apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-10 20:42:39 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

commiting code for some changes to api for calculating ep_gradients, also fixing some issues with gaussian hermite quadrature, no we have both avaialable ...

Browse source
This commit is contained in:
Akash Kumar Dhaka 2017-07-03 18:24:02 +03:00
parent ae27e1225d
commit 8b621a409c
3 changed files with 37 additions and 54 deletions
Download patch file Download diff file Expand all files Collapse all files

15
GPy/inference/latent_function_inference/expectation_propagation.py
View file

@ -179,14 +179,14 @@ class EP(EPBase, ExactGaussianInference):
elif self.ep_mode=="alternated":
if getattr(self, '_ep_approximation', None) is None:
#if we don't yet have the results of runnign EP, run EP and store the computed factors in self._ep_approximation
post_params, ga_approx, log_Z_tilde = self._ep_approximation = self.expectation_propagation(K, Y, likelihood, Y_metadata)
post_params, ga_approx, cav_params, log_Z_tilde = self._ep_approximation = self.expectation_propagation(K, Y, likelihood, Y_metadata)
else:
#if we've already run EP, just use the existing approximation stored in self._ep_approximation
post_params, ga_approx, log_Z_tilde = self._ep_approximation
post_params, ga_approx, cav_params, log_Z_tilde = self._ep_approximation
else:
raise ValueError("ep_mode value not valid")
return self._inference(Y, K, ga_approx, likelihood, Y_metadata=Y_metadata, Z_tilde=log_Z_tilde)
return self._inference(Y, K, ga_approx, cav_params, likelihood, Y_metadata=Y_metadata, Z_tilde=log_Z_tilde)
def expectation_propagation(self, K, Y, likelihood, Y_metadata):
@ -221,7 +221,7 @@ class EP(EPBase, ExactGaussianInference):
# This terms cancel with the coreresponding terms in the marginal loglikelihood
log_Z_tilde = self._log_Z_tilde(marg_moments, ga_approx, cav_params)
# - 0.5*np.log(tau_tilde) + 0.5*(v_tilde*v_tilde*1./tau_tilde)
return (post_params, ga_approx, log_Z_tilde)
return (post_params, ga_approx, cav_params, log_Z_tilde)
def _init_approximations(self, K, num_data):
#initial values - Gaussian factors
@ -281,7 +281,7 @@ class EP(EPBase, ExactGaussianInference):
return log_marginal, post_params
def _inference(self, Y, K, ga_approx, likelihood, Z_tilde, Y_metadata=None):
def _inference(self, Y, K, ga_approx, cav_params, likelihood, Z_tilde, Y_metadata=None):
log_marginal, post_params = self._ep_marginal(K, ga_approx, Z_tilde)
tau_tilde_root = np.sqrt(ga_approx.tau)
@ -294,10 +294,7 @@ class EP(EPBase, ExactGaussianInference):
symmetrify(Wi) #(K + Sigma^(\tilde))^(-1)
dL_dK = 0.5 * (tdot(alpha) - Wi)
if not isinstance(likelihood, Gaussian):
dL_dthetaL = likelihood.ep_gradients(Y, ga_approx.tau, ga_approx.v, Y_metadata=Y_metadata)
else:
dL_dthetaL = likelihood.exact_inference_gradients(np.diag(dL_dK), Y_metadata)
dL_dthetaL = likelihood.ep_gradients(Y, cav_params.tau, cav_params.v, np.diag(dL_dK), Y_metadata=Y_metadata, quad_mode='gh')
return Posterior(woodbury_inv=Wi, woodbury_vector=alpha, K=K), log_marginal, {'dL_dK':dL_dK, 'dL_dthetaL':dL_dthetaL, 'dL_dm':alpha}