args = get_argparser().parse_args()
# Select the data set
data_set_name = args.mode or "skyline"
# Experiment
ex_dir = setup_experiment(TSPred.name, ADNPolicy.name)
# Set seed if desired
pyrado.set_seed(args.seed, verbose=True)
# Load the data
if data_set_name == "skyline":
dt = 0.01
_, vals = skyline(dt=dt,
t_end=20.0,
t_intvl_space=BoxSpace(0.5, 3, shape=(1, )),
val_space=BoxSpace(-2.0, 3.0, shape=(1, )))
data = to.from_numpy(vals).to(dtype=to.get_default_dtype()).view(-1, 1)
else:
data = pd.read_csv(
osp.join(pyrado.PERMA_DIR, "misc", f"{data_set_name}.csv"))
if data_set_name == "daily_min_temperatures":
data = to.tensor(data["Temp"].values,
dtype=to.get_default_dtype()).view(-1, 1)
dt = 1.0
elif data_set_name == "monthly_sunspots":
data = to.tensor(data["Sunspots"].values,
dtype=to.get_default_dtype()).view(-1, 1)
dt = 1.0
elif "oscillation" in data_set_name:
data = to.tensor(data["Positions"].values,
def test_skyline(dt: Union[int, float], t_end: Union[int, float],
t_intvl_space: BoxSpace, val_space: BoxSpace):
# Create the skyline function
t, vals = skyline(dt, t_end, t_intvl_space, val_space)
assert isinstance(t, np.ndarray) and isinstance(vals, np.ndarray)
assert len(t) == len(vals)
def default_randomizer():
return DomainRandomizer(
NormalDomainParam(name="mass", mean=1.2, std=0.1, clip_lo=10, clip_up=100),
UniformDomainParam(name="special", mean=0, halfspan=42, clip_lo=-7.4, roundint=True),
NormalDomainParam(name="length", mean=4, std=0.6, clip_up=50.1),
UniformDomainParam(name="time_delay", mean=13, halfspan=6, clip_up=17, roundint=True),
MultivariateNormalDomainParam(name="multidim", mean=10 * to.ones((2,)), cov=2 * to.eye(2), clip_up=11),
)
# --------------
# Other Fixtures
# --------------
@pytest.fixture(
scope="function",
params=[
(skyline(0.02, 20.0, BoxSpace(0.5, 3, shape=(1,)), BoxSpace(-2.0, 3.0, shape=(1,)))[1], 0.7, 50, False, True),
],
ids=["skyline_0.8_50_notstdized_scaled"],
)
def dataset_ts(request):
return TimeSeriesDataSet(
data=to.from_numpy(request.param[0]).to(dtype=to.get_default_dtype()),
ratio_train=request.param[1],
window_size=request.param[2],
standardize_data=request.param[3],
scale_min_max_data=request.param[4],
)
from pyrado.policies.feed_forward.linear import LinearPolicy
from pyrado.policies.features import *
from pyrado.policies.recurrent.two_headed_rnn import TwoHeadedRNNPolicyBase
from pyrado.sampling.rollout import rollout
from pyrado.sampling.step_sequence import StepSequence
from pyrado.utils.data_sets import TimeSeriesDataSet
from pyrado.utils.data_types import RenderMode
from pyrado.utils.functions import skyline
from pyrado.utils.nn_layers import IndiNonlinLayer
from tests.conftest import m_needs_bullet, m_needs_mujoco, m_needs_rcs, m_needs_libtorch
from tests.environment_wrappers.mock_env import MockEnv
@pytest.fixture(scope='function',
params=[
(skyline(0.02, 20., BoxSpace(0.5, 3, shape=(1,)), BoxSpace(-2., 3., shape=(1,)))[1],
0.7, 50, False, True),
],
ids=['skyline_0.8_50_notstdized_scaled'])
def tsdataset(request):
return TimeSeriesDataSet(
data=to.from_numpy(request.param[0]),
ratio_train=request.param[1],
window_size=request.param[2],
standardize_data=request.param[3],
scale_min_max_data=request.param[4]
)
@pytest.mark.features
@pytest.mark.parametrize(
eq_constraint=
f"{QQubeSwingUpReal.name}, {QCartPoleSwingUpReal.name} or {WAMReal.name}"
)
print_cbt(f"Set the {env_real.name} environment.", "g")
# Create the policy
if args.mode.lower() == "chirp":
def fcn_of_time(t: float):
act = max_amp * chirp(t, f0=3, f1=0, t1=t_end, method="linear")
return act.repeat(env_real.act_space.flat_dim)
elif args.mode.lower() == "skyline":
t_intvl_space = BoxSpace(dt, 100 * dt, shape=(1, ))
val_space = BoxSpace(-abs(max_amp), abs(max_amp), shape=(1, ))
act_predef = skyline(dt, t_end, t_intvl_space, val_space)[1]
def fcn_of_time(t: float):
act = act_predef[int(t / dt)]
return act.repeat(env_real.act_space.flat_dim)
elif args.mode.lower() == "sin":
def fcn_of_time(t: float):
act = max_amp * np.sin(2 * np.pi * t * 2.0) # 2 Hz
return act.repeat(env_real.act_space.flat_dim)
elif args.mode.lower() == "wam_sin":
fcn_of_time = wam_jsp_7dof_sin
else:
