apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-08 11:32:39 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

non essential tidying in stationary

Browse source
This commit is contained in:
James Hensman 2014-02-28 12:06:28 +00:00
parent 7ae9d03c45
commit 7aad39e70e
1 changed files with 10 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

22
GPy/kern/_src/stationary.py
View file

@ -74,16 +74,10 @@ class Stationary(Kern):
r = self._scaled_dist(X, X2) r = self._scaled_dist(X, X2)
return self.K_of_r(r) return self.K_of_r(r)
#@Cache_this(limit=5, ignore_args=(0,))
def _dist(self, X, X2):
if X2 is None:
X2 = X
return X[:, None, :] - X2[None, :, :]
#@Cache_this(limit=5, ignore_args=(0,)) #@Cache_this(limit=5, ignore_args=(0,))
def _unscaled_dist(self, X, X2=None): def _unscaled_dist(self, X, X2=None):
""" """
Compute the square distance between each row of X and X2, or between Compute the Euclidean distance between each row of X and X2, or between
each pair of rows of X if X2 is None. each pair of rows of X if X2 is None.
""" """
if X2 is None: if X2 is None:
@ -99,7 +93,7 @@ class Stationary(Kern):
""" """
Efficiently compute the scaled distance, r. Efficiently compute the scaled distance, r.
r = \sum_{q=1}^Q (x_q - x'q)^2/l_q^2 r = \sqrt( \sum_{q=1}^Q (x_q - x'q)^2/l_q^2 )
Note that if thre is only one lengthscale, l comes outside the sum. In Note that if thre is only one lengthscale, l comes outside the sum. In
this case we compute the unscaled distance first (in a separate this case we compute the unscaled distance first (in a separate
@ -129,10 +123,10 @@ class Stationary(Kern):
dL_dr = self.dK_dr(r) * dL_dK dL_dr = self.dK_dr(r) * dL_dK
if self.ARD: if self.ARD:
#rinv = self._inv_dist(X, X2) #rinv = self._inv_dis# this is rather high memory? Should we loop instead?t(X, X2)
#x_xl3 = np.square(self._dist(X, X2)) # TODO: this is rather high memory? Should we loop instead? #d = X[:, None, :] - X2[None, :, :]
#x_xl3 = np.square(d)
#self.lengthscale.gradient = -((dL_dr*rinv)[:,:,None]*x_xl3).sum(0).sum(0)/self.lengthscale**3 #self.lengthscale.gradient = -((dL_dr*rinv)[:,:,None]*x_xl3).sum(0).sum(0)/self.lengthscale**3
self.lengthscale.gradient = np.zeros(self.input_dim)
tmp = dL_dr*self._inv_dist(X, X2) tmp = dL_dr*self._inv_dist(X, X2)
if X2 is None: X2 = X if X2 is None: X2 = X
[np.copyto(self.lengthscale.gradient[q:q+1], -np.sum(tmp * np.square(X[:,q:q+1] - X2[:,q:q+1].T))/self.lengthscale[q]**3) for q in xrange(self.input_dim)] [np.copyto(self.lengthscale.gradient[q:q+1], -np.sum(tmp * np.square(X[:,q:q+1] - X2[:,q:q+1].T))/self.lengthscale[q]**3) for q in xrange(self.input_dim)]
@ -162,7 +156,9 @@ class Stationary(Kern):
r = self._scaled_dist(X, X2) r = self._scaled_dist(X, X2)
invdist = self._inv_dist(X, X2) invdist = self._inv_dist(X, X2)
dL_dr = self.dK_dr(r) * dL_dK dL_dr = self.dK_dr(r) * dL_dK
#The high-memory numpy way: ret = np.sum((invdist*dL_dr)[:,:,None]*self._dist(X, X2),1)/self.lengthscale**2 #The high-memory numpy way:
#d = X[:, None, :] - X2[None, :, :]
#ret = np.sum((invdist*dL_dr)[:,:,None]*d,1)/self.lengthscale**2
#if X2 is None: #if X2 is None:
#ret *= 2. #ret *= 2.
@ -245,7 +241,7 @@ class Matern52(Stationary):
.. math:: .. math::
k(r) = \sigma^2 (1 + \sqrt{5} r + \\frac53 r^2) \exp(- \sqrt{5} r) \ \ \ \ \ \\text{ where } r = \sqrt{\sum_{i=1}^input_dim \\frac{(x_i-y_i)^2}{\ell_i^2} } k(r) = \sigma^2 (1 + \sqrt{5} r + \\frac53 r^2) \exp(- \sqrt{5} r)
""" """
def __init__(self, input_dim, variance=1., lengthscale=None, ARD=False, name='Mat52'): def __init__(self, input_dim, variance=1., lengthscale=None, ARD=False, name='Mat52'):
super(Matern52, self).__init__(input_dim, variance, lengthscale, ARD, name) super(Matern52, self).__init__(input_dim, variance, lengthscale, ARD, name)