apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-24 14:15:14 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

linear kern variational updates

Browse source
This commit is contained in:
Max Zwiessele 2014-02-11 12:17:07 +00:00
parent ed87e6bbd2
commit 9e1546524e
1 changed files with 26 additions and 58 deletions
Download patch file Download diff file Expand all files Collapse all files

84
GPy/kern/parts/linear.py
View file

@ -8,6 +8,7 @@ from kernpart import Kernpart
from ...util.linalg import tdot from ...util.linalg import tdot
from ...util.misc import fast_array_equal, param_to_array from ...util.misc import fast_array_equal, param_to_array
from ...core.parameterization import Param from ...core.parameterization import Param
from ...core.parameterization.transformations import Logexp
class Linear(Kernpart): class Linear(Kernpart):
""" """
@ -43,8 +44,9 @@ class Linear(Kernpart):
else: else:
variances = np.ones(self.input_dim) variances = np.ones(self.input_dim)
self.variances = Param('variances', variances) self.variances = Param('variances', variances, Logexp())
self.add_parameters(self.variances) self.variances.gradient = np.zeros(self.variances.shape)
self.add_parameter(self.variances)
self.variances.add_observer(self, self.update_variance) self.variances.add_observer(self, self.update_variance)
# initialize cache # initialize cache
@ -57,42 +59,35 @@ class Linear(Kernpart):
def on_input_change(self, X): def on_input_change(self, X):
self._K_computations(X, None) self._K_computations(X, None)
def update_gradients_full(self, dL_dK, X):
#self.variances.gradient[:] = 0
self._param_grad_helper(dL_dK, X, self.variances.gradient)
def update_gradients_sparse(self, dL_dKmm, dL_dKnm, dL_dKdiag, X, Z):
tmp = dL_dKdiag[:, None] * X ** 2
if self.ARD:
self.variances.gradient = tmp.sum(0)
else:
self.variances.gradient = tmp.sum()
self._param_grad_helper(dL_dKmm, Z, None, self.variances.gradient)
self._param_grad_helper(dL_dKnm, X, Z, self.variances.gradient)
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, mu, S, Z): def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, mu, S, Z):
self._psi_computations(Z, mu, S) self._psi_computations(Z, mu, S)
# psi0: # psi0:
tmp = dL_dpsi0[:, None] * self.mu2_S tmp = dL_dpsi0[:, None] * self.mu2_S
if self.ARD: self.variances.gradient = tmp.sum(0) if self.ARD: self.variances.gradient[:] = tmp.sum(0)
else: self.variances.gradient = tmp.sum() else: self.variances.gradient[:] = tmp.sum()
#psi1 #psi1
self.dK_dtheta(dL_dpsi1, mu, Z, self.variances.gradient) self._param_grad_helper(dL_dpsi1, mu, Z, self.variances.gradient)
#psi2
#from psi2
tmp = dL_dpsi2[:, :, :, None] * (self.ZAinner[:, :, None, :] * (2 * Z)[None, None, :, :]) tmp = dL_dpsi2[:, :, :, None] * (self.ZAinner[:, :, None, :] * (2 * Z)[None, None, :, :])
if self.ARD: self.variances.gradient += tmp.sum(0).sum(0).sum(0) if self.ARD: self.variances.gradient += tmp.sum(0).sum(0).sum(0)
else: self.variances.gradient += tmp.sum() else: self.variances.gradient += tmp.sum()
#from Kmm #from Kmm
self._K_computations(Z, None) self._K_computations(Z, None)
self.dK_dtheta(dL_dKmm, Z, None, self.variances.gradient) self._param_grad_helper(dL_dKmm, Z, None, self.variances.gradient)
# def _get_params(self):
# return self.variances
#
# def _set_params(self, x):
# assert x.size == (self.num_params)
# self.variances = x
#def parameters_changed(self):
# self.variances2 = np.square(self.variances)
#
# def _get_param_names(self):
# if self.num_params == 1:
# return ['variance']
# else:
# return ['variance_%i' % i for i in range(self.variances.size)]
def K(self, X, X2, target): def K(self, X, X2, target):
if self.ARD: if self.ARD:
XX = X * np.sqrt(self.variances) XX = X * np.sqrt(self.variances)
@ -109,7 +104,7 @@ class Linear(Kernpart):
def Kdiag(self, X, target): def Kdiag(self, X, target):
np.add(target, np.sum(self.variances * np.square(X), -1), target) np.add(target, np.sum(self.variances * np.square(X), -1), target)
def dK_dtheta(self, dL_dK, X, X2, target): def _param_grad_helper(self, dL_dK, X, X2, target):
if self.ARD: if self.ARD:
if X2 is None: if X2 is None:
[np.add(target[i:i + 1], np.sum(dL_dK * tdot(X[:, i:i + 1])), target[i:i + 1]) for i in range(self.input_dim)] [np.add(target[i:i + 1], np.sum(dL_dK * tdot(X[:, i:i + 1])), target[i:i + 1]) for i in range(self.input_dim)]
@ -121,13 +116,6 @@ class Linear(Kernpart):
self._K_computations(X, X2) self._K_computations(X, X2)
target += np.sum(self._dot_product * dL_dK) target += np.sum(self._dot_product * dL_dK)
def dKdiag_dtheta(self, dL_dKdiag, X, target):
tmp = dL_dKdiag[:, None] * X ** 2
if self.ARD:
target += tmp.sum(0)
else:
target += tmp.sum()
def gradients_X(self, dL_dK, X, X2, target): def gradients_X(self, dL_dK, X, X2, target):
if X2 is None: if X2 is None:
target += 2*(((X[None,:, :] * self.variances)) * dL_dK[:, :, None]).sum(1) target += 2*(((X[None,:, :] * self.variances)) * dL_dK[:, :, None]).sum(1)
@ -145,14 +133,6 @@ class Linear(Kernpart):
self._psi_computations(Z, mu, S) self._psi_computations(Z, mu, S)
target += np.sum(self.variances * self.mu2_S, 1) target += np.sum(self.variances * self.mu2_S, 1)
def dpsi0_dtheta(self, dL_dpsi0, Z, mu, S, target):
self._psi_computations(Z, mu, S)
tmp = dL_dpsi0[:, None] * self.mu2_S
if self.ARD:
target += tmp.sum(0)
else:
target += tmp.sum()
def dpsi0_dmuS(self, dL_dpsi0, Z, mu, S, target_mu, target_S): def dpsi0_dmuS(self, dL_dpsi0, Z, mu, S, target_mu, target_S):
target_mu += dL_dpsi0[:, None] * (2.0 * mu * self.variances) target_mu += dL_dpsi0[:, None] * (2.0 * mu * self.variances)
target_S += dL_dpsi0[:, None] * self.variances target_S += dL_dpsi0[:, None] * self.variances
@ -161,10 +141,6 @@ class Linear(Kernpart):
"""the variance, it does nothing""" """the variance, it does nothing"""
self._psi1 = self.K(mu, Z, target) self._psi1 = self.K(mu, Z, target)
def dpsi1_dtheta(self, dL_dpsi1, Z, mu, S, target):
"""the variance, it does nothing"""
self.dK_dtheta(dL_dpsi1, mu, Z, target)
def dpsi1_dmuS(self, dL_dpsi1, Z, mu, S, target_mu, target_S): def dpsi1_dmuS(self, dL_dpsi1, Z, mu, S, target_mu, target_S):
"""Do nothing for S, it does not affect psi1""" """Do nothing for S, it does not affect psi1"""
self._psi_computations(Z, mu, S) self._psi_computations(Z, mu, S)
@ -185,21 +161,13 @@ class Linear(Kernpart):
def dpsi2_dtheta_new(self, dL_dpsi2, Z, mu, S, target): def dpsi2_dtheta_new(self, dL_dpsi2, Z, mu, S, target):
tmp = np.zeros((mu.shape[0], Z.shape[0])) tmp = np.zeros((mu.shape[0], Z.shape[0]))
self.K(mu,Z,tmp) self.K(mu,Z,tmp)
self.dK_dtheta(2.*np.sum(dL_dpsi2*tmp[:,None,:],2),mu,Z,target) self._param_grad_helper(2.*np.sum(dL_dpsi2*tmp[:,None,:],2),mu,Z,target)
result= 2.*(dL_dpsi2[:,:,:,None]*S[:,None,None,:]*self.variances*Z[None,:,None,:]*Z[None,None,:,:]).sum(0).sum(0).sum(0) result= 2.*(dL_dpsi2[:,:,:,None]*S[:,None,None,:]*self.variances*Z[None,:,None,:]*Z[None,None,:,:]).sum(0).sum(0).sum(0)
if self.ARD: if self.ARD:
target += result.sum(0).sum(0).sum(0) target += result.sum(0).sum(0).sum(0)
else: else:
target += result.sum() target += result.sum()
def dpsi2_dtheta(self, dL_dpsi2, Z, mu, S, target):
self._psi_computations(Z, mu, S)
tmp = dL_dpsi2[:, :, :, None] * (self.ZAinner[:, :, None, :] * (2 * Z)[None, None, :, :])
if self.ARD:
target += tmp.sum(0).sum(0).sum(0)
else:
target += tmp.sum()
def dpsi2_dmuS_new(self, dL_dpsi2, Z, mu, S, target_mu, target_S): def dpsi2_dmuS_new(self, dL_dpsi2, Z, mu, S, target_mu, target_S):
tmp = np.zeros((mu.shape[0], Z.shape[0])) tmp = np.zeros((mu.shape[0], Z.shape[0]))
self.K(mu,Z,tmp) self.K(mu,Z,tmp)
@ -309,11 +277,11 @@ class Linear(Kernpart):
if not (fast_array_equal(X, self._X) and fast_array_equal(X2, self._X2)): if not (fast_array_equal(X, self._X) and fast_array_equal(X2, self._X2)):
self._X = X.copy() self._X = X.copy()
if X2 is None: if X2 is None:
self._dot_product = tdot(X) self._dot_product = tdot(param_to_array(X))
self._X2 = None self._X2 = None
else: else:
self._X2 = X2.copy() self._X2 = X2.copy()
self._dot_product = np.dot(X, X2.T) self._dot_product = np.dot(param_to_array(X), param_to_array(X2.T))
def _psi_computations(self, Z, mu, S): def _psi_computations(self, Z, mu, S):
# here are the "statistics" for psi1 and psi2 # here are the "statistics" for psi1 and psi2