mirror of
https://github.com/SheffieldML/GPy.git
synced 2026-05-21 14:05:14 +02:00
Merge branch 'params' of https://github.com/SheffieldML/GPy into params
This commit is contained in:
Neil Lawrence
commit
86f92869a1
13 changed files with 235 additions and 398 deletions
|
|
@ -1,12 +1,9 @@
|
|||
# Copyright (c) 2012, GPy authors (see AUTHORS.txt).
|
||||
# Licensed under the BSD 3-clause license (see LICENSE.txt)
|
||||
|
||||
import sys
|
||||
import numpy as np
|
||||
import itertools
|
||||
from linear import Linear
|
||||
from ...core.parameterization import Parameterized
|
||||
from ...core.parameterization.param import Param
|
||||
from kern import Kern
|
||||
|
||||
class Add(Kern):
|
||||
|
|
@ -42,8 +39,14 @@ class Add(Kern):
|
|||
else:
|
||||
return sum([p.K(X[:, i_s], X2[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)])
|
||||
|
||||
def update_gradients_full(self, dL_dK, X):
|
||||
[p.update_gradients_full(dL_dK, X[:,i_s]) for p, i_s in zip(self._parameters_, self.input_slices)]
|
||||
def update_gradients_full(self, dL_dK, X, X2=None):
|
||||
if X2 is None:
|
||||
[p.update_gradients_full(dL_dK, X[:,i_s], X2) for p, i_s in zip(self._parameters_, self.input_slices)]
|
||||
else:
|
||||
[p.update_gradients_full(dL_dK, X[:,i_s], X2[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)]
|
||||
|
||||
def update_gradients_diag(self, dL_dKdiag, X):
|
||||
[p.update_gradients_diag(dL_dKdiag, X[:,i_s]) for p, i_s in zip(self._parameters_, self.input_slices)]
|
||||
|
||||
def gradients_X(self, dL_dK, X, X2=None):
|
||||
"""Compute the gradient of the objective function with respect to X.
|
||||
|
|
@ -68,19 +71,18 @@ class Add(Kern):
|
|||
|
||||
|
||||
def psi0(self, Z, variational_posterior):
|
||||
return np.sum([p.psi0(Z[:, i_s], mu[:, i_s], S[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)],0)
|
||||
return np.sum([p.psi0(Z[:, i_s], variational_posterior[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)],0)
|
||||
|
||||
def psi1(self, Z, variational_posterior):
|
||||
return np.sum([p.psi1(Z[:, i_s], mu[:, i_s], S[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)], 0)
|
||||
return np.sum([p.psi1(Z[:, i_s], variational_posterior[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)], 0)
|
||||
|
||||
def psi2(self, Z, variational_posterior):
|
||||
psi2 = np.sum([p.psi2(Z[:, i_s], mu[:, i_s], S[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)], 0)
|
||||
psi2 = np.sum([p.psi2(Z[:, i_s], variational_posterior[:, i_s]) for p, i_s in zip(self._parameters_, self.input_slices)], 0)
|
||||
|
||||
# compute the "cross" terms
|
||||
from white import White
|
||||
from static import White, Bias
|
||||
from rbf import RBF
|
||||
#from rbf_inv import RBFInv
|
||||
from bias import Bias
|
||||
from linear import Linear
|
||||
#ffrom fixed import Fixed
|
||||
|
||||
|
|
@ -91,24 +93,20 @@ class Add(Kern):
|
|||
# rbf X bias
|
||||
#elif isinstance(p1, (Bias, Fixed)) and isinstance(p2, (RBF, RBFInv)):
|
||||
elif isinstance(p1, Bias) and isinstance(p2, (RBF, Linear)):
|
||||
tmp = p2.psi1(Z[:,i2], mu[:,i2], S[:,i2])
|
||||
tmp = p2.psi1(Z[:,i2], variational_posterior[:, i_s])
|
||||
psi2 += p1.variance * (tmp[:, :, None] + tmp[:, None, :])
|
||||
#elif isinstance(p2, (Bias, Fixed)) and isinstance(p1, (RBF, RBFInv)):
|
||||
elif isinstance(p2, Bias) and isinstance(p1, (RBF, Linear)):
|
||||
tmp = p1.psi1(Z[:,i1], mu[:,i1], S[:,i1])
|
||||
tmp = p1.psi1(Z[:,i1], variational_posterior[:, i_s])
|
||||
psi2 += p2.variance * (tmp[:, :, None] + tmp[:, None, :])
|
||||
else:
|
||||
raise NotImplementedError, "psi2 cannot be computed for this kernel"
|
||||
return psi2
|
||||
|
||||
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
|
||||
#from bias import Bias
|
||||
from linear import Linear
|
||||
#ffrom fixed import Fixed
|
||||
|
||||
from static import White, Bias
|
||||
mu, S = variational_posterior.mean, variational_posterior.variance
|
||||
|
||||
for p1, is1 in zip(self._parameters_, self.input_slices):
|
||||
|
||||
#compute the effective dL_dpsi1. Extra terms appear becaue of the cross terms in psi2!
|
||||
|
|
@ -121,20 +119,15 @@ class Add(Kern):
|
|||
elif isinstance(p2, Bias):
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
|
||||
else:
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z[:,is2], mu[:,is2], S[:,is2]) * 2.
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z[:,is2], variational_posterior[:, is1]) * 2.
|
||||
|
||||
|
||||
p1.update_gradients_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], S[:,is1], Z[:,is1])
|
||||
p1.update_gradients_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, Z[:,is1], variational_posterior[:, is1])
|
||||
|
||||
|
||||
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
|
||||
from bias import Bias
|
||||
from linear import Linear
|
||||
#ffrom fixed import fixed
|
||||
|
||||
from static import White, Bias
|
||||
|
||||
target = np.zeros(Z.shape)
|
||||
for p1, is1 in zip(self._parameters_, self.input_slices):
|
||||
|
||||
|
|
@ -148,22 +141,17 @@ class Add(Kern):
|
|||
elif isinstance(p2, Bias):
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
|
||||
else:
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z[:,is2], mu[:,is2], S[:,is2]) * 2.
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z[:,is2], variational_posterior[:, is2]) * 2.
|
||||
|
||||
|
||||
target += p1.gradients_z_variational(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, mu[:,is1], S[:,is1], Z[:,is1])
|
||||
target += p1.gradients_Z_expectations(eff_dL_dpsi1, dL_dpsi2, Z[:,is1], variational_posterior[:, is1])
|
||||
return target
|
||||
|
||||
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
|
||||
#from bias import bias
|
||||
from linear import linear
|
||||
#ffrom fixed import fixed
|
||||
|
||||
target_mu = np.zeros(mu.shape)
|
||||
target_S = np.zeros(S.shape)
|
||||
from static import White, Bias
|
||||
|
||||
target_mu = np.zeros(variational_posterior.shape)
|
||||
target_S = np.zeros(variational_posterior.shape)
|
||||
for p1, is1 in zip(self._parameters_, self.input_slices):
|
||||
|
||||
#compute the effective dL_dpsi1. extra terms appear becaue of the cross terms in psi2!
|
||||
|
|
@ -171,15 +159,15 @@ class Add(Kern):
|
|||
for p2, is2 in zip(self._parameters_, self.input_slices):
|
||||
if p2 is p1:
|
||||
continue
|
||||
if isinstance(p2, white):
|
||||
if isinstance(p2, White):
|
||||
continue
|
||||
elif isinstance(p2, bias):
|
||||
elif isinstance(p2, Bias):
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
|
||||
else:
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(z[:,is2], mu[:,is2], s[:,is2]) * 2.
|
||||
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z[:,is2], variational_posterior[:, is2]) * 2.
|
||||
|
||||
|
||||
a, b = p1.gradients_qX_expectations(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, Z[:,is1], variational_posterior[:, is1])
|
||||
target_mu += a
|
||||
target_S += b
|
||||
return target_mu, target_S
|
||||
|
|
|
|||
|
|
@ -37,6 +37,10 @@ class Kern(Parameterized):
|
|||
def gradients_X_diag(self, dL_dK, X):
|
||||
raise NotImplementedError
|
||||
|
||||
def update_gradients_diag(self, dL_dKdiag, X):
|
||||
""" update the gradients of all parameters when using only the diagonal elements of the covariance matrix"""
|
||||
raise NotImplementedError
|
||||
|
||||
def update_gradients_full(self, dL_dK, X, X2):
|
||||
"""Set the gradients of all parameters when doing full (N) inference."""
|
||||
raise NotImplementedError
|
||||
|
|
@ -89,7 +93,7 @@ class Kern(Parameterized):
|
|||
"""
|
||||
Returns the sensitivity for each dimension of this kernel.
|
||||
"""
|
||||
return self.kern.input_sensitivity()
|
||||
return np.zeros(self.input_dim)
|
||||
|
||||
def __add__(self, other):
|
||||
""" Overloading of the '+' operator. for more control, see self.add """
|
||||
|
|
|
|||
|
|
@ -55,7 +55,7 @@ class White(Static):
|
|||
def psi2(self, Z, variational_posterior):
|
||||
return np.zeros((variational_posterior.shape[0], Z.shape[0], Z.shape[0]), dtype=np.float64)
|
||||
|
||||
def update_gradients_full(self, dL_dK, X):
|
||||
def update_gradients_full(self, dL_dK, X, X2=None):
|
||||
self.variance.gradient = np.trace(dL_dK)
|
||||
|
||||
def update_gradients_diag(self, dL_dKdiag, X):
|
||||
|
|
@ -79,10 +79,10 @@ class Bias(Static):
|
|||
self.variance.gradient = dL_dK.sum()
|
||||
|
||||
def update_gradients_diag(self, dL_dKdiag, X):
|
||||
self.variance.gradient = dL_dK.sum()
|
||||
self.variance.gradient = dL_dKdiag.sum()
|
||||
|
||||
def psi2(self, Z, variational_posterior):
|
||||
ret = np.empty((mu.shape[0], Z.shape[0], Z.shape[0]), dtype=np.float64)
|
||||
ret = np.empty((variational_posterior.shape[0], Z.shape[0], Z.shape[0]), dtype=np.float64)
|
||||
ret[:] = self.variance**2
|
||||
return ret
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue