def test_comparison():
continuous_spaces = [make_box([-1, 2.3], [45, 4.3]), make_box([-10], [45])]
space = SimpleHybrid(continuous_spaces)
continuous_spaces2 = [
make_box([-1, 54], [45, 4.3]),
make_box([-10], [445])
]
space2 = SimpleHybrid(continuous_spaces2)
assert space == space
assert space != "hi"
assert space != space2
assert issubclass(SimpleHybrid, HybridBase)
def test_invalid_arguments():
with pytest.raises(AssertionError):
SimpleHybrid("string")
with pytest.raises(AssertionError):
SimpleHybrid(1.343)
with pytest.raises(ValueError):
SimpleHybrid([])
with pytest.raises(AssertionError):
SimpleHybrid([1, 2])
with pytest.raises(AssertionError):
continuous_spaces = [
make_box([-1, 2.3], [45, 4.3]),
make_box([-10], [45]),
make_box([50, 34, 0], [100, 120, 2]),
]
SimpleHybrid([Discrete(10)] + continuous_spaces)
def test_repr_does_not_throw_error():
from hmlf.spaces import Box
continuous_spaces = [make_box([-1, 2.3], [45, 4.3]), make_box([-10], [45])]
space = SimpleHybrid(continuous_spaces)
represensation_string = repr(space)
represensation_string = represensation_string.replace(", float32", "")
eval(represensation_string)
assert "SimpleHybrid" in represensation_string
def test_dimensions():
space = SimpleHybrid(
[make_box(shape=(1, )),
make_box(shape=(3, )),
make_box(shape=(2, ))])
assert space.n_discrete_options == 3
assert space.get_n_discrete_spaces() == 1
assert space.get_n_discrete_options() == 3
assert isinstance(space._get_continuous_spaces(), list)
assert len(space._get_continuous_spaces()) == 3
assert space.get_n_continuous_spaces() == 3
assert space.get_n_continuous_options() == (1 + 3 + 2)
assert space._get_dimensions_of_continuous_spaces() == [1, 3, 2]
assert np.array_equal(space.split_indices, [1, 4])
assert space.get_dimension() == (1 + 1 + 3 + 2)
def test_build_action():
continuous_spaces = [make_box([-1, 2.3], [45, 4.3]), make_box([-10], [45])]
space = SimpleHybrid(continuous_spaces)
discrete = np.array([2, 0, 1])
parameters = np.array([
[0, 3, 2],
[-10, 3, 63.1],
[0, 3, 50],
])
action = space.build_action(discrete, parameters)
assert isinstance(action, list)
assert len(action) == 3
assert len(action[0]) == 3
assert isinstance(action[0], tuple)
assert action[0][0] == 2
assert action[1][0] == 0
assert action[2][0] == 1
assert np.allclose(action[0][1], [0.0, 3.0])
assert np.allclose(action[0][2], [2])
assert np.allclose(action[1][1], [-1, 3])
assert np.allclose(action[1][2], [45])
def test_low_high_concatination():
continuous_spaces = [
make_box([-1, 2.3], [45, 4.3]),
make_box([-10], [45]),
make_box([50, 34, 0], [100, 120, 2]),
]
space = SimpleHybrid(continuous_spaces)
print(space.continuous_low)
print(space.continuous_high)
assert np.allclose(space.continuous_low,
[-1.0, 2.3, -10.0, 50.0, 34.0, 0.0])
assert np.allclose(space.continuous_high,
[45.0, 4.3, 45.0, 100.0, 120.0, 2.0])
def __init__(self):
self.max_move = np.float32(0.2)
self.max_jump = 0.05
self.n_obstacles = 3
self.goal_position = 0.95
self.obstacle_thickness = 0.01
self.jump_threshold = 0.1
self.max_timesteps = 30
self.goal_threshold = 0.05
self.action_space = SimpleHybrid([
Box(-self.max_move, self.max_move, (1, )),
Box(np.float32(0), np.float32(1), (1, ))
])
self.observation_space = Box(low=np.zeros(7, dtype=np.float32),
high=np.ones(7, dtype=np.float32))
self.position = 0
self.time = 0
self.obstacle_position = np.zeros(self.n_obstacles)
self.obstacle_target_height = np.zeros(self.n_obstacles)
def test_init_simple_hybrid(spaces):
hybrid_space = SimpleHybrid(spaces)
assert hybrid_space.get_n_discrete_options() == len(spaces)
sample = hybrid_space.sample()
# check datatypes
assert type(sample) == tuple
assert type(sample[0]) == int
for s in sample[1:]:
assert type(s) == np.ndarray
# check that sample and fomat_actions work consistently and correctly
sample_list = [hybrid_space.sample(), hybrid_space.sample()]
sample_array = np.vstack([np.hstack(sample_list[0]), np.hstack(sample_list[1])])
new_sample_list = hybrid_space.format_action(sample_array)
for sample, new_sample in zip(sample_list, new_sample_list):
for action, new_action in zip(sample, new_sample):
assert np.sum(np.abs(action - new_action)) < 1e-6
def __init__(
self,
observation_space: Space,
action_space: SimpleHybrid,
lr_schedule_q: Schedule,
lr_schedule_parameter: Schedule,
net_arch_q: Optional[List[int]] = None,
net_arch_parameter: Optional[List[int]] = None,
activation_fn: Type[nn.Module] = nn.ReLU,
features_extractor_class: Type[
BaseFeaturesExtractor] = FlattenExtractor,
features_extractor_kwargs: Optional[Dict[str, Any]] = None,
normalize_images: bool = True,
optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
optimizer_kwargs: Optional[Dict[str, Any]] = None,
):
super(MPDQNPolicy, self).__init__(
observation_space, # :Space,
action_space, # :SimpleHybrid,
lr_schedule_q, # :Schedule,
lr_schedule_parameter, # :Schedule,
net_arch_q, # :Optional[List[int]] = None,
net_arch_parameter, # :Optional[List[int]] = None,
activation_fn, # :Type[nn.Module] = nn.ReLU,
features_extractor_class, # :Type[BaseFeaturesExtractor] = FlattenExtractor,
features_extractor_kwargs, # :Optional[Dict[str, Any]] = None,
normalize_images, # :bool = True,
optimizer_class, # :Type[th.optim.Optimizer] = th.optim.Adam,
optimizer_kwargs, # :Optional[Dict[str, Any]] = None,
)
# For formatting inside forward Q
self.discrete_action_size = self.action_space_q.n
self.state_size = observation_space.shape[0]
self.offsets = np.cumsum(
action_space._get_dimensions_of_continuous_spaces())
self.offsets = np.insert(self.offsets, 0, 0)
def test_input_tuple(spaces):
space_list = SimpleHybrid(spaces)
space_tuple = SimpleHybrid(tuple(spaces))
assert space_list == space_tuple
def _build_action_space(self) -> SimpleHybrid:
parameter_spaces = self._build_parameter_spaces()
spaces = parameter_spaces
return SimpleHybrid(spaces)