apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-18 13:55:14 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

fixed merge conflicts

Browse source
This commit is contained in:
Alan Saul 2013-03-11 12:26:44 +00:00
commit 95cedc7e4e
23 changed files with 612 additions and 209 deletions
Download patch file Download diff file Expand all files Collapse all files

64
GPy/likelihoods/EP.py
View file

@ -17,7 +17,7 @@ class EP(likelihood):
self.epsilon = epsilon
self.eta, self.delta = power_ep
self.data = data
self.N = self.data.size
self.N, self.D = self.data.shape
self.is_heteroscedastic = True
self.Nparams = 0
@ -29,7 +29,7 @@ class EP(likelihood):
#initial values for the GP variables
self.Y = np.zeros((self.N,1))
self.covariance_matrix = np.eye(self.N)
self.precision = np.ones(self.N)
self.precision = np.ones(self.N)[:,None]
self.Z = 0
self.YYT = None
@ -54,18 +54,14 @@ class EP(likelihood):
self.Y = mu_tilde[:,None]
self.YYT = np.dot(self.Y,self.Y.T)
self.precision = self.tau_tilde
self.covariance_matrix = np.diag(1./self.precision)
self.covariance_matrix = np.diag(1./self.tau_tilde)
self.precision = self.tau_tilde[:,None]
def fit_full(self,K):
"""
The expectation-propagation algorithm.
For nomenclature see Rasmussen & Williams 2006.
"""
#Prior distribution parameters: p(f|X) = N(f|0,K)
self.tau_tilde = np.zeros(self.N)
self.v_tilde = np.zeros(self.N)
#Initial values - Posterior distribution parameters: q(f|X,Y) = N(f|mu,Sigma)
mu = np.zeros(self.N)
Sigma = K.copy()
@ -124,13 +120,14 @@ class EP(likelihood):
return self._compute_GP_variables()
def fit_DTC(self, Knn_diag, Kmn, Kmm):
#def fit_DTC(self, Knn_diag, Kmn, Kmm):
def fit_DTC(self, Kmm, Kmn):
"""
The expectation-propagation algorithm with sparse pseudo-input.
For nomenclature see ... 2013.
"""
#TODO: this doesn;t work with uncertain inputs!
#TODO: this doesn't work with uncertain inputs!
"""
Prior approximation parameters:
@ -158,12 +155,12 @@ class EP(likelihood):
sigma_ = 1./tau_
mu_ = v_/tau_
"""
tau_ = np.empty(self.N,dtype=float)
v_ = np.empty(self.N,dtype=float)
self.tau_ = np.empty(self.N,dtype=float)
self.v_ = np.empty(self.N,dtype=float)
#Initial values - Marginal moments
z = np.empty(self.N,dtype=float)
Z_hat = np.empty(self.N,dtype=float)
self.Z_hat = np.empty(self.N,dtype=float)
phi = np.empty(self.N,dtype=float)
mu_hat = np.empty(self.N,dtype=float)
sigma2_hat = np.empty(self.N,dtype=float)
@ -172,21 +169,21 @@ class EP(likelihood):
epsilon_np1 = 1
epsilon_np2 = 1
self.iterations = 0
np1 = [tau_tilde.copy()]
np2 = [v_tilde.copy()]
np1 = [self.tau_tilde.copy()]
np2 = [self.v_tilde.copy()]
while epsilon_np1 > self.epsilon or epsilon_np2 > self.epsilon:
update_order = np.random.permutation(self.N)
for i in update_order:
#Cavity distribution parameters
tau_[i] = 1./Sigma_diag[i] - self.eta*tau_tilde[i]
v_[i] = mu[i]/Sigma_diag[i] - self.eta*v_tilde[i]
self.tau_[i] = 1./Sigma_diag[i] - self.eta*self.tau_tilde[i]
self.v_[i] = mu[i]/Sigma_diag[i] - self.eta*self.v_tilde[i]
#Marginal moments
Z_hat[i], mu_hat[i], sigma2_hat[i] = self.likelihood_function.moments_match(self.data[i],tau_[i],v_[i])
self.Z_hat[i], mu_hat[i], sigma2_hat[i] = self.likelihood_function.moments_match(self.data[i],self.tau_[i],self.v_[i])
#Site parameters update
Delta_tau = delta/self.eta*(1./sigma2_hat[i] - 1./Sigma_diag[i])
Delta_tau = self.delta/self.eta*(1./sigma2_hat[i] - 1./Sigma_diag[i])
Delta_v = self.delta/self.eta*(mu_hat[i]/sigma2_hat[i] - mu[i]/Sigma_diag[i])
tau_tilde[i] = tau_tilde[i] + Delta_tau
v_tilde[i] = v_tilde[i] + Delta_v
self.tau_tilde[i] = self.tau_tilde[i] + Delta_tau
self.v_tilde[i] = self.v_tilde[i] + Delta_v
#Posterior distribution parameters update
LLT = LLT + np.outer(Kmn[:,i],Kmn[:,i])*Delta_tau
L = jitchol(LLT)
@ -196,25 +193,26 @@ class EP(likelihood):
mu = mu + (Delta_v-Delta_tau*mu[i])*si
self.iterations += 1
#Sigma recomputation with Cholesky decompositon
LLT0 = LLT0 + np.dot(Kmn*tau_tilde[None,:],Kmn.T)
LLT0 = LLT0 + np.dot(Kmn*self.tau_tilde[None,:],Kmn.T)
L = jitchol(LLT)
V,info = linalg.lapack.flapack.dtrtrs(L,Kmn,lower=1)
V2,info = linalg.lapack.flapack.dtrtrs(L.T,V,lower=0)
Sigma_diag = np.sum(V*V,-2)
Knmv_tilde = np.dot(Kmn,v_tilde)
Knmv_tilde = np.dot(Kmn,self.v_tilde)
mu = np.dot(V2.T,Knmv_tilde)
epsilon_np1 = sum((tau_tilde-np1[-1])**2)/self.N
epsilon_np2 = sum((v_tilde-np2[-1])**2)/self.N
np1.append(tau_tilde.copy())
np2.append(v_tilde.copy())
epsilon_np1 = sum((self.tau_tilde-np1[-1])**2)/self.N
epsilon_np2 = sum((self.v_tilde-np2[-1])**2)/self.N
np1.append(self.tau_tilde.copy())
np2.append(self.v_tilde.copy())
self._compute_GP_variables()
def fit_FITC(self, Knn_diag, Kmn):
def fit_FITC(self, Kmm, Kmn, Knn_diag):
"""
The expectation-propagation algorithm with sparse pseudo-input.
For nomenclature see Naish-Guzman and Holden, 2008.
"""
M = Kmm.shape[0]
"""
Prior approximation parameters:
@ -235,7 +233,7 @@ class EP(likelihood):
mu = w + P*gamma
"""
self.w = np.zeros(self.N)
self.gamma = np.zeros(self.M)
self.gamma = np.zeros(M)
mu = np.zeros(self.N)
P = P0.copy()
R = R0.copy()
@ -271,7 +269,7 @@ class EP(likelihood):
self.tau_[i] = 1./Sigma_diag[i] - self.eta*self.tau_tilde[i]
self.v_[i] = mu[i]/Sigma_diag[i] - self.eta*self.v_tilde[i]
#Marginal moments
self.Z_hat[i], mu_hat[i], sigma2_hat[i] = self.likelihood_function.moments_match(data[i],self.tau_[i],self.v_[i])
self.Z_hat[i], mu_hat[i], sigma2_hat[i] = self.likelihood_function.moments_match(self.data[i],self.tau_[i],self.v_[i])
#Site parameters update
Delta_tau = self.delta/self.eta*(1./sigma2_hat[i] - 1./Sigma_diag[i])
Delta_v = self.delta/self.eta*(mu_hat[i]/sigma2_hat[i] - mu[i]/Sigma_diag[i])
@ -281,10 +279,10 @@ class EP(likelihood):
dtd1 = Delta_tau*Diag[i] + 1.
dii = Diag[i]
Diag[i] = dii - (Delta_tau * dii**2.)/dtd1
pi_ = P[i,:].reshape(1,self.M)
pi_ = P[i,:].reshape(1,M)
P[i,:] = pi_ - (Delta_tau*dii)/dtd1 * pi_
Rp_i = np.dot(R,pi_.T)
RTR = np.dot(R.T,np.dot(np.eye(self.M) - Delta_tau/(1.+Delta_tau*Sigma_diag[i]) * np.dot(Rp_i,Rp_i.T),R))
RTR = np.dot(R.T,np.dot(np.eye(M) - Delta_tau/(1.+Delta_tau*Sigma_diag[i]) * np.dot(Rp_i,Rp_i.T),R))
R = jitchol(RTR).T
self.w[i] = self.w[i] + (Delta_v - Delta_tau*self.w[i])*dii/dtd1
self.gamma = self.gamma + (Delta_v - Delta_tau*mu[i])*np.dot(RTR,P[i,:].T)
@ -296,7 +294,7 @@ class EP(likelihood):
Diag = Diag0/(1.+ Diag0 * self.tau_tilde)
P = (Diag / Diag0)[:,None] * P0
RPT0 = np.dot(R0,P0.T)
L = jitchol(np.eye(self.M) + np.dot(RPT0,(1./Diag0 - Diag/(Diag0**2))[:,None]*RPT0.T))
L = jitchol(np.eye(M) + np.dot(RPT0,(1./Diag0 - Diag/(Diag0**2))[:,None]*RPT0.T))
R,info = linalg.lapack.flapack.dtrtrs(L,R0,lower=1)
RPT = np.dot(R,P.T)
Sigma_diag = Diag + np.sum(RPT.T*RPT.T,-1)

4
GPy/likelihoods/likelihood_functions.py
View file

@ -37,8 +37,8 @@ class probit(likelihood_function):
:param tau_i: precision of the cavity distribution (float)
:param v_i: mean/variance of the cavity distribution (float)
"""
# TODO: some version of assert np.sum(np.abs(Y)-1) == 0, "Output values must be either -1 or 1"
if data_i == 0: data_i = -1 #NOTE Binary classification works better classes {-1,1}, 1D-plotting works better with classes {0,1}.
if data_i == 0: data_i = -1 #NOTE Binary classification algorithm works better with classes {-1,1}, 1D-plotting works better with classes {0,1}.
# TODO: some version of assert
z = data_i*v_i/np.sqrt(tau_i**2 + tau_i)
Z_hat = stats.norm.cdf(z)
phi = stats.norm.pdf(z)