def burn_tests():
graph = burn()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'M')
print("P(MapoDoufu | Burn=false) = <", mean, ", ", f_mean, ">")
print()
def decMCMC(self, data, parguess, bounds, steps):
#deconvolution with MCMC
self.deconvolute(data, parguess, bounds, False)
parguess = self.pars
Norm = np.max(data.current) / 100
Y = data.current / Norm
corr = np.asarray(
[Norm if n % 3 == 0 else 1 for n in np.arange(len(parguess))])
_, bounds, corr = self.checkParams(parguess, np.asarray(bounds), corr,
self.shape)
parguess = [par / cor for par, cor in zip(self.pars, corr)]
try:
mcmc = MCMC(self.model, parguess, bounds)
mcmc.pg.connect(self.pg.emit)
st = 0.5 * np.ones(len(parguess))
st[[
int(np.sum(self.args[:i]) + 2)
for i in range(0, len(self.args))
]] = 0.1
st[[
int(np.sum(self.args[:i + 1]) - 1)
for i in range(0, len(self.args)) if self.args[i] == 4
]] = 0.01
# mcmc.step = [0.1 if (n-1)%3 == 0 else 0.5 for n in np.arange(len(parguess))]
mcmc.step = st
self.pars, self.perr = mcmc(data.X, Y, steps, corr=corr)
self.pars_u = [
ufloat(p, abs(e)) for p, e in zip(self.pars, self.perr)
]
#Calculate each peak
for i, name in enumerate(self.names):
indx = int(np.sum(self.args[:i]))
self.peaks[name] = self.Peak(self,
data.X,
*self.pars[indx:indx +
self.args[i]],
shape=self.shape[i])
self.peaks_u[name] = self.Peak(self,
data.X,
*self.pars_u[indx:indx +
self.args[i]],
shape=self.shape[i],
uncert=True)
self.peaks['cumulative'] = self.model(
data.X, *self.pars) #save the fit result
self.peaks_u['cumulative'] = self.model(
data.X, *self.pars_u, uncert=True) #save the fit result
self.fitGoodness = self.goodness(data.current,
self.peaks['cumulative'])
self.FWHM() #calculate fwhm
self.Area() #calculate the areas
self.printResult(data) #print the fit report
except Exception as e:
print('exception\n', e)
def tanks_tests():
graph = tanks()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(50, 100000)
Tests.mixing_plot(samples, 'num_tanks')
Tests.plot_distribution(samples, 'num_tanks')
Tests.mixing_plot(samples, 'A')
Tests.plot_distribution(samples, 'A')
def faculty_evaluation_tests():
graph = faculty_evals()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(500, 10000)
Tests.mixing_plot(samples, 'mu')
Tests.mixing_plot(samples, 'sigma2')
Tests.plotposterior([s['mu'] for s in samples], faculty_mean_prior, 'mean', 5.0, 6.5)
Tests.plotposterior([s['sigma2'] for s in samples], faculty_var_prior, 'var', 0.0001, 1.0)
def progress_tests():
graph = progress()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(2000, 50000)
Tests.plot_distribution(samples, 'Encounters')
encounters = [s['Encounters'] for s in samples]
median = np.median(encounters)
print('median', median)
def golfer_network_tests():
graph = golf()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(1000, 100000)
print('woohoo! the golfers finished!')
with open('golf_samples.pickle', 'wb') as gs:
pickle.dump(samples, gs)
with open('golf_graph.pickle', 'wb') as gg:
pickle.dump(graph, gg)
def __init__(self, NP, means, mins, maxs, sds, outfile, errlev=0.1, goodchi2=350.0):
"""
Instantiates the class by synthetically generating data.
"""
MCMC.__init__(self, TargetAcceptedPoints=1000, NumberOfParams=NP, Mins=mins, Maxs=maxs, SDs=sds, \
write2file=True, outputfilename=outfile, alpha=0.1, debug=False,\
EstimateCovariance=True, CovNum=100, goodchi2=goodchi2)
lcurv.readmap()
lcurv.mockdata(means,errlev)
def hyper_alarm_inference():
legs = [10, 25, 50, 100, 1000]
model = 'lab'
for n in legs:
name = model + '-inf_' + str(n)
graph = hyper_alarm(val_dict=name, inference=True)
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(2000, 100000)
mean, f_mean = Tests.sample_dim(samples, 'B')
print(mean)
def pareto_poisson_tests():
graph = pareto()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(2000, 100000)
for node in graph.hidden_nodes:
#Tests.mixing_plot(samples, node.name)
Tests.plot_distribution(samples, node.name)
encounters = [s[node.name] for s in samples]
median = np.median(encounters)
print('median', median)
def thomas_tests():
graph = thomas()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'Cross')
print("P(Cross | Diesel=true, Diesel10=true) = <", mean, ", ", f_mean, ">")
graph = thomas()
graph.node_dict['Diesel'].observed = False
graph.node_dict['Diesel10'].observed = False
graph.node_dict['Thomas'].observed = True
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'Cross')
print("P(Cross | Thomas=true) = <", mean, ", ", f_mean, ">")
graph = thomas()
graph.node_dict['Diesel'].observed = False
graph.node_dict['Diesel10'].observed = False
graph.node_dict['Cross'].observed = True
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'Diesel')
print("P(Diesel | Cross=true) = <", mean, ", ", f_mean, ">")
print()
def wacky_network_tests():
graph = wacky()
# graph.node_dict['G'].observed = True
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(20000, 2000000)
with open('wacky_G_samples.pickle', 'wb') as ws:
pickle.dump(samples, ws)
with open('wacky_G_graph.pickle', 'wb') as wg:
pickle.dump(graph, wg)
for node in graph.nodes:
Tests.mixing_plot(samples, node.name)
Tests.plot_distribution(samples, node.name)
def hyper_faculty_tests():
meta_samples = []
for n in range(5):
graph = faculty_evals_1hyper(n)
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 100000)
print(n)
Tests.plot_multi(samples, ['mu', 'sigma2'])
print(graph.nodes)
for node in graph.hidden_nodes:
if node.name != 'mu' and node.name != 'sigma2':
Tests.mixing_plot(samples, node.name)
Tests.plot_distribution(samples, node.name)
def aufg15c():
normal = MCMC(2, -3, 2)
sample = normal.sample(15, 10000)
bins = plt.hist(sample, bins=50, density=True, label="Gezogen")[1]
x = np.linspace(bins[0], bins[-1], 1000)
plt.plot(x, normal.pdf(x), label="PDF")
plt.xlabel("$x$")
plt.ylabel("Wahrscheinlichkeitsdichte")
plt.legend()
# plt.show()
plt.savefig("A15c.pdf")
plt.clf()
return sample
def SMC(n, k, N, n_steps, verbose=False):
# Trigger timer
start = time.time()
# Build the puzzle
puzzle = Puzzle(n, k)
print('Number of pieces in puzzle %s' % puzzle.d)
# Initialize the algorithm
particles = (np.ones((N, puzzle.d))*np.arange(puzzle.d)).astype(int)
particles = np.array([np.random.permutation(particles[i][:]) for i in range(N)])
t = 0
av_loss = []
# Annealing kernel
kernel = Kernel(t)
# Compute initial loss
av_loss.append(np.mean(puzzle.loss_global(particles)))
if verbose:
print('Initial average number of non matching edges: %s' % av_loss[-1])
print('')
while t < max_iter:
if verbose:
print('Step %s' % (t+1))
# Annealing loss
loss = TargetLoss(puzzle.loss, 2*np.log(t + np.exp(1)))
# Re-sample for importance weights
isampler = IS(loss)
isampler.fit(particles)
particles = isampler.resample()
# Forward search
mcmc = MCMC(loss, kernel, n_steps)
particles = mcmc.forward(particles)
# Track performance
av_loss.append(np.mean(puzzle.loss_global(particles)))
if verbose:
print('Current average number of non matching edges: %s' % av_loss[-1])
print('------------------------------------------------------------------')
# Update t
t += 1
end = time.time()
return av_loss, t, end-start
def hyper_alarm_learning_tests():
legs = [1, 10, 25, 50, 75, 100, 250, 500, 750, 1000]
prior = 'lab'
model = '01_50'
for n in legs:
graph = hyper_alarm_learn('alarm-gen-' + model + '.json', n=n, val_dict=prior)
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(1000, 1000)
mean, f_mean = Tests.sample_dim(samples, 'b_B')
print(mean, f_mean)
#mean, f_mean = Tests.sample_dim(samples, 'b_E')
#print(mean, f_mean)
name = 'alarm-' + model + '_' + str(n) + '_samples.pickle'
with open(name, 'wb') as f:
pickle.dump(samples, f)
def setUp(self):
freq = {'A': .292, 'C': .213, 'G': .237, 'T': .258}
k = 22.2
alpha = .0005
tau = 1
draws = 100
filename = 'infile'
self.mcmc = MCMC(freq, k, alpha, tau, draws, filename)
def hyper_alarm_generate():
val_dict = 'orig'
graph = hyper_alarm(val_dict)
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 30000)
mean, f_mean = Tests.sample_dim(samples, 'B')
print(mean, f_mean)
mean, f_mean = Tests.sample_dim(samples, 'E')
print(mean, f_mean)
saved_samples = []
for i, sample in enumerate(samples):
if i < 3001 and i % 3 == 0:
#print(sample)
saved_samples.append(sample)
with open('alarm-gen-' + val_dict + '.json', 'w') as f:
json.dump(saved_samples,f)
def __init__(self,
NP,
means,
mins,
maxs,
sds,
outfile,
errlev=0.1,
goodchi2=350.0):
"""
Instantiates the class by synthetically generating data.
"""
MCMC.__init__(self, TargetAcceptedPoints=1000, NumberOfParams=NP, Mins=mins, Maxs=maxs, SDs=sds, \
write2file=True, outputfilename=outfile, alpha=0.1, debug=False,\
EstimateCovariance=True, CovNum=100, goodchi2=goodchi2)
lcurv.readmap()
lcurv.mockdata(means, errlev)
def beta_bernoulli_tests():
graph = beta_bernoulli()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(5000, 100000)
Tests.plotposterior([s['A'] for s in samples], beta_expected_t, 'beta-bernoulli', 0, 1)
graph = beta_bernoulli(b=0)
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(13000, 100000)
Tests.plotposterior([s['A'] for s in samples], beta_expected_f, 'beta-bernoulli', 0, 1)
def home_or_school_tests():
graph = home_or_school()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'AS')
print("P(AS | IA=false) = <", mean, ", ", f_mean, ">")
graph = home_or_school()
mcmc = MCMC({'IA': 1}, graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'AS')
print("P(AS | IA=false) = <", mean, ", ", f_mean, ">")
print()
def dirty_roommates_tests():
graph = dirty_roommates()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'CK')
print("P(CK | DR=false) = <", mean, ", ", f_mean, ">")
graph = dirty_roommates()
graph.node_dict['DR'].observed = False
graph.node_dict['CK'].observed = True
mcmc = MCMC({'CK': 1}, graph=graph)
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'IT')
print("P(IT | CK=true) = <", mean, ", ", f_mean, ">")
class MCMCTest(unittest.TestCase):
def setUp(self):
freq = {'A': .292, 'C': .213, 'G': .237, 'T': .258}
k = 22.2
alpha = .0005
tau = 1
draws = 100
filename = 'infile'
self.mcmc = MCMC(freq, k, alpha, tau, draws, filename)
def test_constructor(self):
self.assertEqual(self.mcmc.tau, 1)
def test_fun_matrix(self):
self.assertEqual(len(self.mcmc.matrix), 64)
def test_topology(self):
nodes = self.mcmc.tree.get_internal()
self.mcmc.target = nodes[0]
self.mcmc.generate_topology(self.mcmc.target)
# print [node.sequence for node in self.mcmc.old_topo]
# print [node.sequence for node in self.mcmc.new_topo]
def test_calc_matrix(self):
nodes = self.mcmc.tree.nodes()
# print len(nodes)
# for node in nodes:
# print node.sequence
self.mcmc.target = nodes[1]
# print self.mcmc.target.sequence
# print self.mcmc.target.parent.sequence
# print self.mcmc.target.sibling.sequence
# print self.mcmc.target.left.sequence
# print self.mcmc.target.right.sequence
self.mcmc.calc_fun_matrix(self.mcmc.target, self.mcmc.new_topo)
self.assertEqual(len(self.mcmc.matrix), 64)
def test_select_time(self):
nodes = self.mcmc.tree.get_internal()
self.mcmc.target = nodes[0]
topo = self.mcmc.generate_topology(self.mcmc.target)
self.mcmc.calc_fun_matrix(self.mcmc.target, self.mcmc.new_topo)
def test_select_seq(self):
nodes = self.mcmc.tree.get_internal()
self.mcmc.target = nodes[0]
target = self.mcmc.target
topo = self.mcmc.generate_topology(self.mcmc.target)
self.mcmc.calc_fun_matrix(self.mcmc.target, self.mcmc.new_topo)
time = self.mcmc.select_time(target, topo)
result = self.mcmc.select_sequence(target, topo, time)
# print target.sequence
# print result
# print target.time
# print time
def test_accept(self):
nodes = self.mcmc.tree.get_internal()
self.mcmc.target = nodes[0]
target = self.mcmc.target
topo = self.mcmc.generate_topology(self.mcmc.target)
self.mcmc.calc_fun_matrix(self.mcmc.target, self.mcmc.new_topo)
def alarm_tests():
mcmc = MCMC()
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'B')
print("P(Burglary | JohnCalls=true, MaryCalls=true) = <", mean, ", ", f_mean, ">")
mean, f_mean = Tests.sample_dim(samples, 'A')
print("P(Alarm | JohnCalls=true, MaryCalls=true) = <", mean, ", ", f_mean, ">")
mean, f_mean = Tests.sample_dim(samples, 'E')
print("P(Earthquake | JohnCalls=true, MaryCalls=true) = <", mean, ", ", f_mean, ">")
mcmc = MCMC({'J': 1., 'M': 0.})
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'B')
print("P(Burglary | JohnCalls=true, MaryCalls=false) = <", mean, ", ", f_mean, ">")
mcmc = MCMC({'J': 1.})
mcmc.graph.node_dict['M'].observed = False
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'B')
print("P(Burglary | JohnCalls=true) = <", mean, ", ", f_mean, ">")
mcmc = MCMC({'M': 1.})
mcmc.graph.node_dict['J'].observed = False
samples = mcmc.gibbs(10000, 10000)
mean, f_mean = Tests.sample_dim(samples, 'B')
print("P(Burglary | MaryCalls=true) = <", mean, ", ", f_mean, ">")
print()
def normal_normal_tests():
graph = normal_normal()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(500, 10000)
Tests.plotposterior([s['A'] for s in samples], normal_expected, 'normal-normal', -3, 3)
np.savetxt("./mcmc/log_theta_sigma2.txt",
dgp.session.run(dgp.log_theta_sigma2),
fmt="%f",
delimiter='\t')
np.savetxt("./mcmc/log_theta_lengthscale.txt",
dgp.session.run(dgp.log_theta_lengthscale),
fmt="%f",
delimiter='\t')
np.savetxt("./mcmc/log_lambda.txt",
dgp.session.run([dgp.likelihood.log_var]),
fmt="%f",
delimiter='\t')
## Run the MCMC sampler and save some quantities of interest for comparison with the variational approximation
samples_F1, samples_F2, predictions_F1, predictions_F2 = MCMC(
data.X, data.Y, test.X)
np.savetxt("./mcmc/predictions_MCMC_F1.txt",
predictions_F1,
fmt="%f",
delimiter='\t')
np.savetxt("./mcmc/predictions_MCMC_F2.txt",
predictions_F2,
fmt="%f",
delimiter='\t')
np.savetxt("./mcmc/samples_MCMC_F1.txt",
samples_F1,
fmt="%f",
delimiter='\t')
np.savetxt("./mcmc/samples_MCMC_F2.txt",
samples_F2,
fmt="%f",
count = 0
for i in range(len(x)):
xsum = +x[i]
count = +i + 1
cumsum[i] = xsum / count
plt.plot(cumsum, label=u'E(X) Convergence')
plt.show()
#########################################################################
import numpy as np
import matplotlib.pyplot as plt
from mcmc import MCMC, DataDistribution
mu1 = [-2.0, 2.0, -5.0, -10.0]
sd1 = [1.0, 0.5, 0.5, 0.5]
w1 = [0.2, 0.4, 0.3, 0.1]
mixtureDens1 = DataDistribution(mu1, sd1, w1)
x0 = np.random.rand(1)
sigma = 5
n = int(10000)
sim = MCMC(x0=x0, sigma=sigma, n=n)
x_out = sim.mcmc(mixtureDens1.gaussian_mixture)
sim.plotHistorgram(mixtureDens1.gaussian_mixture, x_out, xrange=[-15, 5])
sim.plotConvergence(x_out)
def gamma_poisson_tests():
graph = gamma_poisson()
mcmc = MCMC(graph=graph)
samples = mcmc.gibbs(50000, 10000)
Tests.plotposterior([s['L'] for s in samples], gamma_expected, 'gamma-poisson', 0, 12)