apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-07 19:12:40 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

resolve the requirement of dL_dpsi2 to be symmetric

Browse source
This commit is contained in:
Zhenwen Dai 2015-10-22 17:09:22 +01:00
parent 6b6938bd11
commit e18de2b2c0
4 changed files with 19 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

17
GPy/kern/src/add.py
View file

@ -181,6 +181,8 @@ class Add(CombinationKernel):
return psi2
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
tmp = dL_dpsi2.sum(0)+ dL_dpsi2.sum(1) if len(dL_dpsi2.shape)==2 else dL_dpsi2.sum(2)+ dL_dpsi2.sum(1)
if not self._exact_psicomp: return Kern.update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior)
from .static import White, Bias
for p1 in self.parts:
@ -192,12 +194,13 @@ class Add(CombinationKernel):
if isinstance(p2, White):
continue
elif isinstance(p2, Bias):
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
eff_dL_dpsi1 += tmp * p2.variance
else:# np.setdiff1d(p1._all_dims_active, ar2, assume_unique): # TODO: Careful, not correct for overlapping _all_dims_active
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z, variational_posterior) * 2.
eff_dL_dpsi1 += tmp * p2.psi1(Z, variational_posterior)
p1.update_gradients_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, Z, variational_posterior)
def gradients_Z_expectations(self, dL_psi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
tmp = dL_dpsi2.sum(0)+ dL_dpsi2.sum(1) if len(dL_dpsi2.shape)==2 else dL_dpsi2.sum(2)+ dL_dpsi2.sum(1)
if not self._exact_psicomp: return Kern.gradients_Z_expectations(self, dL_psi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior)
from .static import White, Bias
target = np.zeros(Z.shape)
@ -210,13 +213,15 @@ class Add(CombinationKernel):
if isinstance(p2, White):
continue
elif isinstance(p2, Bias):
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
eff_dL_dpsi1 += tmp * p2.variance
else:
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z, variational_posterior) * 2.
eff_dL_dpsi1 += tmp * p2.psi1(Z, variational_posterior)
target += p1.gradients_Z_expectations(dL_psi0, eff_dL_dpsi1, dL_dpsi2, Z, variational_posterior)
return target
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
tmp = dL_dpsi2.sum(0)+ dL_dpsi2.sum(1) if len(dL_dpsi2.shape)==2 else dL_dpsi2.sum(2)+ dL_dpsi2.sum(1)
if not self._exact_psicomp: return Kern.gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior)
from .static import White, Bias
target_grads = [np.zeros(v.shape) for v in variational_posterior.parameters]
@ -229,9 +234,9 @@ class Add(CombinationKernel):
if isinstance(p2, White):
continue
elif isinstance(p2, Bias):
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
eff_dL_dpsi1 += tmp * p2.variance
else:
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z, variational_posterior) * 2.
eff_dL_dpsi1 += tmp * p2.psi1(Z, variational_posterior)
grads = p1.gradients_qX_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, Z, variational_posterior)
[np.add(target_grads[i],grads[i],target_grads[i]) for i in range(len(grads))]
return target_grads

2
GPy/kern/src/psi_comp/rbf_psi_comp.py
View file

@ -106,6 +106,8 @@ def _psi2compDer(dL_dpsi2, variance, lengthscale, Z, mu, S):
denom = 1./(2*S+lengthscale2)
denom2 = np.square(denom)
if len(dL_dpsi2.shape)==2: dL_dpsi2 = (dL_dpsi2+dL_dpsi2.T)/2
else: dL_dpsi2 = (dL_dpsi2+ np.swapaxes(dL_dpsi2, 1,2))/2
_psi2 = _psi2computations(variance, lengthscale, Z, mu, S) # NxMxM
Lpsi2 = dL_dpsi2*_psi2 # dL_dpsi2 is MxM, using broadcast to multiply N out
Lpsi2sum = Lpsi2.reshape(N,M*M).sum(1) #N

4
GPy/kern/src/psi_comp/rbf_psi_gpucomp.py
View file

@ -370,6 +370,10 @@ class PSICOMP_RBF_GPU(PSICOMP_RBF):
@Cache_this(limit=10, ignore_args=(0,2,3,4))
def _psiDerivativecomputations(self, kern, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
# resolve the requirement of dL_dpsi2 to be symmetric
if len(dL_dpsi2.shape)==2: dL_dpsi2 = (dL_dpsi2+dL_dpsi2.T)/2
else: dL_dpsi2 = (dL_dpsi2+ np.swapaxes(dL_dpsi2, 1,2))/2
variance, lengthscale = kern.variance, kern.lengthscale
from ....util.linalg_gpu import sum_axis
ARD = (len(lengthscale)!=1)