apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-14 06:22:38 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

merged with master

Browse source
This commit is contained in:
Nicolo Fusi 2013-02-15 11:56:37 +00:00
commit 2abba8cf14
59 changed files with 2845 additions and 1608 deletions
Download patch file Download diff file Expand all files Collapse all files

19
GPy/kern/Matern32.py
View file

@ -14,14 +14,14 @@ class Matern32(kernpart):
.. math::
k(r) = \sigma^2 (1 + \sqrt{3} r) \exp(- \sqrt{3} r) \qquad \qquad \\text{ where } r = \sqrt{\sum_{i=1}^D \\frac{(x_i-y_i)^2}{\ell_i^2} }
k(r) = \\sigma^2 (1 + \\sqrt{3} r) \exp(- \sqrt{3} r) \\ \\ \\ \\ \\text{ where } r = \sqrt{\sum_{i=1}^D \\frac{(x_i-y_i)^2}{\ell_i^2} }
:param D: the number of input dimensions
:type D: int
:param variance: the variance :math:`\sigma^2`
:type variance: float
:param lengthscale: the vector of lengthscale :math:`\ell_i`
:type lengthscale: np.ndarray of size (1,) or (D,) depending on ARD
:type lengthscale: array or list of the appropriate size (or float if there is only one lengthscale parameter)
:param ARD: Auto Relevance Determination. If equal to "False", the kernel is isotropic (ie. one single lengthscale parameter \ell), otherwise there is one lengthscale parameter per dimension.
:type ARD: Boolean
:rtype: kernel object
@ -35,17 +35,19 @@ class Matern32(kernpart):
self.Nparam = 2
self.name = 'Mat32'
if lengthscale is not None:
assert lengthscale.shape == (1,)
lengthscale = np.asarray(lengthscale)
assert lengthscale.size == 1, "Only one lengthscale needed for non-ARD kernel"
else:
lengthscale = np.ones(1)
else:
self.Nparam = self.D + 1
self.name = 'Mat32_ARD'
self.name = 'Mat32'
if lengthscale is not None:
assert lengthscale.shape == (self.D,)
lengthscale = np.asarray(lengthscale)
assert lengthscale.size == self.D, "bad number of lengthscales"
else:
lengthscale = np.ones(self.D)
self._set_params(np.hstack((variance,lengthscale)))
self._set_params(np.hstack((variance,lengthscale.flatten())))
def _get_params(self):
"""return the value of the parameters."""
@ -116,9 +118,9 @@ class Matern32(kernpart):
:param F1: vector of derivatives of F
:type F1: np.array
:param F2: vector of second derivatives of F
:type F2: np.array
:type F2: np.array
:param lower,upper: boundaries of the input domain
:type lower,upper: floats
:type lower,upper: floats
"""
assert self.D == 1
def L(x,i):
@ -133,4 +135,3 @@ class Matern32(kernpart):
#print "OLD \n", np.dot(F1lower,F1lower.T), "\n \n"
#return(G)
return(self.lengthscale**3/(12.*np.sqrt(3)*self.variance) * G + 1./self.variance*np.dot(Flower,Flower.T) + self.lengthscale**2/(3.*self.variance)*np.dot(F1lower,F1lower.T))

14
GPy/kern/Matern52.py
View file

@ -13,14 +13,14 @@ class Matern52(kernpart):
.. math::
k(r) = \sigma^2 (1 + \sqrt{5} r + \\frac53 r^2) \exp(- \sqrt{5} r) \qquad \qquad \\text{ where } r = \sqrt{\sum_{i=1}^D \\frac{(x_i-y_i)^2}{\ell_i^2} }
k(r) = \sigma^2 (1 + \sqrt{5} r + \\frac53 r^2) \exp(- \sqrt{5} r) \ \ \ \ \ \\text{ where } r = \sqrt{\sum_{i=1}^D \\frac{(x_i-y_i)^2}{\ell_i^2} }
:param D: the number of input dimensions
:type D: int
:param variance: the variance :math:`\sigma^2`
:type variance: float
:param lengthscale: the vector of lengthscale :math:`\ell_i`
:type lengthscale: np.ndarray of size (1,) or (D,) depending on ARD
:type lengthscale: array or list of the appropriate size (or float if there is only one lengthscale parameter)
:param ARD: Auto Relevance Determination. If equal to "False", the kernel is isotropic (ie. one single lengthscale parameter \ell), otherwise there is one lengthscale parameter per dimension.
:type ARD: Boolean
:rtype: kernel object
@ -33,17 +33,19 @@ class Matern52(kernpart):
self.Nparam = 2
self.name = 'Mat52'
if lengthscale is not None:
assert lengthscale.shape == (1,)
lengthscale = np.asarray(lengthscale)
assert lengthscale.size == 1, "Only one lengthscale needed for non-ARD kernel"
else:
lengthscale = np.ones(1)
else:
self.Nparam = self.D + 1
self.name = 'Mat52_ARD'
self.name = 'Mat52'
if lengthscale is not None:
assert lengthscale.shape == (self.D,)
lengthscale = np.asarray(lengthscale)
assert lengthscale.size == self.D, "bad number of lengthscales"
else:
lengthscale = np.ones(self.D)
self._set_params(np.hstack((variance,lengthscale)))
self._set_params(np.hstack((variance,lengthscale.flatten())))
def _get_params(self):
"""return the value of the parameters."""

25
GPy/kern/exponential.py
View file

@ -13,14 +13,14 @@ class exponential(kernpart):
.. math::
k(r) = \sigma^2 \exp(- r) \qquad \qquad \\text{ where } r = \sqrt{\sum_{i=1}^D \\frac{(x_i-y_i)^2}{\ell_i^2} }
k(r) = \sigma^2 \exp(- r) \ \ \ \ \ \\text{ where } r = \sqrt{\sum_{i=1}^D \\frac{(x_i-y_i)^2}{\ell_i^2} }
:param D: the number of input dimensions
:type D: int
:param variance: the variance :math:`\sigma^2`
:type variance: float
:param lengthscale: the vector of lengthscale :math:`\ell_i`
:type lengthscale: np.ndarray of size (1,) or (D,) depending on ARD
:type lengthscale: array or list of the appropriate size (or float if there is only one lengthscale parameter)
:param ARD: Auto Relevance Determination. If equal to "False", the kernel is isotropic (ie. one single lengthscale parameter \ell), otherwise there is one lengthscale parameter per dimension.
:type ARD: Boolean
:rtype: kernel object
@ -33,17 +33,19 @@ class exponential(kernpart):
self.Nparam = 2
self.name = 'exp'
if lengthscale is not None:
assert lengthscale.shape == (1,)
lengthscale = np.asarray(lengthscale)
assert lengthscale.size == 1, "Only one lengthscale needed for non-ARD kernel"
else:
lengthscale = np.ones(1)
else:
self.Nparam = self.D + 1
self.name = 'exp_ARD'
self.name = 'exp'
if lengthscale is not None:
assert lengthscale.shape == (self.D,)
lengthscale = np.asarray(lengthscale)
assert lengthscale.size == self.D, "bad number of lengthscales"
else:
lengthscale = np.ones(self.D)
self._set_params(np.hstack((variance,lengthscale)))
self._set_params(np.hstack((variance,lengthscale.flatten())))
def _get_params(self):
"""return the value of the parameters."""
@ -87,7 +89,7 @@ class exponential(kernpart):
dl = self.variance*dvar*dist2M.sum(-1)*invdist
target[1] += np.sum(dl*partial)
def dKdiag_dtheta(self,partial,X,target):
def dKdiag_dtheta(self,partial,X,target):
"""derivative of the diagonal of the covariance matrix with respect to the parameters."""
#NB: derivative of diagonal elements wrt lengthscale is 0
target[0] += np.sum(partial)
@ -110,9 +112,9 @@ class exponential(kernpart):
:param F: vector of functions
:type F: np.array
:param F1: vector of derivatives of F
:type F1: np.array
:type F1: np.array
:param lower,upper: boundaries of the input domain
:type lower,upper: floats
:type lower,upper: floats
"""
assert self.D == 1
def L(x,i):
@ -124,8 +126,3 @@ class exponential(kernpart):
G[i,j] = G[j,i] = integrate.quad(lambda x : L(x,i)*L(x,j),lower,upper)[0]
Flower = np.array([f(lower) for f in F])[:,None]
return(self.lengthscale/2./self.variance * G + 1./self.variance * np.dot(Flower,Flower.T))

7
GPy/kern/kern.py
View file

@ -6,7 +6,7 @@ import numpy as np
from ..core.parameterised import parameterised
from kernpart import kernpart
import itertools
from product_orthogonal import product_orthogonal
from product_orthogonal import product_orthogonal
class kern(parameterised):
def __init__(self,D,parts=[], input_slices=None):
@ -155,7 +155,7 @@ class kern(parameterised):
D = K1.D + K2.D
newkernparts = [product_orthogonal(k1,k2).parts[0] for k1, k2 in itertools.product(K1.parts,K2.parts)]
newkernparts = [product_orthogonal(k1,k2) for k1, k2 in itertools.product(K1.parts,K2.parts)]
slices = []
for sl1, sl2 in itertools.product(K1.input_slices,K2.input_slices):
@ -235,6 +235,8 @@ class kern(parameterised):
X2 = X
target = np.zeros(self.Nparam)
[p.dK_dtheta(partial[s1,s2],X[s1,i_s],X2[s2,i_s],target[ps]) for p,i_s,ps,s1,s2 in zip(self.parts, self.input_slices, self.param_slices, slices1, slices2)]
#TODO: transform the gradients here!
return target
def dK_dX(self,partial,X,X2=None,slices1=None,slices2=None):
@ -324,6 +326,7 @@ class kern(parameterised):
[p.psi2(Z[s2,i_s],mu[s1,i_s],S[s1,i_s],target[s1,s2,s2]) for p,i_s,s1,s2 in zip(self.parts,self.input_slices,slices1,slices2)]
# "crossterms". Here we are recomputing psi1 for white (we don't need to), but it's
# not really expensive, since it's just a matrix of zeroes.
# psi1_matrices = [np.zeros((mu.shape[0], Z.shape[0])) for p in self.parts]

19
GPy/kern/linear.py
View file

@ -15,34 +15,33 @@ class linear(kernpart):
:param D: the number of input dimensions
:type D: int
:param variances: the vector of variances :math:`\sigma^2_i`
:type variances: np.ndarray of size (1,) or (D,) depending on ARD
:param ARD: Auto Relevance Determination. If equal to "False", the kernel is isotropic (ie. one single variance parameter \sigma^2), otherwise there is one variance parameter per dimension.
:type variances: array or list of the appropriate size (or float if there is only one variance parameter)
:param ARD: Auto Relevance Determination. If equal to "False", the kernel has only one variance parameter \sigma^2, otherwise there is one variance parameter per dimension.
:type ARD: Boolean
:rtype: kernel object
"""
def __init__(self,D,variances=None,ARD=True):
def __init__(self,D,variances=None,ARD=False):
self.D = D
self.ARD = ARD
if ARD == False:
self.Nparam = 1
self.name = 'linear'
if variances is not None:
if isinstance(variances, float):
variances = np.array([variances])
assert variances.shape == (1,)
variances = np.asarray(variances)
assert variances.size == 1, "Only one variance needed for non-ARD kernel"
else:
variances = np.ones(1)
self._Xcache, self._X2cache = np.empty(shape=(2,))
else:
self.Nparam = self.D
self.name = 'linear_ARD'
self.name = 'linear'
if variances is not None:
assert variances.shape == (self.D,)
variances = np.asarray(variances)
assert variances.size == self.D, "bad number of lengthscales"
else:
variances = np.ones(self.D)
self._set_params(variances)
self._set_params(variances.flatten())
#initialize cache
self._Z, self._mu, self._S = np.empty(shape=(3,1))

5
GPy/kern/rbf.py
View file

@ -12,7 +12,7 @@ class rbf(kernpart):
.. math::
k(r) = \sigma^2 \exp(- \frac{1}{2}r^2) \qquad \qquad \\text{ where } r^2 = \sum_{i=1}^d \frac{ (x_i-x^\prime_i)^2}{\ell_i^2}}
k(r) = \sigma^2 \exp(- \frac{1}{2}r^2) \ \ \ \ \ \\text{ where } r^2 = \sum_{i=1}^d \frac{ (x_i-x^\prime_i)^2}{\ell_i^2}}
where \ell_i is the lengthscale, \sigma^2 the variance and d the dimensionality of the input.
@ -21,7 +21,7 @@ class rbf(kernpart):
:param variance: the variance of the kernel
:type variance: float
:param lengthscale: the vector of lengthscale of the kernel
:type lengthscale: np.ndarray od size (1,) or (D,) depending on ARD
:type lengthscale: array or list of the appropriate size (or float if there is only one lengthscale parameter)
:param ARD: Auto Relevance Determination. If equal to "False", the kernel is isotropic (ie. one single lengthscale parameter \ell), otherwise there is one lengthscale parameter per dimension.
:type ARD: Boolean
:rtype: kernel object
@ -75,6 +75,7 @@ class rbf(kernpart):
def K(self,X,X2,target):
if X2 is None:
X2 = X
self._K_computations(X,X2)
np.add(self.variance*self._K_dvar, target,target)