apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-03 16:52:39 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

effiiency improvements in sparse_GP

Browse source 
the recasting of derivatives through psi2 into psi1 is now only done in
one place
This commit is contained in:
James Hensman 2013-02-26 15:00:18 +00:00
parent 4d79c3c97d
commit 6e809b024f
2 changed files with 13 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

23
GPy/models/sparse_GP.py
View file

@ -106,12 +106,20 @@ class sparse_GP(GP):
self.dL_dpsi2 = 0.5 * self.likelihood.precision[:,None,None] * self.D * self.Kmmi[None,:,:] # dB self.dL_dpsi2 = 0.5 * self.likelihood.precision[:,None,None] * self.D * self.Kmmi[None,:,:] # dB
self.dL_dpsi2 += - 0.5 * self.likelihood.precision[:,None,None]/sf2 * self.D * self.C[None,:,:] # dC self.dL_dpsi2 += - 0.5 * self.likelihood.precision[:,None,None]/sf2 * self.D * self.C[None,:,:] # dC
self.dL_dpsi2 += - 0.5 * self.likelihood.precision[:,None,None]* self.E[None,:,:] # dD self.dL_dpsi2 += - 0.5 * self.likelihood.precision[:,None,None]* self.E[None,:,:] # dD
if not self.has_uncertain_inputs:
raise NotImplementedError, "TODO: recaste derivatibes in psi2 back into psi1"
else: else:
self.dL_dpsi2 = 0.5 * self.likelihood.precision * self.D * self.Kmmi # dB self.dL_dpsi2 = 0.5 * self.likelihood.precision * self.D * self.Kmmi # dB
self.dL_dpsi2 += - 0.5 * self.likelihood.precision/sf2 * self.D * self.C # dC self.dL_dpsi2 += - 0.5 * self.likelihood.precision/sf2 * self.D * self.C # dC
self.dL_dpsi2 += - 0.5 * self.likelihood.precision * self.E # dD self.dL_dpsi2 += - 0.5 * self.likelihood.precision * self.E # dD
if self.has_uncertain_inputs:
#repeat for each of the N psi_2 matrices #repeat for each of the N psi_2 matrices
self.dL_dpsi2 = np.repeat(self.dL_dpsi2[None,:,:],self.N,axis=0) self.dL_dpsi2 = np.repeat(self.dL_dpsi2[None,:,:],self.N,axis=0)
else:
self.dL_dpsi1 += 2.*np.dot(self.dL_dpsi2,self.psi1)
self.dL_dpsi2 = None
# Compute dL_dKmm # Compute dL_dKmm
self.dL_dKmm = -0.5 * self.D * mdot(self.Lmi.T, self.A, self.Lmi)*sf2 # dB self.dL_dKmm = -0.5 * self.D * mdot(self.Lmi.T, self.A, self.Lmi)*sf2 # dB
@ -175,13 +183,7 @@ class sparse_GP(GP):
dL_dtheta += self.kern.dpsi1_dtheta(self.dL_dpsi1.T,self.Z,self.X, self.X_uncertainty) dL_dtheta += self.kern.dpsi1_dtheta(self.dL_dpsi1.T,self.Z,self.X, self.X_uncertainty)
dL_dtheta += self.kern.dpsi2_dtheta(self.dL_dpsi2,self.dL_dpsi1.T, self.Z,self.X, self.X_uncertainty) dL_dtheta += self.kern.dpsi2_dtheta(self.dL_dpsi2,self.dL_dpsi1.T, self.Z,self.X, self.X_uncertainty)
else: else:
#re-cast computations in psi2 back to psi1: dL_dtheta += self.kern.dK_dtheta(self.dL_dpsi1,self.Z,self.X)
#dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2.sum(0),self.psi1)
if not self.likelihood.is_heteroscedastic:
dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2[0,:,:],self.psi1)
else:
raise NotImplementedError, "TODO"
dL_dtheta += self.kern.dK_dtheta(dL_dpsi1,self.Z,self.X)
dL_dtheta += self.kern.dKdiag_dtheta(self.dL_dpsi0, self.X) dL_dtheta += self.kern.dKdiag_dtheta(self.dL_dpsi0, self.X)
return dL_dtheta return dL_dtheta
@ -195,12 +197,7 @@ class sparse_GP(GP):
dL_dZ += self.kern.dpsi1_dZ(self.dL_dpsi1,self.Z,self.X, self.X_uncertainty) dL_dZ += self.kern.dpsi1_dZ(self.dL_dpsi1,self.Z,self.X, self.X_uncertainty)
dL_dZ += 2.*self.kern.dpsi2_dZ(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty) # 'stripes' dL_dZ += 2.*self.kern.dpsi2_dZ(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty) # 'stripes'
else: else:
#re-cast computations in psi2 back to psi1: dL_dZ += self.kern.dK_dX(self.dL_dpsi1,self.Z,self.X)
if not self.likelihood.is_heteroscedastic:
dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2[0,:,:],self.psi1)
else:
raise NotImplementedError, "TODO"
dL_dZ += self.kern.dK_dX(dL_dpsi1,self.Z,self.X)
return dL_dZ return dL_dZ
def _raw_predict(self, Xnew, slices, full_cov=False): def _raw_predict(self, Xnew, slices, full_cov=False):

5
GPy/models/sparse_GPLVM.py
View file

@ -42,11 +42,8 @@ class sparse_GPLVM(sparse_GP_regression, GPLVM):
return sparse_GP_regression.log_likelihood(self) return sparse_GP_regression.log_likelihood(self)
def dL_dX(self): def dL_dX(self):
#dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2,self.psi1)
dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2[0,:,:],self.psi1)
dL_dX = self.kern.dKdiag_dX(self.dL_dpsi0,self.X) dL_dX = self.kern.dKdiag_dX(self.dL_dpsi0,self.X)
dL_dX += self.kern.dK_dX(dL_dpsi1.T,self.X,self.Z) dL_dX += self.kern.dK_dX(self.dL_dpsi1,self.X,self.Z)
return dL_dX return dL_dX