apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-04 17:22:39 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Try calculating dL_dpsi1*psi1 individually for each dimension as we go along

Browse source
This commit is contained in:
Alan Saul 2015-08-31 14:13:35 +03:00
parent c83f56723e
commit 3818aa3745
7 changed files with 47 additions and 31 deletions
Download patch file Download diff file Expand all files Collapse all files

12
GPy/kern/_src/kern.py
View file

@ -117,7 +117,7 @@ class Kern(Parameterized):
raise NotImplementedError
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
"""
Set the gradients of all parameters when doing inference with
uncertain inputs, using expectations of the kernel.
@ -129,26 +129,26 @@ class Kern(Parameterized):
dL_dpsi2 * dpsi2_d{theta_i}
"""
dtheta = self.psicomp.psiDerivativecomputations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=psi0, psi1=psi1, psi2=psi2)[0]
psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)[0]
self.gradient[:] = dtheta
def gradients_Z_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
"""
Returns the derivative of the objective wrt Z, using the chain rule
through the expectation variables.
"""
return self.psicomp.psiDerivativecomputations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=psi0, psi1=psi1, psi2=psi2)[1]
psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)[1]
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
"""
Compute the gradients wrt the parameters of the variational
distruibution q(X), chain-ruling via the expectations of the kernel
"""
return self.psicomp.psiDerivativecomputations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=psi0, psi1=psi1, psi2=psi2)[2:]
psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)[2:]
def plot(self, x=None, fignum=None, ax=None, title=None, plot_limits=None, resolution=None, **mpl_kwargs):
"""

12
GPy/kern/_src/kernel_slice_operations.py
View file

@ -117,30 +117,30 @@ def _slice_psi(f):
def _slice_update_gradients_expectations(f):
@wraps(f)
def wrap(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
with _Slice_wrap(self, Z, variational_posterior) as s:
ret = f(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, s.X, s.X2,
psi0=psi0, psi1=psi1, psi2=psi2)
psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)
return ret
return wrap
def _slice_gradients_Z_expectations(f):
@wraps(f)
def wrap(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
with _Slice_wrap(self, Z, variational_posterior) as s:
ret = s.handle_return_array(f(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, s.X, s.X2,
psi0=psi0, psi1=psi1, psi2=psi2))
psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2))
return ret
return wrap
def _slice_gradients_qX_expectations(f):
@wraps(f)
def wrap(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
with _Slice_wrap(self, variational_posterior, Z) as s:
ret = list(f(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, s.X2, s.X,
psi0=psi0, psi1=psi1, psi2=psi2))
psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2))
r2 = ret[:2]
ret[0] = s.handle_return_array(r2[0])
ret[1] = s.handle_return_array(r2[1])

4
GPy/kern/_src/psi_comp/__init__.py
View file

@ -24,10 +24,10 @@ class PSICOMP_RBF(Pickleable):
@Cache_this(limit=10, ignore_args=(0,1,2,3))
def psiDerivativecomputations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
if isinstance(variational_posterior, variational.NormalPosterior):
return rbf_psi_comp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior,
psi0=psi0, psi1=psi1, psi2=psi2)
psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)
elif isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
return ssrbf_psi_comp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior)
else:

16
GPy/kern/_src/psi_comp/rbf_psi_comp.py
View file

@ -69,11 +69,11 @@ def __psi2computations(variance, lengthscale, Z, mu, S):
return _psi2
def psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscale, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
ARD = (len(lengthscale)!=1)
dvar_psi1, dl_psi1, dZ_psi1, dmu_psi1, dS_psi1 = _psi1compDer(dL_dpsi1, variance, lengthscale, Z, variational_posterior.mean, variational_posterior.variance, psi1=psi1)
dvar_psi2, dl_psi2, dZ_psi2, dmu_psi2, dS_psi2 = _psi2compDer(dL_dpsi2, variance, lengthscale, Z, variational_posterior.mean, variational_posterior.variance, psi2=psi2)
dvar_psi1, dl_psi1, dZ_psi1, dmu_psi1, dS_psi1 = _psi1compDer(dL_dpsi1, variance, lengthscale, Z, variational_posterior.mean, variational_posterior.variance, psi1=psi1, Lpsi1=Lpsi1)
dvar_psi2, dl_psi2, dZ_psi2, dmu_psi2, dS_psi2 = _psi2compDer(dL_dpsi2, variance, lengthscale, Z, variational_posterior.mean, variational_posterior.variance, psi2=psi2, Lpsi2=Lpsi2)
dL_dvar = np.sum(dL_dpsi0) + dvar_psi1 + dvar_psi2
@ -87,7 +87,7 @@ def psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, variance, lengthscal
return dL_dvar, dL_dlengscale, dL_dZ, dL_dmu, dL_dS
def __psi1compDer(dL_dpsi1, variance, lengthscale, Z, mu, S, psi1=None):
def __psi1compDer(dL_dpsi1, variance, lengthscale, Z, mu, S, psi1=None, Lpsi1=None):
"""
dL_dpsi1 - NxM
Z - MxQ
@ -108,7 +108,8 @@ def __psi1compDer(dL_dpsi1, variance, lengthscale, Z, mu, S, psi1=None):
if psi1 is None:
psi1 = _psi1computations(variance, lengthscale, Z, mu, S)
Lpsi1 = dL_dpsi1*psi1
if Lpsi1 is None:
Lpsi1 = dL_dpsi1*psi1
Zmu = Z[None,:,:]-mu[:,None,:] # NxMxQ
denom = 1./(S+lengthscale2)
Zmu2_denom = np.square(Zmu)*denom[:,None,:] #NxMxQ
@ -120,7 +121,7 @@ def __psi1compDer(dL_dpsi1, variance, lengthscale, Z, mu, S, psi1=None):
return _dL_dvar, _dL_dl, _dL_dZ, _dL_dmu, _dL_dS
def __psi2compDer(dL_dpsi2, variance, lengthscale, Z, mu, S, psi2=None):
def __psi2compDer(dL_dpsi2, variance, lengthscale, Z, mu, S, psi2=None, Lpsi2=None):
"""
Z - MxQ
mu - NxQ
@ -143,7 +144,8 @@ def __psi2compDer(dL_dpsi2, variance, lengthscale, Z, mu, S, psi2=None):
if psi2 is None:
psi2 = _psi2computations(variance, lengthscale, Z, mu, S) # NxMxM
Lpsi2 = dL_dpsi2*psi2 # dL_dpsi2 is MxM, using broadcast to multiply N out
if Lpsi2 is None:
Lpsi2 = dL_dpsi2*psi2 # dL_dpsi2 is MxM, using broadcast to multiply N out
Lpsi2sum = np.einsum('nmo->n',Lpsi2) #N
Lpsi2Z = np.einsum('nmo,oq->nq',Lpsi2,Z) #NxQ
Lpsi2Z2 = np.einsum('nmo,oq,oq->nq',Lpsi2,Z,Z) #NxQ

12
GPy/kern/_src/rbf.py
View file

@ -59,16 +59,16 @@ class RBF(Stationary):
return self.psicomp.psicomputations(self.variance, self.lengthscale, Z, variational_posterior, return_psi2_n=self.return_psi2_n)[2]
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
dL_dvar, dL_dlengscale = self.psicomp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, self.variance, self.lengthscale, Z, variational_posterior, psi0=psi0, psi1=psi1, psi2=psi2)[:2]
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
dL_dvar, dL_dlengscale = self.psicomp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, self.variance, self.lengthscale, Z, variational_posterior, psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)[:2]
self.variance.gradient = dL_dvar
self.lengthscale.gradient = dL_dlengscale
def gradients_Z_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
return self.psicomp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, self.variance, self.lengthscale, Z, variational_posterior, psi0=psi0, psi1=psi1, psi2=psi2)[2]
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
return self.psicomp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, self.variance, self.lengthscale, Z, variational_posterior, psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)[2]
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior,
psi0=None, psi1=None, psi2=None):
return self.psicomp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, self.variance, self.lengthscale, Z, variational_posterior, psi0=psi0, psi1=psi1, psi2=psi2)[3:]
psi0=None, psi1=None, psi2=None, Lpsi0=None, Lpsi1=None, Lpsi2=None):
return self.psicomp.psiDerivativecomputations(dL_dpsi0, dL_dpsi1, dL_dpsi2, self.variance, self.lengthscale, Z, variational_posterior, psi0=psi0, psi1=psi1, psi2=psi2, Lpsi0=Lpsi0, Lpsi1=Lpsi1, Lpsi2=Lpsi2)[3:]

8
GPy/models/bayesian_gplvm_minibatch.py
View file

@ -126,7 +126,8 @@ class BayesianGPLVMMiniBatch(SparseGPMiniBatch):
Z=self.Z, dL_dpsi0=full_values['dL_dpsi0'],
dL_dpsi1=full_values['dL_dpsi1'],
dL_dpsi2=full_values['dL_dpsi2'],
psi0=self.psi0, psi1=self.psi1, psi2=self.psi2)
psi0=self.psi0, psi1=self.psi1, psi2=self.psi2,
Lpsi0=full_values['Lpsi0'], Lpsi1=full_values['Lpsi1'], Lpsi2=full_values['Lpsi2'])
full_values['meangrad'] += meangrad_tmp
full_values['vargrad'] += vargrad_tmp
else:
@ -156,6 +157,11 @@ class BayesianGPLVMMiniBatch(SparseGPMiniBatch):
full_values['vargrad'] = np.zeros((self.X.shape[0], self.X.shape[1]))
full_values['dL_dpsi0'] = np.zeros(self.X.shape[0])
full_values['dL_dpsi1'] = np.zeros((self.X.shape[0], self.Z.shape[0]))
full_values['dL_dpsi2'] = np.zeros((self.Z.shape[0], self.Z.shape[0]))
full_values['Lpsi0'] = np.zeros(self.X.shape[0])
full_values['Lpsi1'] = np.zeros((self.X.shape[0], self.Z.shape[0]))
full_values['Lpsi2'] = np.zeros((self.X.shape[0], self.Z.shape[0], self.Z.shape[0]))
return full_values
def parameters_changed(self):

14
GPy/models/sparse_gp_minibatch.py
View file

@ -106,6 +106,10 @@ class SparseGPMiniBatch(SparseGP):
posterior, log_marginal_likelihood, grad_dict = self.inference_method.inference(kern, X, Z, likelihood, Y, Y_metadata, Lm=Lm,
dL_dKmm=dL_dKmm, psi0=psi0, psi1=psi1, psi2=psi2_sum_n, **kwargs)
if self.has_uncertain_inputs():
grad_dict['Lpsi0'] = grad_dict['dL_dpsi0']*psi0
grad_dict['Lpsi1'] = grad_dict['dL_dpsi1']*psi1
grad_dict['Lpsi2'] = grad_dict['dL_dpsi2']*psi2
return posterior, log_marginal_likelihood, grad_dict
def _inner_take_over_or_update(self, full_values=None, current_values=None, value_indices=None):
@ -172,7 +176,8 @@ class SparseGPMiniBatch(SparseGP):
Z=self.Z, dL_dpsi0=full_values['dL_dpsi0'],
dL_dpsi1=full_values['dL_dpsi1'],
dL_dpsi2=full_values['dL_dpsi2'],
psi0=self.psi0, psi1=self.psi1, psi2=self.psi2)
psi0=self.psi0, psi1=self.psi1, psi2=self.psi2,
Lpsi0=full_values['Lpsi0'], Lpsi1=full_values['Lpsi1'], Lpsi2=full_values['Lpsi2'])
#self.kern.update_gradients_expectations(variational_posterior=self.X,
#Z=self.Z,
#dL_dpsi0=full_values['dL_dpsi0'],
@ -187,7 +192,8 @@ class SparseGPMiniBatch(SparseGP):
Z=self.Z, dL_dpsi0=full_values['dL_dpsi0'],
dL_dpsi1=full_values['dL_dpsi1'],
dL_dpsi2=full_values['dL_dpsi2'],
psi0=self.psi0, psi1=self.psi1, psi2=self.psi2)
psi0=self.psi0, psi1=self.psi1, psi2=self.psi2,
Lpsi0=full_values['Lpsi0'], Lpsi1=full_values['Lpsi1'], Lpsi2=full_values['Lpsi2'])
else:
#gradients wrt kernel
self.kern.update_gradients_diag(full_values['dL_dKdiag'], self.X)
@ -267,7 +273,9 @@ class SparseGPMiniBatch(SparseGP):
psi1ni = psi1[ninan]
if self.has_uncertain_inputs():
psi2ni = psi2[ninan]
value_indices = dict(outputs=d, samples=ninan, dL_dpsi0=ninan, dL_dpsi1=ninan, meangrad=ninan, vargrad=ninan)
#value_indices = dict(outputs=d, samples=ninan, dL_dpsi0=ninan, dL_dpsi1=ninan, meangrad=ninan, vargrad=ninan)
value_indices = dict(outputs=d, samples=ninan, dL_dpsi0=ninan, dL_dpsi1=ninan, meangrad=ninan, vargrad=ninan,
Lpsi0=ninan, Lpsi1=ninan, Lpsi2=ninan)
else:
psi2ni = None
value_indices = dict(outputs=d, samples=ninan, dL_dKdiag=ninan, dL_dKnm=ninan)