apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-04-27 05:46:24 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Got the mode finding without computing Ki

Browse source
This commit is contained in:
Alan Saul 2013-04-05 17:34:11 +01:00
parent 2006a94caa
commit 4a14a82dfb
2 changed files with 152 additions and 63 deletions
Download patch file Download diff file Expand all files Collapse all files

85
python/examples/laplace_approximations.py
View file

@ -6,6 +6,38 @@ from coxGP.python.likelihoods.Laplace import Laplace
from coxGP.python.likelihoods.likelihood_function import student_t
def timing():
real_var = 0.1
times = 1000
deg_free = 10
real_sd = np.sqrt(real_var)
the_is = np.zeros(times)
X = np.linspace(0.0, 10.0, 30)[:, None]
for a in xrange(times):
Y = np.sin(X) + np.random.randn(*X.shape)*real_var
Yc = Y.copy()
Yc[10] += 100
Yc[25] += 10
Yc[23] += 10
Yc[24] += 10
edited_real_sd = real_sd
kernel1 = GPy.kern.rbf(X.shape[1])
t_distribution = student_t(deg_free, sigma=edited_real_sd)
corrupt_stu_t_likelihood = Laplace(Yc.copy(), t_distribution, rasm=True)
m = GPy.models.GP(X, corrupt_stu_t_likelihood, kernel1)
m.ensure_default_constraints()
m.update_likelihood_approximation()
m.optimize()
the_is[a] = m.likelihood.i
print the_is
print np.mean(the_is)
import ipdb; ipdb.set_trace() ### XXX BREAKPOINT
def student_t_approx():
"""
Example of regressing with a student t likelihood
@ -80,32 +112,6 @@ def student_t_approx():
plt.suptitle('Student-t likelihood')
edited_real_sd = real_sd
print "Clean student t, ncg"
t_distribution = student_t(deg_free, sigma=edited_real_sd)
stu_t_likelihood = Laplace(Y, t_distribution, rasm=False)
m = GPy.models.GP(X, stu_t_likelihood, kernel3)
m.ensure_default_constraints()
m.update_likelihood_approximation()
m.optimize()
print(m)
plt.subplot(221)
m.plot()
plt.plot(X_full, Y_full)
plt.ylim(-2.5, 2.5)
print "Corrupt student t, ncg"
t_distribution = student_t(deg_free, sigma=edited_real_sd)
corrupt_stu_t_likelihood = Laplace(Yc.copy(), t_distribution, rasm=False)
m = GPy.models.GP(X, corrupt_stu_t_likelihood, kernel5)
m.ensure_default_constraints()
m.update_likelihood_approximation()
m.optimize()
print(m)
plt.subplot(223)
m.plot()
plt.plot(X_full, Y_full)
plt.ylim(-2.5, 2.5)
print "Clean student t, rasm"
t_distribution = student_t(deg_free, sigma=edited_real_sd)
stu_t_likelihood = Laplace(Y.copy(), t_distribution, rasm=True)
@ -133,6 +139,33 @@ def student_t_approx():
plt.ylim(-2.5, 2.5)
import ipdb; ipdb.set_trace() ### XXX BREAKPOINT
print "Clean student t, ncg"
t_distribution = student_t(deg_free, sigma=edited_real_sd)
stu_t_likelihood = Laplace(Y, t_distribution, rasm=False)
m = GPy.models.GP(X, stu_t_likelihood, kernel3)
m.ensure_default_constraints()
m.update_likelihood_approximation()
m.optimize()
print(m)
plt.subplot(221)
m.plot()
plt.plot(X_full, Y_full)
plt.ylim(-2.5, 2.5)
print "Corrupt student t, ncg"
t_distribution = student_t(deg_free, sigma=edited_real_sd)
corrupt_stu_t_likelihood = Laplace(Yc.copy(), t_distribution, rasm=False)
m = GPy.models.GP(X, corrupt_stu_t_likelihood, kernel5)
m.ensure_default_constraints()
m.update_likelihood_approximation()
m.optimize()
print(m)
plt.subplot(223)
m.plot()
plt.plot(X_full, Y_full)
plt.ylim(-2.5, 2.5)
###with a student t distribution, since it has heavy tails it should work well
###likelihood_function = student_t(deg_free, sigma=real_var)
###lap = Laplace(Y, likelihood_function)