def sample_fun(self, model, **sampler_options):
params_array = hyperparameter_utils.params_to_array(self.params)
if model.has_data:
K_XX = model.noiseless_kernel.cov(model.inputs)
current_L = spla.cholesky(K_XX, lower=True)
nu = spla.solve_triangular(current_L, model.latent_values.value-model.mean.value, lower=True)
else:
nu = None # if no data
new_params, current_ll = slice_sample(params_array, self.logprob, model, nu, **sampler_options)
new_latent_values = self._compute_implied_y(model, nu)
return new_params, new_latent_values, current_ll
def sample(self, model):
"""generate a new sample of parameters for the model
Notes
-----
The parameters are stored as self.params which is a list of Params objects.
The values of the parameters are updated on each call. Pesumably the value of
the parameter affects the model (this is not required, but it would be a bit
pointless othewise)
"""
# turn self.params into a 1d numpy array
params_array = hyperparameter_utils.params_to_array(self.params)
for i in range(self.thinning + 1):
# get a new value for the parameter array via slice sampling
params_array, current_ll = slice_sample(params_array, self.logprob, model, **self.sampler_options)
hyperparameter_utils.set_params_from_array(self.params, params_array) # Can this be untabbed safely?
self.current_ll = current_ll # for diagnostics
def sample(self, gp_likelihood):
"""generate a new sample of parameters for the model
Notes
-----
The parameters are stored as self.params which is a list of Params objects.
The values of the parameters are updated on each call. Pesumably the value of
the parameter affects the model (this is not required, but it would be a bit
pointless othewise)
"""
# turn self.params into a 1d numpy array
for i in xrange(self.thinning + 1):
# get a new value for the parameter array via slice sampling
self.params, current_ll = slice_sample(self.params,
self.logprob,
gp_likelihood,
step_out=False)
self.current_ll = current_ll # for diagnostics
return self.params.copy()
import matplotlib.pyplot as plt
n = 10000
# Test on 1D Gaussian
x_samples = np.zeros(n)
x = np.zeros(1)
gsn = priors.Gaussian(mu = -1, sigma = 4)
for i in xrange(n):
if i % 1000 == 0:
print('Sample %d/%d' % (i,n))
x, cur_ll = slice_sample(x, gsn. logprob)
x_samples[i] = x.copy()
print('1D Gaussian actual mean: %f, mean of samples: %f' % (-1, np.mean(x_samples)))
print('1D Gaussian actual sigma: %f, std of samples: %f' % (4, np.std(x_samples)))
plt.figure(1)
plt.clf()
plt.hist(x_samples, 40)
plt.savefig('slice_sampler_test.pdf')
# Test on 2D Gaussian
mu = np.array([-2, 5])
a = npr.rand(2,2)
cov = np.dot(a,a.T)
#r.par(mfrow=[2,2]) # Exhaustive computation to compare to slice sampler log_concs = np.arange(-50,20,.1) lls = np.array([ loglik(log_conc) for log_conc in log_concs ]) mle = log_concs[lls.argmax()] print "MLE", np.exp(mle) lps = np.array([ logpost(log_conc) for log_conc in log_concs ]) map = log_concs[lps.argmax()] print "MAP", np.exp(map) #r.plot(log_concs, lls, xlab='x',ylab='ll') #plot(log_concs, lls, main='loglik') #plot(log_concs, lps, main='unnorm logpost') # For comparisons real_s = r.sample(log_concs,len(histories)*2,prob=np.exp(lps),replace=True) # Slice sampler h = mcmc.slice_sample(logpost, 0.0, 1.0, niter=1000) h = np.array(h) # Replications of single slice-sample #for i in range(500): # init = histories[i][-1] # h = mcmc.slice_sample(logpost, init, 1.0, niter=200) # histories[i] += h