mirror of
https://github.com/SheffieldML/GPy.git
synced 2026-05-21 14:05:14 +02:00
Integral kernels added
This commit is contained in:
Michael T Smith
parent
67e1e1ae13
commit
5fa15037ca
3 changed files with 298 additions and 0 deletions
84
GPy/kern/src/integral.py
Normal file
84
GPy/kern/src/integral.py
Normal file
|
|
@ -0,0 +1,84 @@
|
|||
# Written by Mike Smith michaeltsmith.org.uk
|
||||
|
||||
import numpy as np
|
||||
from .kern import Kern
|
||||
from ...core.parameterization import Param
|
||||
from paramz.transformations import Logexp
|
||||
import math
|
||||
|
||||
class Integral(Kern): #todo do I need to inherit from Stationary
|
||||
"""
|
||||
Integral kernel between...
|
||||
"""
|
||||
|
||||
def __init__(self, input_dim, variances=None, lengthscale=None, ARD=False, active_dims=None, name='integral'):
|
||||
super(Integral, self).__init__(input_dim, active_dims, name)
|
||||
|
||||
if lengthscale is None:
|
||||
lengthscale = np.ones(1)
|
||||
else:
|
||||
lengthscale = np.asarray(lengthscale)
|
||||
|
||||
self.lengthscale = Param('lengthscale', lengthscale, Logexp()) #Logexp - transforms to allow positive only values...
|
||||
self.variances = Param('variances', variances, Logexp()) #and here.
|
||||
self.link_parameters(self.variances, self.lengthscale) #this just takes a list of parameters we need to optimise.
|
||||
|
||||
def h(self, z):
|
||||
return 0.5 * z * np.sqrt(math.pi) * math.erf(z) + np.exp(-(z**2))
|
||||
|
||||
def dk_dl(self, t, tprime, l): #derivative of the kernel wrt lengthscale
|
||||
return l * ( self.h(t/l) - self.h((t - tprime)/l) + self.h(tprime/l) - 1)
|
||||
|
||||
def update_gradients_full(self, dL_dK, X, X2=None):
|
||||
if X2 is None: #we're finding dK_xx/dTheta
|
||||
dK_dl = np.zeros([X.shape[0],X.shape[0]])
|
||||
dK_dv = np.zeros([X.shape[0],X.shape[0]])
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X):
|
||||
dK_dl[i,j] = self.variances[0]*self.dk_dl(x[0],x2[0],self.lengthscale[0]) #TODO Multiple length scales
|
||||
dK_dv[i,j] = self.k_xx(x[0],x2[0],self.lengthscale[0]) #the gradient wrt the variance is k_xx.
|
||||
self.lengthscale.gradient = np.sum(dK_dl * dL_dK)
|
||||
self.variances.gradient = np.sum(dK_dv * dL_dK)
|
||||
#print "V%0.5f" % self.variances.gradient
|
||||
#print "L%0.5f" % self.lengthscale.gradient
|
||||
else: #we're finding dK_xf/Dtheta
|
||||
print("NEED TO HANDLE TODO!")
|
||||
|
||||
#useful little function to help calculate the covariances.
|
||||
def g(self,z):
|
||||
return 1.0 * z * np.sqrt(math.pi) * math.erf(z) + np.exp(-(z**2))
|
||||
|
||||
#covariance between gradients (it's the gradients that we want out... maybe we should have a way of getting K_ff too? Currently you get the diag of K_ff from Kdiag)
|
||||
def k_xx(self,t,tprime,l):
|
||||
return 0.5 * (l**2) * ( self.g(t/l) - self.g((t - tprime)/l) + self.g(tprime/l) - 1)
|
||||
|
||||
def k_ff(self,t,tprime,l):
|
||||
return np.exp(-((t-tprime)**2)/(l**2)) #rbf
|
||||
|
||||
#covariance between the gradient and the actual value
|
||||
def k_xf(self,t,tprime,l):
|
||||
return 0.5 * np.sqrt(math.pi) * l * (math.erf((t-tprime)/l) + math.erf(tprime/l))
|
||||
|
||||
def K(self, X, X2=None):
|
||||
if X2 is None:
|
||||
K_xx = np.zeros([X.shape[0],X.shape[0]])
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X):
|
||||
K_xx[i,j] = self.k_xx(x[0],x2[0],self.lengthscale[0])
|
||||
return K_xx * self.variances[0]
|
||||
else:
|
||||
K_xf = np.zeros([X.shape[0],X2.shape[0]])
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X2):
|
||||
K_xf[i,j] = self.k_xf(x[0],x2[0],self.lengthscale[0])
|
||||
#print self.variances[0]
|
||||
return K_xf * self.variances[0]
|
||||
|
||||
def Kdiag(self, X):
|
||||
"""I've used the fact that we call this method for K_ff when finding the covariance as a hack so
|
||||
I know if I should return K_ff or K_xx. In this case we're returning K_ff!!
|
||||
$K_{ff}^{post} = K_{ff} - K_{fx} K_{xx}^{-1} K_{xf}$"""
|
||||
K_ff = np.zeros(X.shape[0])
|
||||
for i,x in enumerate(X):
|
||||
K_ff[i] = self.k_ff(x[0],x[0],self.lengthscale[0])
|
||||
return K_ff * self.variances[0]
|
||||
94
GPy/kern/src/integral_limits.py
Normal file
94
GPy/kern/src/integral_limits.py
Normal file
|
|
@ -0,0 +1,94 @@
|
|||
# Written by Mike Smith michaeltsmith.org.uk
|
||||
|
||||
import numpy as np
|
||||
from .kern import Kern
|
||||
from ...core.parameterization import Param
|
||||
from paramz.transformations import Logexp
|
||||
import math
|
||||
|
||||
class Integral_Limits(Kern): #todo do I need to inherit from Stationary
|
||||
"""
|
||||
Integral kernel, can include limits on each integral value.
|
||||
"""
|
||||
|
||||
def __init__(self, input_dim, variances=None, lengthscale=None, ARD=False, active_dims=None, name='integral'):
|
||||
super(Integral_Limits, self).__init__(input_dim, active_dims, name)
|
||||
|
||||
if lengthscale is None:
|
||||
lengthscale = np.ones(1)
|
||||
else:
|
||||
lengthscale = np.asarray(lengthscale)
|
||||
|
||||
self.lengthscale = Param('lengthscale', lengthscale, Logexp()) #Logexp - transforms to allow positive only values...
|
||||
self.variances = Param('variances', variances, Logexp()) #and here.
|
||||
self.link_parameters(self.variances, self.lengthscale) #this just takes a list of parameters we need to optimise.
|
||||
|
||||
def h(self, z):
|
||||
return 0.5 * z * np.sqrt(math.pi) * math.erf(z) + np.exp(-(z**2))
|
||||
|
||||
def dk_dl(self, t, tprime, s, sprime, l): #derivative of the kernel wrt lengthscale
|
||||
return l * ( self.h((t-sprime)/l) - self.h((t - tprime)/l) + self.h((tprime-s)/l) - self.h((s-sprime)/l))
|
||||
|
||||
def update_gradients_full(self, dL_dK, X, X2=None):
|
||||
if X2 is None: #we're finding dK_xx/dTheta
|
||||
dK_dl = np.zeros([X.shape[0],X.shape[0]])
|
||||
dK_dv = np.zeros([X.shape[0],X.shape[0]])
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X):
|
||||
dK_dl[i,j] = self.variances[0]*self.dk_dl(x[0],x2[0],x[1],x2[1],self.lengthscale[0])
|
||||
dK_dv[i,j] = self.k_xx(x[0],x2[0],x[1],x2[1],self.lengthscale[0]) #the gradient wrt the variance is k_xx.
|
||||
self.lengthscale.gradient = np.sum(dK_dl * dL_dK)
|
||||
self.variances.gradient = np.sum(dK_dv * dL_dK)
|
||||
#print "V%0.5f" % self.variances.gradient
|
||||
#print "L%0.5f" % self.lengthscale.gradient
|
||||
else: #we're finding dK_xf/Dtheta
|
||||
print("NEED TO HANDLE TODO!")
|
||||
|
||||
#useful little function to help calculate the covariances.
|
||||
def g(self,z):
|
||||
return 1.0 * z * np.sqrt(math.pi) * math.erf(z) + np.exp(-(z**2))
|
||||
|
||||
def k_xx(self,t,tprime,s,sprime,l):
|
||||
"""Covariance between observed values.
|
||||
|
||||
s and t are one domain of the integral (i.e. the integral between s and t)
|
||||
sprime and tprime are another domain of the integral (i.e. the integral between sprime and tprime)
|
||||
|
||||
We're interested in how correlated these two integrals are.
|
||||
|
||||
Note: We've not multiplied by the variance, this is done in K."""
|
||||
return 0.5 * (l**2) * ( self.g((t-sprime)/l) + self.g((tprime-s)/l) - self.g((t - tprime)/l) - self.g((s-sprime)/l))
|
||||
|
||||
def k_ff(self,t,tprime,l):
|
||||
"""Doesn't need s or sprime as we're looking at the 'derivatives', so no domains over which to integrate are required"""
|
||||
return np.exp(-((t-tprime)**2)/(l**2)) #rbf
|
||||
|
||||
def k_xf(self,t,tprime,s,l):
|
||||
"""Covariance between the gradient (latent value) and the actual (observed) value.
|
||||
|
||||
Note that sprime isn't actually used in this expression, presumably because the 'primes' are the gradient (latent) values which don't
|
||||
involve an integration, and thus there is no domain over which they're integrated, just a single value that we want."""
|
||||
return 0.5 * np.sqrt(math.pi) * l * (math.erf((t-tprime)/l) + math.erf((tprime-s)/l))
|
||||
|
||||
def K(self, X, X2=None):
|
||||
if X2 is None:
|
||||
K_xx = np.zeros([X.shape[0],X.shape[0]])
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X):
|
||||
K_xx[i,j] = self.k_xx(x[0],x2[0],x[1],x2[1],self.lengthscale[0])
|
||||
return K_xx * self.variances[0]
|
||||
else:
|
||||
K_xf = np.zeros([X.shape[0],X2.shape[0]])
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X2):
|
||||
K_xf[i,j] = self.k_xf(x[0],x2[0],x[1],self.lengthscale[0]) #x2[1] unused, see k_xf docstring for explanation.
|
||||
return K_xf * self.variances[0]
|
||||
|
||||
def Kdiag(self, X):
|
||||
"""I've used the fact that we call this method for K_ff when finding the covariance as a hack so
|
||||
I know if I should return K_ff or K_xx. In this case we're returning K_ff!!
|
||||
$K_{ff}^{post} = K_{ff} - K_{fx} K_{xx}^{-1} K_{xf}$"""
|
||||
K_ff = np.zeros(X.shape[0])
|
||||
for i,x in enumerate(X):
|
||||
K_ff[i] = self.k_ff(x[0],x[0],self.lengthscale[0])
|
||||
return K_ff * self.variances[0]
|
||||
120
GPy/kern/src/multidimensional_integral_limits.py
Normal file
120
GPy/kern/src/multidimensional_integral_limits.py
Normal file
|
|
@ -0,0 +1,120 @@
|
|||
# Written by Mike Smith michaeltsmith.org.uk
|
||||
|
||||
import numpy as np
|
||||
from .kern import Kern
|
||||
from ...core.parameterization import Param
|
||||
from paramz.transformations import Logexp
|
||||
import math
|
||||
|
||||
class Multidimensional_Integral_Limits(Kern): #todo do I need to inherit from Stationary
|
||||
"""
|
||||
Integral kernel, can include limits on each integral value.
|
||||
"""
|
||||
|
||||
def __init__(self, input_dim, variances=None, lengthscale=None, ARD=False, active_dims=None, name='integral'):
|
||||
super(Multidimensional_Integral_Limits, self).__init__(input_dim, active_dims, name)
|
||||
|
||||
if lengthscale is None:
|
||||
lengthscale = np.ones(1)
|
||||
else:
|
||||
lengthscale = np.asarray(lengthscale)
|
||||
|
||||
self.lengthscale = Param('lengthscale', lengthscale, Logexp()) #Logexp - transforms to allow positive only values...
|
||||
self.variances = Param('variances', variances, Logexp()) #and here.
|
||||
self.link_parameters(self.variances, self.lengthscale) #this just takes a list of parameters we need to optimise.
|
||||
|
||||
def h(self, z):
|
||||
return 0.5 * z * np.sqrt(math.pi) * math.erf(z) + np.exp(-(z**2))
|
||||
|
||||
def dk_dl(self, t, tprime, s, sprime, l): #derivative of the kernel wrt lengthscale
|
||||
return l * ( self.h((t-sprime)/l) - self.h((t - tprime)/l) + self.h((tprime-s)/l) - self.h((s-sprime)/l))
|
||||
|
||||
def update_gradients_full(self, dL_dK, X, X2=None):
|
||||
#print self.variances
|
||||
if X2 is None: #we're finding dK_xx/dTheta
|
||||
dK_dl_term = np.zeros([X.shape[0],X.shape[0],self.lengthscale.shape[0]])
|
||||
k_term = np.zeros([X.shape[0],X.shape[0],self.lengthscale.shape[0]])
|
||||
dK_dl = np.zeros([X.shape[0],X.shape[0],self.lengthscale.shape[0]])
|
||||
dK_dv = np.zeros([X.shape[0],X.shape[0]])
|
||||
for il,l in enumerate(self.lengthscale):
|
||||
idx = il*2
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X):
|
||||
dK_dl_term[i,j,il] = self.dk_dl(x[idx],x2[idx],x[idx+1],x2[idx+1],l)
|
||||
k_term[i,j,il] = self.k_xx(x[idx],x2[idx],x[idx+1],x2[idx+1],l)
|
||||
for il,l in enumerate(self.lengthscale):
|
||||
dK_dl = self.variances[0] * dK_dl_term[:,:,il]
|
||||
for jl, l in enumerate(self.lengthscale):
|
||||
if jl!=il:
|
||||
dK_dl *= k_term[:,:,jl]
|
||||
#dK_dl = np.dot(dK_dl,k_term[:,:,il])
|
||||
#print k_term[:,:,il]
|
||||
self.lengthscale.gradient[il] = np.sum(dK_dl * dL_dK)
|
||||
dK_dv = self.calc_K_xx_wo_variance(X) #the gradient wrt the variance is k_xx.
|
||||
self.variances.gradient = np.sum(dK_dv * dL_dK)
|
||||
else: #we're finding dK_xf/Dtheta
|
||||
print("NEED TO HANDLE TODO!")
|
||||
#print self.variances[0],self.lengthscale[0],self.lengthscale[1] #np.sum(dK_dv*dL_dK)
|
||||
|
||||
|
||||
#useful little function to help calculate the covariances.
|
||||
def g(self,z):
|
||||
return 1.0 * z * np.sqrt(math.pi) * math.erf(z) + np.exp(-(z**2))
|
||||
|
||||
def k_xx(self,t,tprime,s,sprime,l):
|
||||
"""Covariance between observed values.
|
||||
|
||||
s and t are one domain of the integral (i.e. the integral between s and t)
|
||||
sprime and tprime are another domain of the integral (i.e. the integral between sprime and tprime)
|
||||
|
||||
We're interested in how correlated these two integrals are.
|
||||
|
||||
Note: We've not multiplied by the variance, this is done in K."""
|
||||
return 0.5 * (l**2) * ( self.g((t-sprime)/l) + self.g((tprime-s)/l) - self.g((t - tprime)/l) - self.g((s-sprime)/l))
|
||||
|
||||
def k_ff(self,t,tprime,l):
|
||||
"""Doesn't need s or sprime as we're looking at the 'derivatives', so no domains over which to integrate are required"""
|
||||
return np.exp(-((t-tprime)**2)/(l**2)) #rbf
|
||||
|
||||
def k_xf(self,t,tprime,s,l):
|
||||
"""Covariance between the gradient (latent value) and the actual (observed) value.
|
||||
|
||||
Note that sprime isn't actually used in this expression, presumably because the 'primes' are the gradient (latent) values which don't
|
||||
involve an integration, and thus there is no domain over which they're integrated, just a single value that we want."""
|
||||
return 0.5 * np.sqrt(math.pi) * l * (math.erf((t-tprime)/l) + math.erf((tprime-s)/l))
|
||||
|
||||
def calc_K_xx_wo_variance(self,X):
|
||||
"""Calculates K_xx without the variance term"""
|
||||
K_xx = np.ones([X.shape[0],X.shape[0]]) #ones now as a product occurs over each dimension
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X):
|
||||
for il,l in enumerate(self.lengthscale):
|
||||
idx = il*2 #each pair of input dimensions describe the limits on one actual dimension in the data
|
||||
K_xx[i,j] *= self.k_xx(x[idx],x2[idx],x[idx+1],x2[idx+1],l)
|
||||
return K_xx
|
||||
|
||||
def K(self, X, X2=None):
|
||||
if X2 is None:
|
||||
#print "X x X"
|
||||
K_xx = self.calc_K_xx_wo_variance(X)
|
||||
return K_xx * self.variances[0]
|
||||
else:
|
||||
#print "X x X2"
|
||||
K_xf = np.ones([X.shape[0],X2.shape[0]])
|
||||
for i,x in enumerate(X):
|
||||
for j,x2 in enumerate(X2):
|
||||
for il,l in enumerate(self.lengthscale):
|
||||
idx = il*2
|
||||
K_xf[i,j] *= self.k_xf(x[idx],x2[idx],x[idx+1],l)
|
||||
return K_xf * self.variances[0]
|
||||
|
||||
def Kdiag(self, X):
|
||||
"""I've used the fact that we call this method for K_ff when finding the covariance as a hack so
|
||||
I know if I should return K_ff or K_xx. In this case we're returning K_ff!!
|
||||
$K_{ff}^{post} = K_{ff} - K_{fx} K_{xx}^{-1} K_{xf}$"""
|
||||
K_ff = np.ones(X.shape[0])
|
||||
for i,x in enumerate(X):
|
||||
for il,l in enumerate(self.lengthscale):
|
||||
idx = il*2
|
||||
K_ff[i] *= self.k_ff(x[idx],x[idx],l)
|
||||
return K_ff * self.variances[0]
|
||||
Loading…
Add table
Add a link
Reference in a new issue