Changed some parameters of the laplace, tidied up examples

This commit is contained in:
Alan Saul 2013-11-28 15:23:39 +00:00
parent 50e9034a6d
commit 0a43329150
2 changed files with 105 additions and 96 deletions

View file

@ -2,22 +2,21 @@ import GPy
import numpy as np
import matplotlib.pyplot as plt
from GPy.util import datasets
#np.random.seed(1)
def student_t_approx():
def student_t_approx(optimize=True, plot=True):
"""
Example of regressing with a student t likelihood
Example of regressing with a student t likelihood using Laplace
"""
real_std = 0.1
#Start a function, any function
X = np.linspace(0.0, np.pi*2, 100)[:, None]
Y = np.sin(X) + np.random.randn(*X.shape)*real_std
Y = Y/Y.max()
Yc = Y.copy()
X_full = np.linspace(0.0, np.pi*2, 500)[:, None]
Y_full = np.sin(X_full)
Y = Y/Y.max()
Y_full = Y_full/Y_full.max()
#Slightly noisy data
Yc[75:80] += 1
@ -34,94 +33,93 @@ def student_t_approx():
deg_free = 5
print "Real noise: ", real_std
initial_var_guess = 0.5
edited_real_sd = initial_var_guess
#t_rv = t(deg_free, loc=0, scale=real_var)
#noise = t_rvrvs(size=Y.shape)
#Y += noise
plt.figure(1)
plt.suptitle('Gaussian likelihood')
# Kernel object
kernel1 = GPy.kern.rbf(X.shape[1]) + GPy.kern.white(X.shape[1])
kernel2 = kernel1.copy()
kernel3 = kernel1.copy()
kernel4 = kernel1.copy()
kernel5 = kernel1.copy()
kernel6 = kernel1.copy()
print "Clean Gaussian"
#A GP should completely break down due to the points as they get a lot of weight
# create simple GP model
m = GPy.models.GPRegression(X, Y, kernel=kernel1)
#Gaussian GP model on clean data
m1 = GPy.models.GPRegression(X, Y.copy(), kernel=kernel1)
# optimize
m.ensure_default_constraints()
m.constrain_fixed('white', 1e-4)
m.randomize()
m.optimize()
# plot
ax = plt.subplot(211)
m.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Gaussian clean')
print m
m1.ensure_default_constraints()
m1.constrain_fixed('white', 1e-5)
m1.randomize()
#Corrupt
print "Corrupt Gaussian"
m = GPy.models.GPRegression(X, Yc, kernel=kernel2)
m.ensure_default_constraints()
m.constrain_fixed('white', 1e-4)
m.randomize()
m.optimize()
ax = plt.subplot(212)
m.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Gaussian corrupt')
print m
#Gaussian GP model on corrupt data
m2 = GPy.models.GPRegression(X, Yc.copy(), kernel=kernel2)
m2.ensure_default_constraints()
m2.constrain_fixed('white', 1e-5)
m2.randomize()
plt.figure(2)
plt.suptitle('Student-t likelihood')
edited_real_sd = initial_var_guess
print "Clean student t, rasm"
#Student t GP model on clean data
t_distribution = GPy.likelihoods.noise_model_constructors.student_t(deg_free=deg_free, sigma2=edited_real_sd)
stu_t_likelihood = GPy.likelihoods.Laplace(Y.copy(), t_distribution)
m = GPy.models.GPRegression(X, Y.copy(), kernel6, likelihood=stu_t_likelihood)
m.ensure_default_constraints()
m.constrain_positive('t_noise')
m.constrain_fixed('white', 1e-4)
m.randomize()
#m.update_likelihood_approximation()
m.optimize()
print(m)
ax = plt.subplot(211)
m.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Student-t rasm clean')
m3 = GPy.models.GPRegression(X, Y.copy(), kernel3, likelihood=stu_t_likelihood)
m3.ensure_default_constraints()
m3.constrain_bounded('t_noise', 1e-6, 10.)
m3.constrain_fixed('white', 1e-5)
m3.randomize()
print "Corrupt student t, rasm"
#Student t GP model on corrupt data
t_distribution = GPy.likelihoods.noise_model_constructors.student_t(deg_free=deg_free, sigma2=edited_real_sd)
corrupt_stu_t_likelihood = GPy.likelihoods.Laplace(Yc.copy(), t_distribution)
m = GPy.models.GPRegression(X, Yc.copy(), kernel4, likelihood=corrupt_stu_t_likelihood)
m.ensure_default_constraints()
m.constrain_bounded('t_noise', 1e-6, 10.)
m.constrain_fixed('white', 1e-4)
m.randomize()
for a in range(1):
m.randomize()
m_start = m.copy()
print m
m.optimize('scg', messages=1)
print(m)
ax = plt.subplot(212)
m.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Student-t rasm corrupt')
m4 = GPy.models.GPRegression(X, Yc.copy(), kernel4, likelihood=corrupt_stu_t_likelihood)
m4.ensure_default_constraints()
m4.constrain_bounded('t_noise', 1e-6, 10.)
m4.constrain_fixed('white', 1e-5)
m4.randomize()
return m
if optimize:
optimizer='scg'
print "Clean Gaussian"
m1.optimize(optimizer, messages=1)
print "Corrupt Gaussian"
m2.optimize(optimizer, messages=1)
print "Clean student t"
m3.optimize(optimizer, messages=1)
print "Corrupt student t"
m4.optimize(optimizer, messages=1)
if False:
print m1
print m3
plt.figure(3)
plt.scatter(X, m1.likelihood.Y, c='g')
plt.scatter(X, m3.likelihood.Y, c='r')
if plot:
plt.figure(1)
plt.suptitle('Gaussian likelihood')
ax = plt.subplot(211)
m1.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Gaussian clean')
ax = plt.subplot(212)
m2.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Gaussian corrupt')
plt.figure(2)
plt.suptitle('Student-t likelihood')
ax = plt.subplot(211)
m3.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Student-t rasm clean')
ax = plt.subplot(212)
m4.plot(ax=ax)
plt.plot(X_full, Y_full)
plt.ylim(-1.5, 1.5)
plt.title('Student-t rasm corrupt')
return m1, m2, m3, m4
def boston_example():
import sklearn
@ -294,3 +292,4 @@ def precipitation_example():
for n, (train, test) in enumerate(kf):
X_train, X_test, Y_train, Y_test = X[train], X[test], Y[train], Y[test]
print "Fold {}".format(n)

View file

@ -15,6 +15,7 @@ import scipy as sp
from likelihood import likelihood
from ..util.linalg import mdot, jitchol, pddet, dpotrs
from functools import partial as partial_func
import warnings
class Laplace(likelihood):
"""Laplace approximation to a posterior"""
@ -64,6 +65,7 @@ class Laplace(likelihood):
self.YYT = None
self.old_Ki_f = None
self.bad_fhat = False
def predictive_values(self,mu,var,full_cov,**noise_args):
if full_cov:
@ -198,18 +200,16 @@ class Laplace(likelihood):
Y_tilde = Wi*self.Ki_f + self.f_hat
self.Wi_K_i = self.W12BiW12
self.ln_det_Wi_K = pddet(self.Sigma_tilde + self.K)
self.lik = self.noise_model.logpdf(self.f_hat, self.data, extra_data=self.extra_data)
self.y_Wi_Ki_i_y = mdot(Y_tilde.T, self.Wi_K_i, Y_tilde)
ln_det_Wi_K = pddet(self.Sigma_tilde + self.K)
lik = self.noise_model.logpdf(self.f_hat, self.data, extra_data=self.extra_data)
y_Wi_K_i_y = mdot(Y_tilde.T, self.Wi_K_i, Y_tilde)
Z_tilde = (+ self.lik
Z_tilde = (+ lik
- 0.5*self.ln_B_det
+ 0.5*self.ln_det_Wi_K
+ 0.5*ln_det_Wi_K
- 0.5*self.f_Ki_f
+ 0.5*self.y_Wi_Ki_i_y
+ 0.5*y_Wi_K_i_y
)
#print "Term, {}, {}, {}, {}, {}".format(self.lik, - 0.5*self.ln_B_det, + 0.5*self.ln_det_Wi_K, - 0.5*self.f_Ki_f, + 0.5*self.y_Wi_Ki_i_y)
#Convert to float as its (1, 1) and Z must be a scalar
self.Z = np.float64(Z_tilde)
self.Y = Y_tilde
@ -247,7 +247,10 @@ class Laplace(likelihood):
#At this point get the hessian matrix (or vector as W is diagonal)
self.W = -self.noise_model.d2logpdf_df2(self.f_hat, self.data, extra_data=self.extra_data)
#TODO: Could save on computation when using rasm by returning these, means it isn't just a "mode finder" though
if not self.noise_model.log_concave:
#print "Under 1e-10: {}".format(np.sum(self.W < 1e-6))
self.W[self.W < 1e-6] = 1e-6 # FIXME-HACK: This is a hack since GPy can't handle negative variances which can occur
self.W12BiW12, self.ln_B_det = self._compute_B_statistics(self.K, self.W, np.eye(self.N))
self.Ki_f = self.Ki_f
@ -283,11 +286,11 @@ class Laplace(likelihood):
except:
import ipdb; ipdb.set_trace()
W12BiW12 = W_12*dpotrs(L, np.asfortranarray(W_12*a), lower=1)[0]
W12BiW12a = W_12*dpotrs(L, np.asfortranarray(W_12*a), lower=1)[0]
ln_B_det = 2*np.sum(np.log(np.diag(L)))
return W12BiW12, ln_B_det
return W12BiW12a, ln_B_det
def rasm_mode(self, K, MAX_ITER=30):
def rasm_mode(self, K, MAX_ITER=40):
"""
Rasmussen's numerically stable mode finding
For nomenclature see Rasmussen & Williams 2006
@ -302,9 +305,10 @@ class Laplace(likelihood):
"""
#old_Ki_f = np.zeros((self.N, 1))
#Start f's at zero originally
if self.old_Ki_f is None:
old_Ki_f = np.zeros((self.N, 1))
#Start f's at zero originally of if we have gone off track, try restarting
if self.old_Ki_f is None or self.bad_fhat:
old_Ki_f = np.random.rand(self.N, 1)/50.0
#old_Ki_f = self.Y
f = np.dot(K, old_Ki_f)
else:
#Start at the old best point
@ -318,7 +322,7 @@ class Laplace(likelihood):
return -0.5*np.dot(Ki_f.T, f) + self.noise_model.logpdf(f, self.data, extra_data=self.extra_data)
difference = np.inf
epsilon = 1e-5
epsilon = 1e-7
#step_size = 1
#rs = 0
i = 0
@ -381,14 +385,20 @@ class Laplace(likelihood):
#difference = abs(new_obj - old_obj)
#old_obj = new_obj.copy()
difference = np.abs(np.sum(f - f_old))
#difference = np.abs(np.sum(Ki_f - old_Ki_f))
difference = np.abs(np.sum(f - f_old)) + np.abs(np.sum(Ki_f - old_Ki_f))
#difference = np.abs(np.sum(Ki_f - old_Ki_f))/np.float(self.N)
old_Ki_f = Ki_f.copy()
i += 1
self.old_Ki_f = old_Ki_f.copy()
#Warn of bad fits
if difference > epsilon:
print "Not perfect f_hat fit difference: {}".format(difference)
self.bad_fhat = True
warnings.warn("Not perfect f_hat fit difference: {}".format(difference))
elif self.bad_fhat:
self.bad_fhat = False
warnings.warn("f_hat now perfect again")
self.Ki_f = Ki_f
return f