mirror of
https://github.com/SheffieldML/GPy.git
synced 2026-05-21 14:05:14 +02:00
messing with kernels
This commit is contained in:
James Hensman
parent
6a667e749f
commit
80acca640f
8 changed files with 66 additions and 57 deletions
|
|
@ -101,7 +101,7 @@ class Add(Kern):
|
|||
raise NotImplementedError, "psi2 cannot be computed for this kernel"
|
||||
return psi2
|
||||
|
||||
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, variational_posterior, Z):
|
||||
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
from white import White
|
||||
from rbf import RBF
|
||||
#from rbf_inv import RBFInv
|
||||
|
|
@ -124,10 +124,10 @@ class Add(Kern):
|
|||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z[:,is2], mu[:,is2], S[:,is2]) * 2.
|
||||
|
||||
|
||||
p1.update_gradients_variational(dL_dKmm, dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], S[:,is1], Z[:,is1])
|
||||
p1.update_gradients_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], S[:,is1], Z[:,is1])
|
||||
|
||||
|
||||
def gradients_Z_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, mu, S, Z):
|
||||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
from white import White
|
||||
from rbf import RBF
|
||||
#from rbf_inv import rbfinv
|
||||
|
|
@ -151,10 +151,10 @@ class Add(Kern):
|
|||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z[:,is2], mu[:,is2], S[:,is2]) * 2.
|
||||
|
||||
|
||||
target += p1.gradients_z_variational(dL_dKmm, dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], S[:,is1], Z[:,is1])
|
||||
target += p1.gradients_z_variational(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], S[:,is1], Z[:,is1])
|
||||
return target
|
||||
|
||||
def gradients_muS_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, mu, S, Z):
|
||||
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
from white import white
|
||||
from rbf import rbf
|
||||
#from rbf_inv import rbfinv
|
||||
|
|
@ -179,7 +179,7 @@ class Add(Kern):
|
|||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(z[:,is2], mu[:,is2], s[:,is2]) * 2.
|
||||
|
||||
|
||||
a, b = p1.gradients_muS_variational(dL_dkmm, dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], s[:,is1], z[:,is1])
|
||||
a, b = p1.gradients_qX_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], s[:,is1], z[:,is1])
|
||||
target_mu += a
|
||||
target_S += b
|
||||
return target_mu, target_S
|
||||
|
|
|
|||
|
|
@ -39,28 +39,21 @@ class Kern(Parameterized):
|
|||
def update_gradients_full(self, dL_dK, X, X2):
|
||||
"""Set the gradients of all parameters when doing full (N) inference."""
|
||||
raise NotImplementedError
|
||||
def update_gradients_sparse(self, dL_dKmm, dL_dKnm, dL_dKdiag, X, Z):
|
||||
target = np.zeros(self.size)
|
||||
self.update_gradients_diag(dL_dKdiag, X)
|
||||
self._collect_gradient(target)
|
||||
self.update_gradients_full(dL_dKnm, X, Z)
|
||||
self._collect_gradient(target)
|
||||
self.update_gradients_full(dL_dKmm, Z, None)
|
||||
self._collect_gradient(target)
|
||||
self._set_gradient(target)
|
||||
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
"""
|
||||
Set the gradients of all parameters when doing inference with
|
||||
uncertain inputs, using expectations of the kernel.
|
||||
"""
|
||||
raise NotImplementedError
|
||||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
raise NotImplementedError
|
||||
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
"""
|
||||
Compute the gradients wrt the parameters of the variational
|
||||
distruibution q(X), chain-ruling via the expectations of the kernel
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
"""Set the gradients of all parameters when doing variational (M) inference with uncertain inputs."""
|
||||
raise NotImplementedError
|
||||
def gradients_Z_sparse(self, dL_dKmm, dL_dKnm, dL_dKdiag, X, Z):
|
||||
grad = self.gradients_X(dL_dKmm, Z)
|
||||
grad += self.gradients_X(dL_dKnm.T, Z, X)
|
||||
return grad
|
||||
def gradients_Z_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
raise NotImplementedError
|
||||
def gradients_q_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
raise NotImplementedError
|
||||
|
||||
def plot_ARD(self, *args, **kw):
|
||||
if "matplotlib" in sys.modules:
|
||||
from ...plotting.matplot_dep import kernel_plots
|
||||
|
|
@ -68,13 +61,13 @@ class Kern(Parameterized):
|
|||
assert "matplotlib" in sys.modules, "matplotlib package has not been imported."
|
||||
from ...plotting.matplot_dep import kernel_plots
|
||||
return kernel_plots.plot_ARD(self,*args,**kw)
|
||||
|
||||
|
||||
def input_sensitivity(self):
|
||||
"""
|
||||
Returns the sensitivity for each dimension of this kernel.
|
||||
"""
|
||||
return np.zeros(self.input_dim)
|
||||
|
||||
|
||||
def __add__(self, other):
|
||||
""" Overloading of the '+' operator. for more control, see self.add """
|
||||
return self.add(other)
|
||||
|
|
|
|||
|
|
@ -117,7 +117,7 @@ class Linear(Kern):
|
|||
ZAinner = self._ZAinner(variational_posterior, Z)
|
||||
return np.dot(ZAinner, ZA.T)
|
||||
|
||||
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, variational_posterior, Z):
|
||||
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
mu, S = variational_posterior.mean, variational_posterior.variance
|
||||
# psi0:
|
||||
tmp = dL_dpsi0[:, None] * self._mu2S(variational_posterior)
|
||||
|
|
@ -130,20 +130,15 @@ class Linear(Kern):
|
|||
tmp = dL_dpsi2[:, :, :, None] * (self._ZAinner(variational_posterior, Z)[:, :, None, :] * (2. * Z)[None, None, :, :])
|
||||
if self.ARD: grad += tmp.sum(0).sum(0).sum(0)
|
||||
else: grad += tmp.sum()
|
||||
#from Kmm
|
||||
self.update_gradients_full(dL_dKmm, Z, None)
|
||||
self.variances.gradient += grad
|
||||
|
||||
def gradients_Z_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, variational_posterior, Z):
|
||||
# Kmm
|
||||
grad = self.gradients_X(dL_dKmm, Z, None)
|
||||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
#psi1
|
||||
grad += self.gradients_X(dL_dpsi1.T, Z, variational_posterior.mean)
|
||||
grad = self.gradients_X(dL_dpsi1.T, Z, variational_posterior.mean)
|
||||
#psi2
|
||||
self._weave_dpsi2_dZ(dL_dpsi2, Z, variational_posterior, grad)
|
||||
return grad
|
||||
|
||||
def gradients_q_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, variational_posterior, Z):
|
||||
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
grad_mu, grad_S = np.zeros(variational_posterior.mean.shape), np.zeros(variational_posterior.mean.shape)
|
||||
# psi0
|
||||
grad_mu += dL_dpsi0[:, None] * (2.0 * variational_posterior.mean * self.variances)
|
||||
|
|
|
|||
|
|
@ -55,10 +55,7 @@ class RBF(Stationary):
|
|||
self._psi_computations(Z, mu, S)
|
||||
return self._psi2
|
||||
|
||||
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
#contributions from Kmm
|
||||
sself.update_gradients_full(dL_dKmm, Z)
|
||||
|
||||
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
mu = variational_posterior.mean
|
||||
S = variational_posterior.variance
|
||||
self._psi_computations(Z, mu, S)
|
||||
|
|
@ -87,7 +84,7 @@ class RBF(Stationary):
|
|||
else:
|
||||
self.lengthscale.gradient += dpsi2_dlength.sum(0).sum(0).sum(0)
|
||||
|
||||
def gradients_Z_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
mu = variational_posterior.mean
|
||||
S = variational_posterior.variance
|
||||
self._psi_computations(Z, mu, S)
|
||||
|
|
@ -104,11 +101,9 @@ class RBF(Stationary):
|
|||
dZ = self._psi2[:, :, :, None] * (term1[None] + term2)
|
||||
grad += 2*(dL_dpsi2[:, :, :, None] * dZ).sum(0).sum(0)
|
||||
|
||||
grad += self.gradients_X(dL_dKmm, Z, None)
|
||||
|
||||
return grad
|
||||
|
||||
def gradients_q_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
mu = variational_posterior.mean
|
||||
S = variational_posterior.variance
|
||||
self._psi_computations(Z, mu, S)
|
||||
|
|
|
|||
|
|
@ -25,10 +25,10 @@ class Static(Kern):
|
|||
def gradients_X_diag(self, dL_dKdiag, X):
|
||||
return np.zeros(X.shape)
|
||||
|
||||
def gradients_Z_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
return np.zeros(Z.shape)
|
||||
|
||||
def gradients_muS_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
return np.zeros(variational_posterior.shape), np.zeros(variational_posterior.shape)
|
||||
|
||||
def psi0(self, Z, variational_posterior):
|
||||
|
|
@ -61,8 +61,8 @@ class White(Static):
|
|||
def update_gradients_diag(self, dL_dKdiag, X):
|
||||
self.variance.gradient = dL_dKdiag.sum()
|
||||
|
||||
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
self.variance.gradient = np.trace(dL_dKmm) + dL_dpsi0.sum()
|
||||
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
self.variance.gradient = dL_dpsi0.sum()
|
||||
|
||||
|
||||
class Bias(Static):
|
||||
|
|
@ -86,6 +86,6 @@ class Bias(Static):
|
|||
ret[:] = self.variance**2
|
||||
return ret
|
||||
|
||||
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
self.variance.gradient = dL_dKmm.sum() + dL_dpsi0.sum() + dL_dpsi1.sum() + 2.*self.variance*dL_dpsi2.sum()
|
||||
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
|
||||
self.variance.gradient = dL_dpsi0.sum() + dL_dpsi1.sum() + 2.*self.variance*dL_dpsi2.sum()
|
||||
|
||||
|
|
|
|||
|
|
@ -312,4 +312,8 @@ class RatQuad(Stationary):
|
|||
grad = np.sum(dL_dK*dK_dpow)
|
||||
self.power.gradient = grad
|
||||
|
||||
def update_gradients_diag(self, dL_dKdiag, X):
|
||||
super(RatQuad, self).update_gradients_diag(dL_dKdiag, X)
|
||||
self.power.gradient = 0.
|
||||
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue