apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-24 14:15:14 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

[parallel vardtc] minor adjustments to work with current implementation of

Browse source 
psi stats
This commit is contained in:
mzwiessele 2014-06-18 08:44:29 -07:00
parent c62dd85418
commit 1e006f63b5
1 changed files with 84 additions and 67 deletions
Download patch file Download diff file Expand all files Collapse all files

49
GPy/inference/latent_function_inference/var_dtc_parallel.py
View file

@ -112,12 +112,12 @@ class VarDTC_minibatch(LatentFunctionInference):
if het_noise: if het_noise:
psi2_full += beta_slice*psi2 psi2_full += beta_slice*psi2
else: else:
psi2_full += psi2 psi2_full += psi2.sum(0)
else: else:
if het_noise: if het_noise:
psi2_full += beta_slice*np.outer(psi1,psi1) psi2_full += beta_slice*np.outer(psi1,psi1)
else: else:
psi2_full += np.outer(psi1,psi1) psi2_full += np.einsum('nm,jk->mk',psi1,psi1)
if not het_noise: if not het_noise:
psi0_full *= beta psi0_full *= beta
@ -128,7 +128,7 @@ class VarDTC_minibatch(LatentFunctionInference):
#====================================================================== #======================================================================
# Compute Common Components # Compute Common Components
#====================================================================== #======================================================================
self.psi1Y = psi1Y_full
Kmm = kern.K(Z).copy() Kmm = kern.K(Z).copy()
diag.add(Kmm, self.const_jitter) diag.add(Kmm, self.const_jitter)
Lm = jitchol(Kmm) Lm = jitchol(Kmm)
@ -159,7 +159,10 @@ class VarDTC_minibatch(LatentFunctionInference):
logL_R = -np.log(beta).sum() logL_R = -np.log(beta).sum()
else: else:
logL_R = -num_data*np.log(beta) logL_R = -num_data*np.log(beta)
logL = -(output_dim*(num_data*log_2_pi+logL_R+psi0_full-np.trace(LmInvPsi2LmInvT))+YRY_full-bbt)/2.-output_dim*(-np.log(np.diag(Lm)).sum()+np.log(np.diag(LL)).sum()) logL = (
-(output_dim*(num_data*log_2_pi+logL_R+psi0_full-np.trace(LmInvPsi2LmInvT))+YRY_full-bbt)/2.
-output_dim*(-np.log(np.diag(Lm)).sum()+np.log(np.diag(LL)).sum())
)
#====================================================================== #======================================================================
# Compute dL_dKmm # Compute dL_dKmm
@ -256,14 +259,14 @@ class VarDTC_minibatch(LatentFunctionInference):
if het_noise: if het_noise:
if uncertain_inputs: if uncertain_inputs:
psiR = np.einsum('mo,nmo->n',dL_dpsi2R,psi2) psiR = np.einsum('mo,nmo->',dL_dpsi2R,psi2)
else: else:
psiR = np.einsum('nm,no,mo->n',psi1,psi1,dL_dpsi2R) psiR = np.einsum('nm,no,mo->',psi1,psi1,dL_dpsi2R)
dL_dthetaL = ((np.square(betaY)).sum(axis=-1) + np.square(beta)*(output_dim*psi0)-output_dim*beta)/2. - np.square(beta)*psiR- (betaY*np.dot(betapsi1,v)).sum(axis=-1) dL_dthetaL = ((np.square(betaY)).sum(axis=-1) + np.square(beta)*(output_dim*psi0)-output_dim*beta)/2. - np.square(beta)*psiR- (betaY*np.dot(betapsi1,v)).sum(axis=-1)
else: else:
if uncertain_inputs: if uncertain_inputs:
psiR = np.einsum('mo,mo->',dL_dpsi2R,psi2) psiR = np.einsum('mo,nmo->',dL_dpsi2R,psi2)
else: else:
psiR = np.einsum('nm,no,mo->',psi1,psi1,dL_dpsi2R) psiR = np.einsum('nm,no,mo->',psi1,psi1,dL_dpsi2R)
@ -305,30 +308,44 @@ def update_gradients(model):
if isinstance(model.X, VariationalPosterior): if isinstance(model.X, VariationalPosterior):
X_slice = model.X[n_range[0]:n_range[1]] X_slice = model.X[n_range[0]:n_range[1]]
dL_dpsi1 = grad_dict['dL_dpsi1']#[None, :]
dL_dpsi2 = grad_dict['dL_dpsi2'][None, :, :]
#gradients w.r.t. kernel #gradients w.r.t. kernel
model.kern.update_gradients_expectations(variational_posterior=X_slice, Z=model.Z, dL_dpsi0=grad_dict['dL_dpsi0'], dL_dpsi1=grad_dict['dL_dpsi1'], dL_dpsi2=grad_dict['dL_dpsi2']) model.kern.update_gradients_expectations(variational_posterior=X_slice,Z=model.Z,dL_dpsi0=grad_dict['dL_dpsi0'],dL_dpsi1=dL_dpsi1,dL_dpsi2=dL_dpsi2)
kern_grad += model.kern.gradient kern_grad += model.kern.gradient
#gradients w.r.t. Z #gradients w.r.t. Z
model.Z.gradient += model.kern.gradients_Z_expectations( model.Z.gradient += model.kern.gradients_Z_expectations(
dL_dpsi0=grad_dict['dL_dpsi0'], dL_dpsi1=grad_dict['dL_dpsi1'], dL_dpsi2=grad_dict['dL_dpsi2'], Z=model.Z, variational_posterior=X_slice) dL_dpsi0=grad_dict['dL_dpsi0'],
dL_dpsi1=dL_dpsi1,
dL_dpsi2=dL_dpsi2,
Z=model.Z, variational_posterior=X_slice)
#gradients w.r.t. posterior parameters of X #gradients w.r.t. posterior parameters of X
X_grad = model.kern.gradients_qX_expectations(variational_posterior=X_slice, Z=model.Z, dL_dpsi0=grad_dict['dL_dpsi0'], dL_dpsi1=grad_dict['dL_dpsi1'], dL_dpsi2=grad_dict['dL_dpsi2']) X_grad = model.kern.gradients_qX_expectations(
model.set_X_gradients(X_slice, X_grad) variational_posterior=X_slice,
Z=model.Z,
dL_dpsi0=grad_dict['dL_dpsi0'],
dL_dpsi1=dL_dpsi1,
dL_dpsi2=dL_dpsi2)
model.X.mean[n_range[0]:n_range[1]].gradient = X_grad[0]
model.X.variance[n_range[0]:n_range[1]].gradient = X_grad[1]
if het_noise: if het_noise:
dL_dthetaL[n_range[0]:n_range[1]] = grad_dict['dL_dthetaL'] dL_dthetaL[n_range[0]:n_range[1]] = grad_dict['dL_dthetaL']
else: else:
dL_dthetaL += grad_dict['dL_dthetaL'] dL_dthetaL += grad_dict['dL_dthetaL']
#import ipdb;ipdb.set_trace()
model.grad_dict = grad_dict
if isinstance(model.X, VariationalPosterior):
# Update Log-likelihood
model._log_marginal_likelihood -= model.variational_prior.KL_divergence(model.X)
# update for the KL divergence
model.variational_prior.update_gradients_KL(model.X)
# Set the gradients w.r.t. kernel # Set the gradients w.r.t. kernel
model.kern.gradient = kern_grad model.kern.gradient = kern_grad
# Update Log-likelihood
model._log_marginal_likelihood -= model.variational_prior.KL_divergence(model.X)
# update for the KL divergence
model.variational_prior.update_gradients_KL(model.X)
# dL_dthetaL # dL_dthetaL
model.likelihood.update_gradients(dL_dthetaL) model.likelihood.update_gradients(dL_dthetaL)