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

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

Browse source 
Conflicts:
	GPy/core/parameterization/param.py
	GPy/core/parameterization/parameter_core.py
	GPy/core/parameterization/parameterized.py
This commit is contained in:
Max Zwiessele 2014-02-10 15:21:09 +00:00
commit 6a068775f5
10 changed files with 404 additions and 301 deletions
Download patch file Download diff file Expand all files Collapse all files

102
GPy/inference/latent_function_inference/laplace.py
View file

@ -32,7 +32,7 @@ class LaplaceInference(object):
self._mode_finding_tolerance = 1e-7
self._mode_finding_max_iter = 40
self.bad_fhat = True
self._previous_Ki_fhat = None
def inference(self, kern, X, likelihood, Y, Y_metadata=None):
"""
@ -50,16 +50,17 @@ class LaplaceInference(object):
Ki_f_init = np.zeros_like(Y)
else:
Ki_f_init = self._previous_Ki_fhat
f_hat, Ki_fhat = self.rasm_mode(K, Y, likelihood, Ki_f_init, Y_metadata=Y_metadata)
#Compute hessian and other variables at mode
log_marginal, Ki_W_i, K_Wi_i, dL_dK, woodbury_vector = self.mode_computations(f_hat, Ki_fhat, K, Y, likelihood, Y_metadata)
log_marginal, woodbury_vector, woodbury_inv, dL_dK, dL_dthetaL = self.mode_computations(f_hat, Ki_fhat, K, Y, likelihood, kern, Y_metadata)
#likelihood.gradient = self.likelihood_gradients()
kern.update_gradients_full(dL_dK, X)
likelihood.update_gradients(dL_dthetaL)
self._previous_Ki_fhat = Ki_fhat.copy()
return Posterior(woodbury_vector=woodbury_vector, woodbury_inv = K_Wi_i, K=K), log_marginal, {'dL_dK':dL_dK}
return Posterior(woodbury_vector=woodbury_vector, woodbury_inv=woodbury_inv, K=K), log_marginal, {'dL_dK':dL_dK}
def rasm_mode(self, K, Y, likelihood, Ki_f_init, Y_metadata=None):
"""
@ -133,13 +134,15 @@ class LaplaceInference(object):
return f, Ki_f
def mode_computations(self, f_hat, Ki_f, K, Y, likelihood, Y_metadata):
def mode_computations(self, f_hat, Ki_f, K, Y, likelihood, kern, Y_metadata):
"""
At the mode, compute the hessian and effective covariance matrix.
returns: logZ : approximation to the marginal likelihood
Cov : the approximation to the covariance matrix
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)
@ -153,45 +156,62 @@ class LaplaceInference(object):
#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)))
#compute dL_dK
explicit_part = 0.5*(np.dot(Ki_f, Ki_f.T) - K_Wi_i)
#Implicit
d3lik_d3fhat = likelihood.d3logpdf_df3(f_hat, Y, extra_data=Y_metadata)
dL_dfhat = 0.5*(np.diag(Ki_W_i)[:, None]*d3lik_d3fhat) #why isn't this -0.5? s2 in R&W p126 line 9.
#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)
implicit_part = np.dot(woodbury_vector, dL_dfhat.T).dot(np.eye(Y.shape[0]) - np.dot(K, K_Wi_i))
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
I_KW_i = np.eye(Y.shape[0]) - np.dot(K, K_Wi_i)
dL_dK = explicit_part + implicit_part
return log_marginal, Ki_W_i, K_Wi_i, dL_dK, woodbury_vector
def likelihood_gradients(self):
"""
Gradients with respect to likelihood parameters (dL_dthetaL)
:rtype: array of derivatives (1 x num_likelihood_params)
"""
dL_dfhat, I_KW_i = self._shared_gradients_components()
dlik_dthetaL, dlik_grad_dthetaL, dlik_hess_dthetaL = likelihood._laplace_gradients(self.f_hat, self.data, extra_data=self.extra_data)
num_params = len(self._get_param_names())
# make space for one derivative for each likelihood parameter
dL_dthetaL = np.zeros(num_params)
for thetaL_i in range(num_params):
####################
#compute dL_dK#
####################
if kern.size > 0 and not kern.is_fixed:
#Explicit
dL_dthetaL_exp = ( np.sum(dlik_dthetaL[:, thetaL_i])
#- 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])
)
explicit_part = 0.5*(np.dot(Ki_f, Ki_f.T) - K_Wi_i)
#Implicit
dfhat_dthetaL = mdot(I_KW_i, self.K, dlik_grad_dthetaL[:, thetaL_i])
dL_dthetaL_imp = np.dot(dL_dfhat, dfhat_dthetaL)
dL_dthetaL[thetaL_i] = dL_dthetaL_exp + dL_dthetaL_imp
implicit_part = np.dot(woodbury_vector, dL_dfhat.T).dot(I_KW_i)
return dL_dthetaL
dL_dK = explicit_part + implicit_part
else:
dL_dK = np.zeros(likelihood.size)
####################
#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)
num_params = likelihood.size
# make space for one derivative for each likelihood parameter
dL_dthetaL = np.zeros(num_params)
for thetaL_i in range(num_params):
#Explicit
dL_dthetaL_exp = ( np.sum(dlik_dthetaL[thetaL_i])
# The + comes from the fact that dlik_hess_dthetaL == -dW_dthetaL
+ 0.5*np.sum(np.diag(Ki_W_i).flatten()*dlik_hess_dthetaL[:, thetaL_i].flatten())
)
#Implicit
dfhat_dthetaL = mdot(I_KW_i, K, dlik_grad_dthetaL[:, thetaL_i])
#dfhat_dthetaL = mdot(Ki_W_i, dlik_grad_dthetaL[:, thetaL_i])
dL_dthetaL_imp = np.dot(dL_dfhat.T, dfhat_dthetaL)
dL_dthetaL[thetaL_i] = dL_dthetaL_exp + dL_dthetaL_imp
else:
dL_dthetaL = np.zeros(likelihood.size)
return log_marginal, woodbury_vector, K_Wi_i, dL_dK, dL_dthetaL
#def likelihood_gradients(self, f_hat, K, Y, Ki_W_i, dL_dfhat, I_KW_i, likelihood, Y_metadata):
#"""
#Gradients with respect to likelihood parameters (dL_dthetaL)
#:rtype: array of derivatives (1 x num_likelihood_params)
#"""
def _compute_B_statistics(self, K, W, log_concave):
"""
@ -219,7 +239,7 @@ class LaplaceInference(object):
LiW12, _ = dtrtrs(L, np.diagflat(W_12), lower=1, trans=0)
K_Wi_i = np.dot(LiW12.T, LiW12) # R = W12BiW12, in R&W p 126, eq 5.25
#here's a better way to compute the required matrix.
#here's a better way to compute the required matrix.
# you could do the model finding witha backsub, instead of a dot...
#L2 = L/W_12
#K_Wi_i_2 , _= dpotri(L2)