mirror of
https://github.com/SheffieldML/GPy.git
synced 2026-05-24 14:15:14 +02:00
[GPU] update gradients rest
This commit is contained in:
Zhenwen Dai
parent
73f690a4c9
commit
b90a867232
2 changed files with 106 additions and 38 deletions
|
|
@ -246,7 +246,7 @@ try:
|
|||
dpsi2_dgamma[IDX_NMMQ(n,m1,m2,q)] = var2*neq*(psi2exp1_c/denom_sqrt - psi2exp2_c);
|
||||
dpsi2_dmu[IDX_NMMQ(n,m1,m2,q)] = var2*neq*(-2.0*psi2_common*muZ*psi2exp1_c);
|
||||
dpsi2_dS[IDX_NMMQ(n,m1,m2,q)] = var2*neq*(psi2_common*(2.0*muZ*muZ/(2.0*S_c+l_c)-1.0)*psi2exp1_c);
|
||||
dpsi2_dZ[IDX_NMMQ(n,m1,m2,q)] = var2*neq*(psi2_common*(dZ*denom/-2.0+muZ)*psi2exp1_c-gamma1*Z1_c/l_c*psi2exp2_c)*2.0;
|
||||
dpsi2_dZ[IDX_NMMQ(n,m1,m2,q)] = var2*neq*(psi2_common*(dZ*denom/-2.0+muZ)*psi2exp1_c-gamma1*Z2_c/l_c*psi2exp2_c)*2.0;
|
||||
return var2*neq*(psi2_common*(S_c/l_c+dZ*dZ*denom/(4.0*l_c)+muZ*muZ/(2.0*S_c+l_c))*psi2exp1_c+gamma1*Z2/(2.0*l_c)*psi2exp2_c)*l_sqrt_c*2.0;
|
||||
}
|
||||
""")
|
||||
|
|
@ -411,7 +411,7 @@ class PSICOMP_SSRBF(object):
|
|||
dpsi2_dl_gpu = self.gpuCache['dpsi2_dl_gpu']
|
||||
psi1_comb_gpu = self.gpuCache['psi1_neq_gpu']
|
||||
psi2_comb_gpu = self.gpuCache['psi2_neq_gpu']
|
||||
grad_dl_gpu = self.gpuCache['grad_l_gpu']
|
||||
grad_l_gpu = self.gpuCache['grad_l_gpu']
|
||||
|
||||
# variance
|
||||
variance.gradient = gpuarray.sum(dL_dpsi0).get() \
|
||||
|
|
@ -420,22 +420,78 @@ class PSICOMP_SSRBF(object):
|
|||
|
||||
# lengscale
|
||||
if ARD:
|
||||
grad_dl_gpu.fill(0.)
|
||||
grad_l_gpu.fill(0.)
|
||||
linalg_gpu.mul_bcast(psi1_comb_gpu, dL_dpsi1, dpsi1_dl_gpu, dL_dpsi1.size)
|
||||
linalg_gpu.sum_axis(grad_dl_gpu, psi1_comb_gpu, 1, N*M)
|
||||
linalg_gpu.sum_axis(grad_l_gpu, psi1_comb_gpu, 1, N*M)
|
||||
linalg_gpu.mul_bcast(psi2_comb_gpu, dL_dpsi2, dpsi2_dl_gpu, dL_dpsi2.size)
|
||||
linalg_gpu.sum_axis(grad_dl_gpu, psi2_comb_gpu, 1, N*M*M)
|
||||
lengthscale.gradient = grad_dl_gpu.get()
|
||||
linalg_gpu.sum_axis(grad_l_gpu, psi2_comb_gpu, 1, N*M*M)
|
||||
lengthscale.gradient = grad_l_gpu.get()
|
||||
else:
|
||||
linalg_gpu.mul_bcast(psi1_comb_gpu, dL_dpsi1, dpsi1_dl_gpu, dL_dpsi1.size)
|
||||
linalg_gpu.mul_bcast(psi2_comb_gpu, dL_dpsi2, dpsi2_dl_gpu, dL_dpsi2.size)
|
||||
lengthscale.gradient = gpuarray.sum(psi1_comb_gpu).get() + gpuarray.sum(psi2_comb_gpu).get()
|
||||
|
||||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, mu, S, gamma):
|
||||
pass
|
||||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior):
|
||||
mu = variational_posterior.mean
|
||||
S = variational_posterior.variance
|
||||
gamma = variational_posterior.binary_prob
|
||||
self._psiDercomputations(variance, lengthscale, Z, mu, S, gamma)
|
||||
N, M, Q = mu.shape[0],Z.shape[0], mu.shape[1]
|
||||
|
||||
dpsi1_dZ_gpu = self.gpuCache['dpsi1_dZ_gpu']
|
||||
dpsi2_dZ_gpu = self.gpuCache['dpsi2_dZ_gpu']
|
||||
psi1_comb_gpu = self.gpuCache['psi1_neq_gpu']
|
||||
psi2_comb_gpu = self.gpuCache['psi2_neq_gpu']
|
||||
grad_Z_gpu = self.gpuCache['grad_Z_gpu']
|
||||
|
||||
grad_Z_gpu.fill(0.)
|
||||
linalg_gpu.mul_bcast(psi1_comb_gpu, dL_dpsi1, dpsi1_dZ_gpu, dL_dpsi1.size)
|
||||
linalg_gpu.sum_axis(grad_Z_gpu, psi1_comb_gpu, 1, N)
|
||||
linalg_gpu.mul_bcast(psi2_comb_gpu, dL_dpsi2, dpsi2_dZ_gpu, dL_dpsi2.size)
|
||||
linalg_gpu.sum_axis(grad_Z_gpu, psi2_comb_gpu, 1, N*M)
|
||||
return grad_Z_gpu.get()
|
||||
|
||||
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, mu, S, gamma):
|
||||
pass
|
||||
def gradients_qX_expectations(self, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior):
|
||||
mu = variational_posterior.mean
|
||||
S = variational_posterior.variance
|
||||
gamma = variational_posterior.binary_prob
|
||||
self._psiDercomputations(variance, lengthscale, Z, mu, S, gamma)
|
||||
N, M, Q = mu.shape[0],Z.shape[0], mu.shape[1]
|
||||
|
||||
dpsi1_dmu_gpu = self.gpuCache['dpsi1_dmu_gpu']
|
||||
dpsi2_dmu_gpu = self.gpuCache['dpsi2_dmu_gpu']
|
||||
dpsi1_dS_gpu = self.gpuCache['dpsi1_dS_gpu']
|
||||
dpsi2_dS_gpu = self.gpuCache['dpsi2_dS_gpu']
|
||||
dpsi1_dgamma_gpu = self.gpuCache['dpsi1_dgamma_gpu']
|
||||
dpsi2_dgamma_gpu = self.gpuCache['dpsi2_dgamma_gpu']
|
||||
psi1_comb_gpu = self.gpuCache['psi1_neq_gpu']
|
||||
psi2_comb_gpu = self.gpuCache['psi2_neq_gpu']
|
||||
grad_mu_gpu = self.gpuCache['grad_mu_gpu']
|
||||
grad_S_gpu = self.gpuCache['grad_S_gpu']
|
||||
grad_gamma_gpu = self.gpuCache['grad_gamma_gpu']
|
||||
|
||||
# mu gradients
|
||||
grad_mu_gpu.fill(0.)
|
||||
linalg_gpu.mul_bcast(psi1_comb_gpu, dL_dpsi1, dpsi1_dmu_gpu, dL_dpsi1.size)
|
||||
linalg_gpu.sum_axis(grad_mu_gpu, psi1_comb_gpu, N, M)
|
||||
linalg_gpu.mul_bcast(psi2_comb_gpu, dL_dpsi2, dpsi2_dmu_gpu, dL_dpsi2.size)
|
||||
linalg_gpu.sum_axis(grad_mu_gpu, psi2_comb_gpu, N, M*M)
|
||||
|
||||
# S gradients
|
||||
grad_S_gpu.fill(0.)
|
||||
linalg_gpu.mul_bcast(psi1_comb_gpu, dL_dpsi1, dpsi1_dS_gpu, dL_dpsi1.size)
|
||||
linalg_gpu.sum_axis(grad_S_gpu, psi1_comb_gpu, N, M)
|
||||
linalg_gpu.mul_bcast(psi2_comb_gpu, dL_dpsi2, dpsi2_dS_gpu, dL_dpsi2.size)
|
||||
linalg_gpu.sum_axis(grad_S_gpu, psi2_comb_gpu, N, M*M)
|
||||
|
||||
# gamma gradients
|
||||
grad_gamma_gpu.fill(0.)
|
||||
linalg_gpu.mul_bcast(psi1_comb_gpu, dL_dpsi1, dpsi1_dgamma_gpu, dL_dpsi1.size)
|
||||
linalg_gpu.sum_axis(grad_gamma_gpu, psi1_comb_gpu, N, M)
|
||||
linalg_gpu.mul_bcast(psi2_comb_gpu, dL_dpsi2, dpsi2_dgamma_gpu, dL_dpsi2.size)
|
||||
linalg_gpu.sum_axis(grad_gamma_gpu, psi2_comb_gpu, N, M*M)
|
||||
|
||||
return grad_mu_gpu.get(), grad_S_gpu.get(), grad_gamma_gpu.get()
|
||||
|
||||
@Cache_this(limit=1)
|
||||
def _Z_distances(Z):
|
||||
|
|
|
|||
|
|
@ -73,36 +73,33 @@ class RBF(Stationary):
|
|||
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
# Spike-and-Slab GPLVM
|
||||
if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
|
||||
dL_dpsi0_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi0))
|
||||
dL_dpsi1_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi1))
|
||||
dL_dpsi2_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi2))
|
||||
self.psicomp.update_gradients_expectations(dL_dpsi0_gpu, dL_dpsi1_gpu, dL_dpsi2_gpu, self.variance, self.lengthscale, Z, variational_posterior)
|
||||
vg = self.variance.gradient.copy()
|
||||
lg = self.lengthscale.gradient.copy()
|
||||
|
||||
_, _dpsi1_dvariance, _, _, _, _, _dpsi1_dlengthscale = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
|
||||
_, _dpsi2_dvariance, _, _, _, _, _dpsi2_dlengthscale = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
|
||||
|
||||
#contributions from psi0:
|
||||
self.variance.gradient = np.sum(dL_dpsi0)
|
||||
|
||||
#from psi1
|
||||
self.variance.gradient += np.sum(dL_dpsi1 * _dpsi1_dvariance)
|
||||
if self.ARD:
|
||||
self.lengthscale.gradient = (dL_dpsi1[:,:,None]*_dpsi1_dlengthscale).reshape(-1,self.input_dim).sum(axis=0)
|
||||
if self.useGPU:
|
||||
dL_dpsi0_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi0))
|
||||
dL_dpsi1_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi1))
|
||||
dL_dpsi2_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi2))
|
||||
self.psicomp.update_gradients_expectations(dL_dpsi0_gpu, dL_dpsi1_gpu, dL_dpsi2_gpu, self.variance, self.lengthscale, Z, variational_posterior)
|
||||
else:
|
||||
self.lengthscale.gradient = (dL_dpsi1[:,:,None]*_dpsi1_dlengthscale).sum()
|
||||
|
||||
#from psi2
|
||||
self.variance.gradient += (dL_dpsi2 * _dpsi2_dvariance).sum()
|
||||
if self.ARD:
|
||||
self.lengthscale.gradient += (dL_dpsi2[:,:,:,None] * _dpsi2_dlengthscale).reshape(-1,self.input_dim).sum(axis=0)
|
||||
else:
|
||||
self.lengthscale.gradient += (dL_dpsi2[:,:,:,None] * _dpsi2_dlengthscale).sum()
|
||||
|
||||
print np.abs(vg-self.variance.gradient)
|
||||
print np.abs(lg-self.lengthscale.gradient)
|
||||
|
||||
_, _dpsi1_dvariance, _, _, _, _, _dpsi1_dlengthscale = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
|
||||
_, _dpsi2_dvariance, _, _, _, _, _dpsi2_dlengthscale = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
|
||||
|
||||
#contributions from psi0:
|
||||
self.variance.gradient = np.sum(dL_dpsi0)
|
||||
|
||||
#from psi1
|
||||
self.variance.gradient += np.sum(dL_dpsi1 * _dpsi1_dvariance)
|
||||
if self.ARD:
|
||||
self.lengthscale.gradient = (dL_dpsi1[:,:,None]*_dpsi1_dlengthscale).reshape(-1,self.input_dim).sum(axis=0)
|
||||
else:
|
||||
self.lengthscale.gradient = (dL_dpsi1[:,:,None]*_dpsi1_dlengthscale).sum()
|
||||
|
||||
#from psi2
|
||||
self.variance.gradient += (dL_dpsi2 * _dpsi2_dvariance).sum()
|
||||
if self.ARD:
|
||||
self.lengthscale.gradient += (dL_dpsi2[:,:,:,None] * _dpsi2_dlengthscale).reshape(-1,self.input_dim).sum(axis=0)
|
||||
else:
|
||||
self.lengthscale.gradient += (dL_dpsi2[:,:,:,None] * _dpsi2_dlengthscale).sum()
|
||||
|
||||
elif isinstance(variational_posterior, variational.NormalPosterior):
|
||||
l2 = self.lengthscale**2
|
||||
if l2.size != self.input_dim:
|
||||
|
|
@ -141,6 +138,12 @@ class RBF(Stationary):
|
|||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
# Spike-and-Slab GPLVM
|
||||
if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
|
||||
dL_dpsi1_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi1))
|
||||
dL_dpsi2_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi2))
|
||||
gZ = self.psicomp.gradients_Z_expectations(dL_dpsi1_gpu, dL_dpsi2_gpu, self.variance, self.lengthscale, Z, variational_posterior)
|
||||
|
||||
|
||||
|
||||
_, _, _, _, _, _dpsi1_dZ, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
|
||||
_, _, _, _, _, _dpsi2_dZ, _ = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
|
||||
|
||||
|
|
@ -150,6 +153,8 @@ class RBF(Stationary):
|
|||
#psi2
|
||||
grad += (dL_dpsi2[:, :, :, None] * _dpsi2_dZ).sum(axis=0).sum(axis=1)
|
||||
|
||||
print np.abs(gZ - grad).max()
|
||||
|
||||
return grad
|
||||
|
||||
elif isinstance(variational_posterior, variational.NormalPosterior):
|
||||
|
|
@ -174,6 +179,11 @@ class RBF(Stationary):
|
|||
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
# Spike-and-Slab GPLVM
|
||||
if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
|
||||
dL_dpsi1_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi1))
|
||||
dL_dpsi2_gpu = gpuarray.to_gpu(np.asfortranarray(dL_dpsi2))
|
||||
gmu,gS,gg = self.psicomp.gradients_qX_expectations(dL_dpsi1_gpu, dL_dpsi2_gpu, self.variance, self.lengthscale, Z, variational_posterior)
|
||||
|
||||
|
||||
ndata = variational_posterior.mean.shape[0]
|
||||
|
||||
_, _, _dpsi1_dgamma, _dpsi1_dmu, _dpsi1_dS, _, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
|
||||
|
|
@ -191,6 +201,8 @@ class RBF(Stationary):
|
|||
|
||||
if self.group_spike_prob:
|
||||
grad_gamma[:] = grad_gamma.mean(axis=0)
|
||||
|
||||
print np.abs(gmu-grad_mu).max(),np.abs(gS-grad_S).max(),np.abs(gg-grad_gamma).max()
|
||||
|
||||
return grad_mu, grad_S, grad_gamma
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue