def test_lognormal():
import biggles
import esutil as eu
from esutil.random import LogNormal, srandu
from esutil.stat import histogram
n=1000
nwalkers=100
burnin=100
nstep=100
mean=8
sigma=3
ln=LogNormal(mean,sigma)
vals=ln.sample(n)
binsize=0.5
plt=eu.plotting.bhist(vals, binsize=binsize,show=False)
h=histogram(vals, binsize=binsize,more=True)
herr=sqrt(h['hist'])
herr=herr.clip(1.0, herr.max())
guess=[n*(1. + .1*srandu()),
mean*(1. + .1*srandu()),
sigma*(1. + .1*srandu())]
guess=[n*binsize,mean,sigma]
print 'guess:',guess
nlf=LogNormalFitter(h['center'], h['hist'], guess, nwalkers, burnin, nstep,
yerr=herr)
print nlf
res=nlf.get_result()
model=nlf.get_model()
yvals=model.scaled(h['center'])
plt.add(biggles.Curve(h['center'], yvals, color='blue'))
plt.show()
def test_combined(show=False,
nwalkers=20,
burnin=25,
nrand=500,
ntry=1,
gerr=0.1):
import esutil as eu
from esutil.random import LogNormal, Normal, NormalND, srandu
import time
npars=6
cen_prior = CenPrior([15.0, 16.0], [0.1, 0.2])
g_prior = GPriorBA(0.3)
print 'pre-generating g values from prior (for seed to sampler)'
ng_pre = 10000
g1vals_pre,g2vals_pre = g_prior.sample2d(ng_pre)
print 'done'
g_dist_gauss = NormalND( [0.2,0.1], [gerr,gerr] )
gg_prior = GPriorTimesGauss(g_prior, g_dist_gauss)
T_prior = LogNormal(16.0, 3.0)
counts_prior = LogNormal(100.0, 10.0)
comb=CombinedPriorSimple(cen_prior,
gg_prior,
T_prior,
counts_prior)
nwalkers=20
print 'nstep per:',nrand/nwalkers
start=numpy.zeros( (nwalkers,npars) )
minerr2=0.01**2
tm=time.time()
for i in xrange(ntry):
start[:,0:2] = cen_prior.sample(nwalkers)
# g start
g_err2 = g_dist_gauss.sigma[0]**2 + g_dist_gauss.sigma[1]**2
if g_err2 < minerr2:
print 'error small enough from max like, skipping resample'
continue
nsig_g = 3
g1_range = numpy.array([g_dist_gauss.mean[0]-nsig_g*g_dist_gauss.sigma[0],
g_dist_gauss.mean[0]+nsig_g*g_dist_gauss.sigma[0]])
g2_range = numpy.array([g_dist_gauss.mean[1]-nsig_g*g_dist_gauss.sigma[1],
g_dist_gauss.mean[1]+nsig_g*g_dist_gauss.sigma[1]])
g1_range.clip(-1.0,1.0,g1_range)
g2_range.clip(-1.0,1.0,g2_range)
g1_width = g1_range.max() - g1_range.min()
g2_width = g2_range.max() - g2_range.min()
if g1_width==0 or g2_width==0:
# crazy, probably near e==0. Just draw from prior
randi = eu.numpy_util.randind(ng_pre, nwalkers)
g1rand=g1vals_pre[randi]
g2rand=g2vals_pre[randi]
else:
w,=numpy.where( (g1vals_pre > g1_range[0])
& (g1vals_pre < g1_range[1])
& (g2vals_pre > g1_range[0])
& (g2vals_pre < g1_range[1]) )
if w.size < nwalkers:
raise ValueError("too few")
randi = eu.numpy_util.randind(w.size, nwalkers)
randi=w[randi]
g1rand=g1vals_pre[randi]
g2rand=g2vals_pre[randi]
"""
sig2=g_dist_gauss.sigma[0]**2 + g_dist_gauss.sigma[1]**2
if sig2 < 0.3**2:
g1rand = g_dist_gauss.mean[0]*(1.0 + 0.01*srandu(nwalkers))
g2rand = g_dist_gauss.mean[1]*(1.0 + 0.01*srandu(nwalkers))
else:
g1rand,g2rand=g_prior_ba.sample2d(nwalkers)
"""
start[:,2] = g1rand
start[:,3] = g2rand
start[:,4] = T_prior.sample(nwalkers)
start[:,5] = counts_prior.sample(nwalkers)
prand = comb.sample(start,
nrand,
burnin=burnin,
nwalkers=nwalkers)
print 'time per:',(time.time()-tm)/ntry
print 'g1: %g +/- %g' % (prand[:,2].mean(), prand[:,2].std())
print 'g2: %g +/- %g' % (prand[:,3].mean(), prand[:,3].std())
if show:
plot_many_hist(prand)
def _sample_dist_with_gauss(self):
from esutil.random import NormalND, LogNormal
nwalkers=20
burnin=25
pars=self._lm_result['pars']
cov=self._lm_result['pcov']*4
cen_prior = priors.CenPrior(pars[0:2], [cov[0,0], cov[1,1]])
g1_sigma = numpy.sqrt(cov[2,2])
g2_sigma = numpy.sqrt(cov[3,3])
if g1_sigma < 0.001:
g1_sigma=0.001
if g2_sigma < 0.001:
g2_sigma=0.001
g_dist_gauss = NormalND( pars[2:4], [g1_sigma,g2_sigma])
gg_prior = priors.GPriorTimesGauss(self.gprior, g_dist_gauss)
Tg=pars[4]
T_sigma = numpy.sqrt(cov[4,4])
if Tg < 0.01:
Tg=0.01
if T_sigma < 0.01:
T_sigma=0.01
counts_g=pars[5]
counts_sigma = numpy.sqrt(cov[5,5])
if counts_g < 0.01:
counts_g=0.01
if counts_sigma < 0.001:
counts_sigma=0.001
T_prior = LogNormal(Tg, T_sigma)
counts_prior = LogNormal(counts_g, counts_sigma)
comb=priors.CombinedPriorSimple(cen_prior,
gg_prior,
T_prior,
counts_prior)
g1rand,g2rand=self._get_gstart(g_dist_gauss,nwalkers)
if g1rand is None:
# we will just use the lm fit
return None,None
start=numpy.zeros( (nwalkers,self.npars) )
start[:,0:2] = cen_prior.sample(nwalkers)
start[:,2] = g1rand
start[:,3] = g2rand
start[:,4] = T_prior.sample(nwalkers)
start[:,5] = counts_prior.sample(nwalkers)
sampler = comb.sample(start,
self.nsample,
burnin=burnin,
nwalkers=nwalkers,
get_sampler=True)
prand = sampler.flatchain
lnp = sampler.lnprobability
lnp = lnp.reshape(lnp.shape[0]*lnp.shape[1])
probs = numpy.exp(lnp)
return prand,probs