]
if dnm in lrdnms:
from model_lr import *
else:
from model_poiss import *
print('running ' + str(dnm) + ' ' + str(alg) + ' ' + str(ID))
if not os.path.exists('results/'):
os.mkdir('results')
if not os.path.exists('results/' + dnm + '_samples.npy'):
print('No MCMC samples found -- running STAN')
# run sampler
N_samples = 10000
sampler(dnm, dnm in lrdnms, '../data/', 'results/', N_samples)
print('Loading dataset ' + dnm)
Z, Zt, D = load_data('../data/' + dnm + '.npz')
print('Loading posterior samples for ' + dnm)
samples = np.load('results/' + dnm + '_samples.npy')
# TODO FIX SAMPLER TO NOT HAVE TO DO THIS
samples = np.hstack((samples[:, 1:], samples[:, 0][:, np.newaxis]))
# fit a gaussian to the posterior samples
# used for pihat computation for Hilbert coresets with noise to simulate uncertainty in a good pihat
mup = samples.mean(axis=0)
Sigp = np.cov(samples, rowvar=False)
# create the prior -- also used for the above purpose
mu0 = np.zeros(mup.shape[0])
if not stan_samples: # use laplace approximation (save, load from results/)
if not os.path.exists('results/' + dnm + '_samples.npy'):
print('sampling using laplace')
mup_laplace, LSigp_laplace, LSigpInv_laplace = get_laplace(
np.ones(Z.shape[0]), Z, Z.mean(axis=0)[:D], diag=samplediag)
samples_laplace = mup_laplace + np.random.randn(
N_samples, mup_laplace.shape[0]).dot(LSigp_laplace.T)
np.save(os.path.join('results/' + dnm + '_samples.npy'),
samples_laplace)
else:
print('Loading posterior samples for ' + dnm)
samples = np.load('results/' + dnm + '_samples.npy', allow_pickle=True)
else: # use stan sampler (save, load from results/pystan_samples/)
if not os.path.exists('results/' + dnm + '_samples.npy'):
print('No MCMC samples found -- running STAN')
sampler(dnm, True, '../data/', 'results/pystan_samples/', N_samples)
else:
print('Loading posterior samples for ' + dnm)
samples = np.load('results/' + dnm + '_samples.npy', allow_pickle=True)
samples = np.hstack((samples[:, 1:], samples[:, 0][:, np.newaxis]))
#fit a gaussian to the posterior samples
#used for pihat computation for Hilbert coresets with noise to simulate uncertainty in a good pihat
mup = samples.mean(axis=0)
Sigp = np.cov(samples, rowvar=False)
LSigp = np.linalg.cholesky(Sigp)
LSigpInv = sl.solve_triangular(LSigp,
np.eye(LSigp.shape[0]),
lower=True,
overwrite_b=True,
check_finite=False)
X = np.linspace(0, 50, 1000)
y = [0] * len(events)
plt.plot(X, f1(X))
plt.plot(events, y, 'r+')
plt.show()
theta = [1, 10, 0.002]
s_f = theta[0]
l = theta[1]
s_n = theta[2]
V = 50
upper_bound = 2
t1 = 0
t2 = 50
mcmc_sample = sampler(t1, t2, theta, upper_bound)
#obs_test, M_test, thinned_pos_test,gp_thinned_test,gp_obs_test = M_sampler(burnin = 100, niter = 400)
obs_test, M_test, thinned_pos_test, gp_thinned_test, gp_obs_test = mcmc_sample.M_sampler(
events, f1(np.array(events)), burnin=100, niter=1)
def sigmoid(z):
return 1 / (1 + np.exp(-z))
def extract_position(length):
# extract samples postition and gp function value
pos= list(filter(lambda x:x != None, \
[thinned_pos_test[i] if x==length else None for i,x in enumerate(M_test)]))
gp_pos= list(filter(lambda x:x != None, \