def test_batch_constrain(self):
# Test that batch constrain returns the correct boolean array for
# the given model, obs and params
obs_uncertainty = self.training_ensemble.data.std(axis=0)
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
sampler = ABCSampler(self.m, obs,
obs_uncertainty=obs_uncertainty/obs.data,
interann_uncertainty=0.,
repres_uncertainty=0.,
struct_uncertainty=0.)
# Calculate the implausbility of the training points from a perturbed
# training point. The emulator variance should be zero making testing
# easier.
valid_samples = sampler.batch_constrain(self.training_params,
tolerance=0., threshold=2.)
# The implausibility for the 10th sample (the one we perturbed around)
# should be around one (and hence valid), some neighbouring points are
# also valid, the rest should be invalid
expected = np.asarray([True, False, False, True, False,
True, False, False, True, False,
True, False, False, True, True,
True, True, True, True, True,
False, False, False, False, False])
assert_array_equal(valid_samples, expected)
def test_implausibility_scalar_uncertainty_interface(self):
# Test the interface correctly deals with 2d model and obs
obs_uncertainty = 5.
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
sampler = ABCSampler(self.m, obs,
obs_uncertainty=obs_uncertainty/obs.data.mean(),
interann_uncertainty=0.,
repres_uncertainty=0.,
struct_uncertainty=0.)
implausibility = sampler.get_implausibility(self.training_params)
expected_shape = (len(self.training_params), ) + obs.shape
assert implausibility.shape == expected_shape
def test_sample(self):
# Test that batch constrain returns the correct boolean array for
# the given model, obs and params
obs_uncertainty = self.training_ensemble.data.std(axis=0)
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
sampler = ABCSampler(self.m, obs,
obs_uncertainty=obs_uncertainty/obs.data,
interann_uncertainty=0.,
repres_uncertainty=0.,
struct_uncertainty=0.)
# Generate only valid samples
valid_samples = sampler.sample(n_samples=100, tolerance=0., threshold=2.)
assert valid_samples.shape == (100, 3)
def test_implausibility_repres(self):
# Test the implausibility is correct
obs_uncertainty = 5.
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
sampler = ABCSampler(self.m, obs,
obs_uncertainty=0.,
interann_uncertainty=0.,
repres_uncertainty=obs_uncertainty/obs.data.mean(),
struct_uncertainty=0.)
# Calculate the implausbility of the training points from a perturbed
# training point. The emulator variance should be zero making testing
# easier.
implausibility = sampler.get_implausibility(self.training_params)
# The implausibility for the 10th sample (the one we perturbed around)
# should be one - on average
assert_allclose(implausibility.data[10, :].mean(), 1., rtol=1e-2)
def test_implausibility_vector_uncertainty(self):
# Test with a vector obs uncertainty
obs_uncertainty = self.training_ensemble.data.std(axis=0)
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
sampler = ABCSampler(self.m, obs,
obs_uncertainty=obs_uncertainty/obs.data,
interann_uncertainty=0.,
repres_uncertainty=0.,
struct_uncertainty=0.)
# Calculate the implausbility of the training points from a perturbed
# training point. The emulator variance should be zero making testing
# easier.
implausibility = sampler.get_implausibility(self.training_params)
expected_shape = (len(self.training_params),) + obs.shape
assert implausibility.shape == expected_shape
def test_struct_uncertainty_specified_twice(self):
# Test with a vector obs uncertainty
obs_uncertainty = self.training_ensemble.data.std(axis=0)
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
with pytest.raises(ValueError):
sampler = ABCSampler(self.m, obs,
struct_uncertainty=obs_uncertainty,
abs_struct_uncertainty=obs_uncertainty)
def test_batch_constrain(self):
# Test that batch constrain returns the correct boolean array for
# the given model, obs and params
obs_uncertainty = self.training_ensemble.data.std(axis=0)
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
sampler = ABCSampler(self.m, obs,
obs_uncertainty=obs_uncertainty/obs.data,
interann_uncertainty=0.,
repres_uncertainty=0.,
struct_uncertainty=0.)
# Calculate the implausbility of the training points from a perturbed
# training point. The emulator variance should be zero making testing
# easier.
valid_samples = sampler.batch_constrain(self.training_params,
tolerance=0., threshold=2.)
expected_shape = (len(self.training_params),)
assert valid_samples.shape == expected_shape
def test_absolute_struct_uncertainty(self):
# Test with a vector obs uncertainty
obs_uncertainty = self.training_ensemble.data.std(axis=0)
# Perturbing the obs by one sd should lead to an implausibility of 1.
obs = self.training_ensemble[10].copy() + obs_uncertainty
sampler = ABCSampler(self.m, obs,
abs_obs_uncertainty=0.,
interann_uncertainty=0.,
repres_uncertainty=0.,
abs_struct_uncertainty=obs_uncertainty)
# Calculate the implausbility of the training points from a perturbed
# training point. The emulator variance should be zero making testing
# easier.
implausibility = sampler.get_implausibility(self.training_params)
# The implausibility for the 10th sample (the one we perturbed around)
# should be one.
expected = np.ones((100,))
# The first element has zero variance
expected[0] = 0.
assert_allclose(implausibility.data[10, :], expected, rtol=1e-1)