def brute_force_marginal_likelihood(self, X, n_samples=10000, gen_seed=0):
"""
Calculates the log marginal likelihood via brute force method in which
parameters (mu and kappa) are repeatedly drawn from the prior, the
likelihood is calculated for each set of parameters, then the average is
taken.
Inputs:
X: A column of data (numpy)
n_samples: the number of draws
gen_Seed: seed for the rng
"""
check_data_type_column_data(X)
if type(n_samples) is not int:
raise TypeError("n_samples should be an int")
if n_samples <= 0:
raise ValueError("n_samples should be greater than 0")
if type(gen_seed) is not int:
raise TypeError("gen_seed should be an int")
N = float(len(X))
rng = random.Random(gen_seed)
log_likelihoods = [0]*n_samples
for i in range(n_samples):
params = self.sample_parameters_given_hyper(gen_seed=next_seed(rng))
log_likelihoods[i] = self.log_likelihood(X, params)
log_marginal_likelihood = logmeanexp(log_likelihoods)
return log_marginal_likelihood
def brute_force_marginal_likelihood(self, X, n_samples=10000, gen_seed=0):
"""
Calculates the log marginal likelihood via brute force method in which
parameters (weights) are repeatedly drawn from the prior, the
likelihood is calculated for each set of parameters, then the average is
taken.
Inputs:
X: A column of data (numpy)
n_samples: the number of draws
gen_Seed: seed for the rng
"""
check_data_type_column_data(X)
if type(n_samples) is not int:
raise TypeError("n_samples should be an int")
if n_samples <= 0:
raise ValueError("n_samples should be greater than 0")
if type(gen_seed) is not int:
raise TypeError("gen_seed should be an int")
hypers = self.get_hypers()
K = hypers[b'K']
check_data_vs_k(X,K)
rng = numpy.random.RandomState(gen_seed)
log_likelihoods = [0]*n_samples
for i in range(n_samples):
params = self.sample_parameters_given_hyper(gen_seed=next_seed(rng))
log_likelihoods[i] = self.log_likelihood(X, params)
log_marginal_likelihood = logmeanexp(log_likelihoods)
return log_marginal_likelihood
def predictive_probability_multistate(M_c, X_L_list, X_D_list, Y, Q):
"""
Returns the predictive probability, averaged over each sample.
"""
logprobs = [float(predictive_probability(M_c, X_L, X_D, Y, Q))
for X_L, X_D in zip(X_L_list, X_D_list)]
return logmeanexp(logprobs)
def test_logmeanexp():
inf = float('inf')
nan = float('nan')
assert gu.logmeanexp([]) == -inf
assert relerr(992.550919866405, gu.logmeanexp(range(1000))) < 1e-15
assert gu.logmeanexp([-1000., -1000.]) == -1000.
assert relerr(math.log(0.5 * (1 + math.exp(-1.))),
gu.logmeanexp([0., -1.])) \
< 1e-15
assert relerr(math.log(0.5), gu.logmeanexp([0., -1000.])) < 1e-15
assert relerr(-3 - math.log(2.), gu.logmeanexp([-inf, -3])) < 1e-15
assert relerr(-3 - math.log(2.), gu.logmeanexp([-3, -inf])) < 1e-15
assert gu.logmeanexp([+inf, -3]) == +inf
assert gu.logmeanexp([-3, +inf]) == +inf
assert gu.logmeanexp([-inf, 0, +inf]) == +inf
assert math.isnan(gu.logmeanexp([nan, inf]))
assert math.isnan(gu.logmeanexp([nan, -3]))
assert math.isnan(gu.logmeanexp([nan]))
def test_logmeanexp():
inf = float('inf')
nan = float('nan')
assert gu.logmeanexp([]) == -inf
assert relerr(992.550919866405, gu.logmeanexp(range(1000))) < 1e-15
assert gu.logmeanexp([-1000., -1000.]) == -1000.
assert relerr(math.log(0.5 * (1 + math.exp(-1.))),
gu.logmeanexp([0., -1.])) \
< 1e-15
assert relerr(math.log(0.5), gu.logmeanexp([0., -1000.])) < 1e-15
assert relerr(-3 - math.log(2.), gu.logmeanexp([-inf, -3])) < 1e-15
assert relerr(-3 - math.log(2.), gu.logmeanexp([-3, -inf])) < 1e-15
assert gu.logmeanexp([+inf, -3]) == +inf
assert gu.logmeanexp([-3, +inf]) == +inf
assert gu.logmeanexp([-inf, 0, +inf]) == +inf
assert math.isnan(gu.logmeanexp([nan, inf]))
assert math.isnan(gu.logmeanexp([nan, -3]))
assert math.isnan(gu.logmeanexp([nan]))
