def test_adaptive_parzen_normal_estimator():
"""Test adaptive parzen estimator"""
low = -1
high = 5
obs_mus = [1.2]
mus, sigmas, weights = adaptive_parzen_estimator(
obs_mus, low, high, prior_weight=1.0, equal_weight=False, flat_num=25
)
assert list(mus) == [1.2, 2]
assert list(sigmas) == [3, 6]
assert list(weights) == [1.0 / 2, 1.0 / 2]
obs_mus = [3.4]
mus, sigmas, weights = adaptive_parzen_estimator(
obs_mus, low, high, prior_weight=0.5, equal_weight=False, flat_num=25
)
assert list(mus) == [2, 3.4]
assert list(sigmas) == [6, 3]
assert list(weights) == [0.5 / 1.5, 1.0 / 1.5]
obs_mus = numpy.linspace(-1, 5, num=30, endpoint=False)
mus, sigmas, weights = adaptive_parzen_estimator(
obs_mus, low, high, prior_weight=1.0, equal_weight=False, flat_num=25
)
ramp = numpy.linspace(1.0 / 30, 1.0, num=30 - 25)
full = numpy.ones(25 + 1)
all_weights = numpy.concatenate([ramp, full])
assert len(mus) == len(sigmas) == len(weights) == 30 + 1
assert numpy.all(weights[: 30 - 25] == ramp / all_weights.sum())
assert numpy.all(weights[30 - 25 :] == 1 / all_weights.sum())
assert numpy.all(sigmas == 6 / 10)
def test_adaptive_parzen_normal_estimator_weight():
"""Test the weight for the normal components"""
obs_mus = numpy.linspace(-1, 5, num=30, endpoint=False)
low = -1
high = 5
# equal weight
mus, sigmas, weights = adaptive_parzen_estimator(obs_mus,
low,
high,
prior_weight=1.0,
equal_weight=True,
flat_num=25)
assert numpy.all(weights == 1 / 31)
assert numpy.all(sigmas == 6 / 10)
# prior weight
mus, sigmas, weights = adaptive_parzen_estimator(obs_mus,
low,
high,
prior_weight=0.5,
equal_weight=False,
flat_num=25)
ramp = numpy.linspace(1.0 / 30, 1.0, num=30 - 25)
full = numpy.ones(25 + 1)
all_weights = (numpy.concatenate([ramp, full]))
prior_pos = numpy.searchsorted(mus, 2)
all_weights[prior_pos] = 0.5
assert numpy.all(
weights[:30 - 25] == (numpy.linspace(1.0 / 30, 1.0, num=30 - 25) /
all_weights.sum()))
assert numpy.all(weights[33 - 25:prior_pos] == 1 / all_weights.sum())
assert weights[prior_pos] == 0.5 / all_weights.sum()
assert numpy.all(weights[prior_pos + 1:] == 1 / all_weights.sum())
assert numpy.all(sigmas == 6 / 10)
# full weights number
mus, sigmas, weights = adaptive_parzen_estimator(obs_mus,
low,
high,
prior_weight=1.0,
equal_weight=False,
flat_num=15)
ramp = numpy.linspace(1.0 / 30, 1.0, num=30 - 15)
full = numpy.ones(15 + 1)
all_weights = (numpy.concatenate([ramp, full]))
prior_pos = numpy.searchsorted(mus, 2)
all_weights[prior_pos] = 1.0
assert numpy.all(
weights[:30 - 15] == (numpy.linspace(1.0 / 30, 1.0, num=30 - 15) /
all_weights.sum()))
assert numpy.all(weights[30 - 15:] == 1 / all_weights.sum())
assert numpy.all(sigmas == 6 / 10)
def test_adaptive_parzen_normal_estimator_sigma_clip():
"""Test that the magic clip of sigmas for parzen estimator"""
low = -1
high = 5
obs_mus = numpy.linspace(-1, 5, num=8, endpoint=False)
mus, sigmas, weights = adaptive_parzen_estimator(obs_mus,
low,
high,
prior_weight=1.0,
equal_weight=False,
flat_num=25)
assert len(mus) == len(sigmas) == len(weights) == 8 + 1
assert numpy.all(weights == 1 / 9)
assert numpy.all(sigmas == 6 / 8)
obs_mus = numpy.random.uniform(-1, 5, 30)
mus, sigmas, weights = adaptive_parzen_estimator(obs_mus,
low,
high,
prior_weight=1.0,
equal_weight=False,
flat_num=25)
assert len(mus) == len(sigmas) == len(weights) == 30 + 1
assert numpy.all(weights[-25:] == weights[-1])
assert numpy.all(sigmas <= 6) and numpy.all(sigmas >= 6 / 10)
obs_mus = numpy.random.uniform(-1, 5, 400)
mus, sigmas, weights = adaptive_parzen_estimator(obs_mus,
low,
high,
prior_weight=1.0,
equal_weight=False,
flat_num=25)
assert len(mus) == len(sigmas) == len(weights) == 400 + 1
assert numpy.all(weights[-25:] == weights[-1])
assert numpy.all(sigmas <= 6) and numpy.all(sigmas >= 6 / 20)
obs_mus = numpy.random.uniform(-1, 5, 10000)
mus, sigmas, weights = adaptive_parzen_estimator(obs_mus,
low,
high,
prior_weight=1.0,
equal_weight=False,
flat_num=25)
assert len(mus) == len(sigmas) == len(weights) == 10000 + 1
assert numpy.all(weights[-25:] == weights[-1])
assert numpy.all(sigmas <= 6) and numpy.all(sigmas >= 6 / 100)