyImpl(observation_shape, action_size)
base_tester(ope, impl, observation_shape, action_size)
ope._algo.impl = impl
ope.impl = impl
# check save policy
impl.save_policy = Mock()
ope.save_policy("policy.pt", False)
impl.save_policy.assert_called_with("policy.pt", False)
# check predict
x = np.random.random((2, 3)).tolist()
ref_y = np.random.random((2, action_size)).tolist()
impl.predict_best_action = Mock(return_value=ref_y)
y = ope.predict(x)
assert y == ref_y
impl.predict_best_action.assert_called_with(x)
# check predict_value
action = np.random.random((2, action_size)).tolist()
ref_value = np.random.random((2, 3)).tolist()
impl.predict_value = Mock(return_value=ref_value)
value = ope.predict_value(x, action)
assert value == ref_value
impl.predict_value.assert_called_with(x, action, False)
# check sample_action
impl.sample_action = Mock(return_value=ref_y)
try:
y = ope.sample_action(x)
assert y == ref_y
impl.sample_action.assert_called_with(x)
except NotImplementedError:
pass
ope.impl = None
ope._algo.impl = None
def dynamics_tester(dynamics, observation_shape, action_size=2):
# dummy impl object
impl = DummyImpl(observation_shape, action_size)
base_tester(dynamics, impl, observation_shape, action_size)
dynamics._impl = impl
# check predict
x = np.random.random((2, 3)).tolist()
action = np.random.random((2, 3)).tolist()
ref_y = np.random.random((2, 3)).tolist()
ref_reward = np.random.random((2, 1)).tolist()
ref_variance = np.random.random((2, 1)).tolist()
impl.predict = Mock(return_value=(ref_y, ref_reward, ref_variance))
y, reward = dynamics.predict(x, action)
assert y == ref_y
assert reward == ref_reward
impl.predict.assert_called_with(x, action)
# check with_variance
y, reward, variance = dynamics.predict(x, action, with_variance=True)
assert variance == ref_variance
def dynamics_tester(
dynamics, observation_shape, action_size=2, discrete_action=False
):
# dummy impl object
impl = DummyImpl(observation_shape, action_size)
base_tester(dynamics, impl, observation_shape, action_size)
dynamics.create_impl(observation_shape, action_size)
# check predict
x = np.random.random((2, *observation_shape))
if discrete_action:
action = np.random.randint(action_size, size=2)
else:
action = np.random.random((2, action_size))
y, reward = dynamics.predict(x, action)
assert y.shape == (2, *observation_shape)
assert reward.shape == (2, 1)
# check with_variance
y, reward, variance = dynamics.predict(x, action, with_variance=True)
assert variance.shape == (2, 1)
def algo_tester(
algo,
observation_shape,
imitator=False,
action_size=2,
state_value=False,
test_policy_copy=False,
test_q_function_copy=False,
):
# dummy impl object
impl = DummyImpl(observation_shape, action_size)
base_tester(algo, impl, observation_shape, action_size)
algo._impl = impl
# check save policy
impl.save_policy = Mock()
algo.save_policy("policy.pt", False)
impl.save_policy.assert_called_with("policy.pt", False)
# check predict
x = np.random.random((2, 3)).tolist()
ref_y = np.random.random((2, action_size)).tolist()
impl.predict_best_action = Mock(return_value=ref_y)
y = algo.predict(x)
assert y == ref_y
impl.predict_best_action.assert_called_with(x)
# check predict_value
if not imitator and not state_value:
action = np.random.random((2, action_size)).tolist()
ref_value = np.random.random((2, 3)).tolist()
impl.predict_value = Mock(return_value=ref_value)
value = algo.predict_value(x, action)
assert value == ref_value
impl.predict_value.assert_called_with(x, action, False)
# check sample_action
impl.sample_action = Mock(return_value=ref_y)
try:
y = algo.sample_action(x)
assert y == ref_y
impl.sample_action.assert_called_with(x)
except NotImplementedError:
pass
algo.create_impl(observation_shape, action_size)
if test_policy_copy:
algo2 = algo.__class__(**algo.get_params())
algo2.create_impl(observation_shape, action_size)
algo.copy_policy_from(algo2)
observations = np.random.random((100, *observation_shape))
action1 = algo.predict(observations)
action2 = algo.predict(observations)
assert np.all(action1 == action2)
if test_q_function_copy:
algo2 = algo.__class__(**algo.get_params())
algo2.create_impl(observation_shape, action_size)
algo.copy_q_function_from(algo2)
observations = np.random.random((100, *observation_shape))
if algo.get_action_type() == ActionSpace.CONTINUOUS:
actions = np.random.random((100, action_size))
else:
actions = np.random.randint(action_size, size=100)
value1 = algo.predict_value(observations, actions)
value2 = algo2.predict_value(observations, actions)
assert np.all(value1 == value2)
algo._impl = None
