def test_autocorrelation_short_chain():
"""Check that the autocorrelation
reacts nicely to small sample numbers."""
problem = gaussian_problem()
sampler = sample.MetropolisSampler()
# optimization
result = optimize.minimize(problem, n_starts=3)
# sample
result = sample.sample(
problem, sampler=sampler, n_samples=10, result=result)
# manually set burn in to chain length (only for testing!!)
chain_length = result.sample_result.trace_x.shape[1]
result.sample_result.burn_in = chain_length
# run auto-correlation
ac = sample.auto_correlation(result)
assert ac is None
# run effective sample size
ess = sample.effective_sample_size(result)
assert ess is None
def test_autocorrelation_pipeline():
"""Check that the autocorrelation test works."""
problem = gaussian_problem()
sampler = sample.MetropolisSampler()
# optimization
result = optimize.minimize(problem, n_starts=3)
# sample
result = sample.sample(
problem, sampler=sampler, n_samples=1000, result=result)
# run auto-correlation with previous geweke
sample.geweke_test(result)
ac1 = sample.auto_correlation(result)
# run auto-correlation without previous geweke
result.sample_result.burn_in = None
ac2 = sample.auto_correlation(result)
assert ac1 == ac2
# run effective sample size with previous geweke
# and autocorrelation
ess1 = sample.effective_sample_size(result)
# run effective sample size without previous geweke
# and autocorrelation
result.sample_result.burn_in = None
result.sample_result.auto_correlation = None
ess2 = sample.effective_sample_size(result)
assert ess1 == ess2
def test_multiple_startpoints():
problem = gaussian_problem()
x0s = [np.array([0]), np.array([1])]
sampler = sample.ParallelTemperingSampler(
internal_sampler=sample.MetropolisSampler(), n_chains=2)
result = sample.sample(problem, n_samples=10, x0=x0s, sampler=sampler)
assert result.sample_result.trace_neglogpost.shape[0] == 2
assert [
result.sample_result.trace_x[0][0], result.sample_result.trace_x[1][0]
] == x0s
def test_geweke_test_unconverged():
"""Check that the geweke test reacts nicely to small sample numbers."""
problem = gaussian_problem()
sampler = sample.MetropolisSampler()
# optimization
result = optimize.minimize(problem, n_starts=3)
# sample
result = sample.sample(
problem, sampler=sampler, n_samples=100, result=result)
# run geweke test (should not fail!)
sample.geweke_test(result)
def test_samples_cis():
"""
Test whether :py:func:`pypesto.sample.calculate_ci_mcmc_sample` produces
percentile-based credibility intervals correctly.
"""
# load problem
problem = gaussian_problem()
# set a sampler
sampler = sample.MetropolisSampler()
# optimization
result = optimize.minimize(problem, n_starts=3, filename=None)
# sample
result = sample.sample(problem,
sampler=sampler,
n_samples=2000,
result=result,
filename=None)
# run geweke test
sample.geweke_test(result)
# get converged chain
converged_chain = np.asarray(
result.sample_result.trace_x[0, result.sample_result.burn_in:, :])
# set confidence levels
alpha_values = [0.99, 0.95, 0.68]
# loop over confidence levels
for alpha in alpha_values:
# calculate parameter samples confidence intervals
lb, ub = sample.calculate_ci_mcmc_sample(result, ci_level=alpha)
# get corresponding percentiles to alpha
percentiles = 100 * np.array([(1 - alpha) / 2, 1 - (1 - alpha) / 2])
# check result agreement
diff = np.percentile(converged_chain, percentiles, axis=0) - [lb, ub]
assert (diff == 0).all()
# check if lower bound is smaller than upper bound
assert (lb < ub).all()
# check if dimmensions agree
assert lb.shape == ub.shape