def test_adaptivepnorm_all_particles():
"""Test using rejected particles or not for weighting."""
abc = MockABC(
[
{
's1': -1,
's2': -1,
's3': -1
},
{
's1': -1,
's2': 0,
's3': 1
},
{
's1': -2,
's2': 0.5,
's3': 3
},
],
accepted=[True, True, False],
)
x_0 = {'s1': 0, 's2': 0, 's3': 1}
x_1 = {'s1': 0.5, 's2': 0.4, 's3': -5}
# check that distance values calculated when using rejected particles
# or not differ
dist_all = AdaptivePNormDistance(all_particles_for_scale=True)
dist_all.initialize(0, abc.sample_from_prior, x_0=x_0, total_sims=0)
dist_acc = AdaptivePNormDistance(all_particles_for_scale=False)
dist_acc.initialize(0, abc.sample_from_prior, x_0=x_0, total_sims=0)
assert dist_all(x_1, x_0, t=0) != dist_acc(x_1, x_0, t=0)
def test_adaptivepnormdistance():
"""
Only tests basic running.
"""
abc = MockABC([{
's1': -1,
's2': -1,
's3': -1
}, {
's1': -1,
's2': 0,
's3': 1
}])
x_0 = {'s1': 0, 's2': 0, 's3': 1}
scale_functions = [
median_absolute_deviation, mean_absolute_deviation, standard_deviation,
bias, root_mean_square_deviation,
median_absolute_deviation_to_observation,
mean_absolute_deviation_to_observation,
combined_median_absolute_deviation, combined_mean_absolute_deviation,
standard_deviation_to_observation
]
for scale_function in scale_functions:
dist_f = AdaptivePNormDistance(scale_function=scale_function)
dist_f.initialize(0, abc.sample_from_prior, x_0=x_0)
dist_f(abc.sample_from_prior()[0], abc.sample_from_prior()[1], t=0)
# test max weight ratio
for scale_function in scale_functions:
dist_f = AdaptivePNormDistance(scale_function=scale_function,
max_weight_ratio=20)
dist_f.initialize(0, abc.sample_from_prior, x_0=x_0)
dist_f(abc.sample_from_prior()[0], abc.sample_from_prior()[1], t=0)
def test_store_weights():
"""Test whether storing distance weights works."""
abc = MockABC([{'s1': -1, 's2': -1, 's3': -1},
{'s1': -1, 's2': 0, 's3': 1}])
x_0 = {'s1': 0, 's2': 0, 's3': 1}
weights_file = tempfile.mkstemp(suffix=".json")[1]
print(weights_file)
def distance0(x, x_0):
return abs(x['s1'] - x_0['s1'])
def distance1(x, x_0):
return np.sqrt((x['s2'] - x_0['s2'])**2)
for distance in [AdaptivePNormDistance(log_file=weights_file),
AdaptiveAggregatedDistance(
[distance0, distance1], log_file=weights_file)]:
distance.initialize(0, abc.sample_from_prior, x_0=x_0)
distance.update(1, abc.sample_from_prior)
distance.update(2, abc.sample_from_prior)
weights = load_dict_from_json(weights_file)
assert set(weights.keys()) == {0, 1, 2}
expected = distance.weights
for key, val in expected.items():
if isinstance(val, np.ndarray):
expected[key] = val.tolist()
assert weights == expected
def test_adaptivepnormdistance_initial_weights():
abc = MockABC([{
's1': -1,
's2': -1,
's3': -1
}, {
's1': -1,
's2': 0,
's3': 1
}])
x_0 = {'s1': 0, 's2': 0, 's3': 1}
# first test that for PNormDistance, the weights stay constant
initial_weights = {'s1': 1, 's2': 2, 's3': 3}
dist_f = AdaptivePNormDistance(p=2, initial_scale_weights=initial_weights)
dist_f.initialize(0, abc.sample_from_prior, x_0=x_0, total_sims=0)
assert (dist_f.scale_weights[0] == np.array([1, 2, 3])).all()
# call distance function
d = dist_f(abc.sumstats[0], abc.sumstats[1], t=0)
expected = pow(sum([(2 * 1)**2, (3 * 2)**2]), 1 / 2)
assert expected == d
# check updating works
dist_f.update(1, abc.sample_from_prior, total_sims=0)
assert (dist_f.scale_weights[1] != dist_f.scale_weights[0]).any()
def test_adaptivepnormdistance_initial_weights():
abc = MockABC([{
's1': -1,
's2': -1,
's3': -1
}, {
's1': -1,
's2': 0,
's3': 1
}])
x_0 = {'s1': 0, 's2': 0, 's3': 1}
# first test that for PNormDistance, the weights stay constant
initial_weights = {'s1': 1, 's2': 2, 's3': 3}
dist_f = AdaptivePNormDistance(initial_weights=initial_weights)
dist_f.initialize(0, abc.sample_from_prior, x_0=x_0)
assert dist_f.weights[0] == initial_weights
# call distance function
d = dist_f(abc.sample_from_prior()[0], abc.sample_from_prior()[1], t=0)
expected = pow(sum([(2 * 1)**2, (3 * 2)**2]), 1 / 2)
assert expected == d
# check updating works
dist_f.update(1, abc.sample_from_prior)
assert dist_f.weights[1] != dist_f.weights[0]
def test_adaptivepnormdistance():
"""
Only tests basic running.
"""
# TODO it could be checked that the scale functions lead to the expected
# values
abc = MockABC([{
's1': -1,
's2': -1,
's3': -1
}, {
's1': -1,
's2': 0,
's3': 1
}])
x_0 = {'s1': 0, 's2': 0, 's3': 1}
scale_functions = [
median_absolute_deviation,
mean_absolute_deviation,
standard_deviation,
bias,
root_mean_square_deviation,
std_or_rmsd,
median_absolute_deviation_to_observation,
mean_absolute_deviation_to_observation,
combined_median_absolute_deviation,
mad_or_cmad,
combined_mean_absolute_deviation,
standard_deviation_to_observation,
]
for scale_function in scale_functions:
dist_f = AdaptivePNormDistance(scale_function=scale_function)
dist_f.initialize(0, abc.sample_from_prior, x_0=x_0, total_sims=0)
dist_f(abc.sumstats[0], abc.sumstats[1], t=0)
assert (dist_f.scale_weights[0] != np.ones(3)).any()
# test max weight ratio
for scale_function in scale_functions:
dist_f = AdaptivePNormDistance(scale_function=scale_function,
max_scale_weight_ratio=20)
dist_f.initialize(0, abc.sample_from_prior, x_0=x_0, total_sims=0)
dist_f(abc.sumstats[0], abc.sumstats[1], t=0)
weights = dist_f.scale_weights[0]
assert (weights != np.ones(3)).any()
assert np.max(weights) / np.min(weights[~np.isclose(weights, 0)]) <= 20