def example_imm_1d():
n = 100
dim = 1
alpha = .5
x = np.random.randn(n, dim)
x[:.3*n] *= 2
x[:.1*n] += 3
igmm = IMM(alpha, dim)
igmm.set_priors(x)
# warming
igmm.sample(x, niter=100, init=True)
print 'number of components: ', igmm.k
from gmm import GridDescriptor
gd = GridDescriptor(1, [-9,11], 201)
#sampling
like = igmm.sample(x, niter=1000, sampling_points=gd.make_grid())
print 'number of components: ', igmm.k
print 'density sum', 0.1*like.sum()
igmm.show(x, gd, density=like)
return igmm
def example_igmm_wnc():
"""
Example of IMM with a null class
points are generated in [0,1] interval witha higher density close to 0
the null points are assumed to occur uniformly over [0,1]
the algorithm correctly identifies the points close to 0
as the 'active' set
"""
n = 50
dim = 1
alpha = .5
g0 = 1.
x = np.random.rand(n, dim)
x[:.3*n] *= .2
x[:.1*n] *= .3
migmm = MixedIMM(alpha, dim)
migmm.set_priors(x)
migmm.set_constant_densities(null_dens=g0)
# burn-in
ncp = 0.5*np.ones(n)
migmm.sample(x, null_class_proba=ncp, niter=100, init=True)
print 'number of components: ', migmm.k
from gmm import GridDescriptor
gd = GridDescriptor(1, [0,1], 101)
#sampling
like, pproba = migmm.sample(x, null_class_proba=ncp, niter=1000,
sampling_points=gd.make_grid(), kfold=10)
print 'number of components: ', migmm.k
# the density should sum to 1
print 'density sum', 0.01*like.sum()
migmm.show(x, gd, density=like)
import pylab
pylab.figure()
pylab.plot(x, pproba, '.')
pylab.title('posterior probability of being in the null class')
pylab.xlabel('data')
pylab.ylabel('proba of being in the null class')
pylab.show()
return migmm
