apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-15 06:52:39 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'params' of github.com:SheffieldML/GPy into params

Browse source 
Conflicts:
	GPy/likelihoods/mixed_noise.py
This commit is contained in:
Ricardo 2014-03-19 09:39:35 +00:00
commit fcb6742b60
21 changed files with 197 additions and 228 deletions
Download patch file Download diff file Expand all files Collapse all files

2
GPy/inference/latent_function_inference/exact_gaussian_inference.py
View file

@ -43,7 +43,7 @@ class ExactGaussianInference(object):
K = kern.K(X)
Ky = K.copy()
diag.add(Ky, likelihood.gaussian_variance(Y, Y_metadata))
diag.add(Ky, likelihood.gaussian_variance(Y_metadata))
Wi, LW, LWi, W_logdet = pdinv(Ky)
alpha, _ = dpotrs(LW, YYT_factor, lower=1)

23
GPy/inference/latent_function_inference/laplace.py
View file

@ -51,12 +51,11 @@ class Laplace(object):
Ki_f_init = self._previous_Ki_fhat
f_hat, Ki_fhat = self.rasm_mode(K, Y, likelihood, Ki_f_init, Y_metadata=Y_metadata)
self.f_hat = f_hat
self.Ki_fhat = Ki_fhat
self.K = K.copy()
#Compute hessian and other variables at mode
log_marginal, woodbury_vector, woodbury_inv, dL_dK, dL_dthetaL = self.mode_computations(f_hat, Ki_fhat, K, Y, likelihood, kern, Y_metadata)
log_marginal, woodbury_inv, dL_dK, dL_dthetaL = self.mode_computations(f_hat, Ki_fhat, K, Y, likelihood, kern, Y_metadata)
self._previous_Ki_fhat = Ki_fhat.copy()
return Posterior(woodbury_vector=Ki_fhat, woodbury_inv=woodbury_inv, K=K), log_marginal, {'dL_dK':dL_dK, 'dL_dthetaL':dL_dthetaL}
@ -86,13 +85,13 @@ class Laplace(object):
#define the objective function (to be maximised)
def obj(Ki_f, f):
return -0.5*np.dot(Ki_f.flatten(), f.flatten()) + likelihood.logpdf(f, Y, extra_data=Y_metadata)
return -0.5*np.dot(Ki_f.flatten(), f.flatten()) + likelihood.logpdf(f, Y, Y_metadata=Y_metadata)
difference = np.inf
iteration = 0
while difference > self._mode_finding_tolerance and iteration < self._mode_finding_max_iter:
W = -likelihood.d2logpdf_df2(f, Y, extra_data=Y_metadata)
grad = likelihood.dlogpdf_df(f, Y, extra_data=Y_metadata)
W = -likelihood.d2logpdf_df2(f, Y, Y_metadata=Y_metadata)
grad = likelihood.dlogpdf_df(f, Y, Y_metadata=Y_metadata)
W_f = W*f
@ -136,13 +135,12 @@ class Laplace(object):
At the mode, compute the hessian and effective covariance matrix.
returns: logZ : approximation to the marginal likelihood
woodbury_vector : variable required for calculating the approximation to the covariance matrix
woodbury_inv : variable required for calculating the approximation to the covariance matrix
dL_dthetaL : array of derivatives (1 x num_kernel_params)
dL_dthetaL : array of derivatives (1 x num_likelihood_params)
"""
#At this point get the hessian matrix (or vector as W is diagonal)
W = -likelihood.d2logpdf_df2(f_hat, Y, extra_data=Y_metadata)
W = -likelihood.d2logpdf_df2(f_hat, Y, Y_metadata=Y_metadata)
K_Wi_i, L, LiW12 = self._compute_B_statistics(K, W, likelihood.log_concave)
@ -151,11 +149,10 @@ class Laplace(object):
Ki_W_i = K - C.T.dot(C) #Could this be wrong?
#compute the log marginal
log_marginal = -0.5*np.dot(Ki_f.flatten(), f_hat.flatten()) + likelihood.logpdf(f_hat, Y, extra_data=Y_metadata) - np.sum(np.log(np.diag(L)))
log_marginal = -0.5*np.dot(Ki_f.flatten(), f_hat.flatten()) + likelihood.logpdf(f_hat, Y, Y_metadata=Y_metadata) - np.sum(np.log(np.diag(L)))
#Compute vival matrices for derivatives
dW_df = -likelihood.d3logpdf_df3(f_hat, Y, extra_data=Y_metadata) # -d3lik_d3fhat
woodbury_vector = likelihood.dlogpdf_df(f_hat, Y, extra_data=Y_metadata)
dW_df = -likelihood.d3logpdf_df3(f_hat, Y, Y_metadata=Y_metadata) # -d3lik_d3fhat
dL_dfhat = -0.5*(np.diag(Ki_W_i)[:, None]*dW_df) #why isn't this -0.5? s2 in R&W p126 line 9.
#BiK, _ = dpotrs(L, K, lower=1)
#dL_dfhat = 0.5*np.diag(BiK)[:, None]*dW_df
@ -169,7 +166,7 @@ class Laplace(object):
explicit_part = 0.5*(np.dot(Ki_f, Ki_f.T) - K_Wi_i)
#Implicit
implicit_part = np.dot(woodbury_vector, dL_dfhat.T).dot(I_KW_i)
implicit_part = np.dot(Ki_f, dL_dfhat.T).dot(I_KW_i)
dL_dK = explicit_part + implicit_part
else:
@ -179,7 +176,7 @@ class Laplace(object):
#compute dL_dthetaL#
####################
if likelihood.size > 0 and not likelihood.is_fixed:
dlik_dthetaL, dlik_grad_dthetaL, dlik_hess_dthetaL = likelihood._laplace_gradients(f_hat, Y, extra_data=Y_metadata)
dlik_dthetaL, dlik_grad_dthetaL, dlik_hess_dthetaL = likelihood._laplace_gradients(f_hat, Y, Y_metadata=Y_metadata)
num_params = likelihood.size
# make space for one derivative for each likelihood parameter
@ -200,7 +197,7 @@ class Laplace(object):
else:
dL_dthetaL = np.zeros(likelihood.size)
return log_marginal, woodbury_vector, K_Wi_i, dL_dK, dL_dthetaL
return log_marginal, K_Wi_i, dL_dK, dL_dthetaL
def _compute_B_statistics(self, K, W, log_concave):
"""

5
GPy/inference/latent_function_inference/var_dtc.py
View file

@ -65,7 +65,7 @@ class VarDTC(object):
_, output_dim = Y.shape
#see whether we've got a different noise variance for each datum
beta = 1./np.fmax(likelihood.gaussian_variance(Y, Y_metadata), 1e-6)
beta = 1./np.fmax(likelihood.gaussian_variance(Y_metadata), 1e-6)
# VVT_factor is a matrix such that tdot(VVT_factor) = VVT...this is for efficiency!
#self.YYTfactor = self.get_YYTfactor(Y)
#VVT_factor = self.get_VVTfactor(self.YYTfactor, beta)
@ -176,7 +176,6 @@ class VarDTC(object):
#construct a posterior object
post = Posterior(woodbury_inv=woodbury_inv, woodbury_vector=woodbury_vector, K=Kmm, mean=None, cov=None, K_chol=Lm)
return post, log_marginal, grad_dict
class VarDTCMissingData(object):
@ -365,7 +364,7 @@ class VarDTCMissingData(object):
return post, log_marginal, grad_dict
def _compute_dL_dpsi(num_inducing, num_data, output_dim, beta, Lm, VVT_factor, Cpsi1Vf, DBi_plus_BiPBi, psi1, het_noise, uncertain_inputs):
dL_dpsi0 = -0.5 * output_dim * (beta * np.ones([num_data, 1])).flatten()
dL_dpsi0 = -0.5 * output_dim * (beta[:,None] * np.ones([num_data, 1])).flatten()
dL_dpsi1 = np.dot(VVT_factor, Cpsi1Vf.T)
dL_dpsi2_beta = 0.5 * backsub_both_sides(Lm, output_dim * np.eye(num_inducing) - DBi_plus_BiPBi)
if het_noise: