apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-21 14:05:14 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Student t likelihood function checkgrads (summed gradients wrt to sigma2), maybe some numerical instability in laplace

Browse source
This commit is contained in:
Alan Saul 2013-09-13 18:01:41 +01:00
parent b663fff622
commit 5e88a885b1
3 changed files with 34 additions and 17 deletions
Download patch file Download diff file Expand all files Collapse all files

5
GPy/likelihoods/laplace.py
View file

@ -127,7 +127,6 @@ class Laplace(likelihood):
#- 0.5*np.trace(mdot(self.Ki_W_i, (self.K, np.diagflat(dlik_hess_dthetaL[thetaL_i]))))
+ np.dot(0.5*np.diag(self.Ki_W_i)[:,None].T, dlik_hess_dthetaL[thetaL_i])
)
import ipdb; ipdb.set_trace() # XXX BREAKPOINT
#Implicit
df_hat_dthetaL = mdot(I_KW_i, self.K, dlik_grad_dthetaL[thetaL_i])
@ -203,7 +202,7 @@ class Laplace(likelihood):
#self.cC = 0.5*self.y_Wi_Ki_i_y
#self.dD = -0.5*self.ln_B_det
#print "Ztilde: {} lik: {} a: {} b: {} c: {} d:".format(Z_tilde, self.lik, self.aA, self.bB, self.cC, self.dD)
print "param value: {}".format(self.likelihood_function._get_params())
#print "param value: {}".format(self.likelihood_function._get_params())
#Convert to float as its (1, 1) and Z must be a scalar
self.Z = np.float64(Z_tilde)
@ -330,7 +329,6 @@ class Laplace(likelihood):
self.old_before_s = self.likelihood_function._get_params()
#print "before: ", self.old_before_s
#if self.old_before_s < 1e-5:
#import ipdb; ipdb.set_trace() ### XXX BREAKPOINT
#old_a = np.zeros((self.N, 1))
if self.old_a is None:
@ -384,7 +382,6 @@ class Laplace(likelihood):
new_obj = sp.optimize.minimize_scalar(i_o, method='brent', tol=1e-4, options={'maxiter':20, 'disp':True}).fun
f = self.f.copy()
a = self.a.copy()
#import ipdb; ipdb.set_trace() ### XXX BREAKPOINT
#f_old = f.copy()
#update_passed = False

18
GPy/likelihoods/likelihood_functions.py
View file

@ -218,16 +218,11 @@ class StudentT(LikelihoodFunction):
"""
assert y.shape == f.shape
e = y - f
#A = gammaln((self.v + 1) * 0.5)
#B = - gammaln(self.v * 0.5)
#C = - 0.5*np.log(self.sigma2 * self.v * np.pi)
#D = + (-(self.v + 1)*0.5)*np.log(1 + ((e**2)/self.sigma2)/np.float(self.v))
objective = (+ gammaln((self.v + 1) * 0.5)
- gammaln(self.v * 0.5)
- 0.5*np.log(self.sigma2 * self.v * np.pi)
+ (-(self.v + 1)*0.5)*np.log(1 + ((e**2)/self.sigma2)/np.float(self.v))
- 0.5*(self.v + 1)*np.log(1 + (1/np.float(self.v))*((e**2)/self.sigma2))
)
#print "C: {} D: {} obj: {}".format(C, np.sum(D), objective.sum())
return np.sum(objective)
def dlik_df(self, y, f, extra_data=None):
@ -291,9 +286,13 @@ class StudentT(LikelihoodFunction):
"""
assert y.shape == f.shape
e = y - f
#FIXME: OUT BY SOME FUNCTION OF N
#FIXME: OUT BY SOME FUNCTION OF N, or the fact that we are summing over several things in the objective?
dlik_dvar = self.v*(e**2 - self.sigma2)/(2*self.sigma2*(self.sigma2*self.v + e**2))
return dlik_dvar
#dlik_dvar = ( 0.5*(1/float(self.sigma2))
#-0.5*(self.v + 1)*(-(1/float(self.v))*(e**2)/(1/(float(self.sigma2**2))))
#/ (1 + (1/float(self.v))*((e**2)/float(self.sigma2)))
#)
return np.sum(dlik_dvar) #May not want to sum over all dimensions if using many D?
def dlik_df_dvar(self, y, f, extra_data=None):
"""
@ -516,8 +515,7 @@ class Gaussian(LikelihoodFunction):
e = y - f
s_4 = 1.0/(self._variance**2)
dlik_dsigma = -0.5*self.N/self._variance + 0.5*s_4*np.dot(e.T, e)
#dlik_dsigma = -0.5*self.N + 0.5*s_4*np.dot(e.T, e)
return dlik_dsigma
return np.sum(dlik_dsigma) # Sure about this sum?
def dlik_df_dvar(self, y, f, extra_data=None):
"""