def test_logsumexp():
from galpy.util import logsumexp
sumthis= numpy.array([[0.,1.]])
sum= numpy.log(numpy.exp(0.)+numpy.exp(1.))
assert numpy.all(numpy.fabs(logsumexp(sumthis,axis=0)-sumthis) < 10.**-10.), 'galpy.util.logsumexp did not work as expected'
assert numpy.fabs(logsumexp(sumthis,axis=1)-sum) < 10.**-10., 'galpy.util.logsumexp did not work as expected'
assert numpy.fabs(logsumexp(sumthis,axis=None)-sum) < 10.**-10., 'galpy.util.logsumexp did not work as expected'
return None
def test_logsumexp():
from galpy.util import logsumexp
sumthis = numpy.array([[0., 1.]])
sum = numpy.log(numpy.exp(0.) + numpy.exp(1.))
assert numpy.all(
numpy.fabs(logsumexp(sumthis, axis=0) - sumthis) < 10.**-10.
), 'galpy.util.logsumexp did not work as expected'
assert numpy.fabs(
logsumexp(sumthis, axis=1) -
sum) < 10.**-10., 'galpy.util.logsumexp did not work as expected'
assert numpy.fabs(
logsumexp(sumthis, axis=None) -
sum) < 10.**-10., 'galpy.util.logsumexp did not work as expected'
return None
def __call__(self, x, h=None, log=False, scale=True, sx2=None):
"""
NAME:
__call__
PURPOSE:
return the density
INPUT:
x - evaluate at this x [nx,dim]
log= (False) if True, return the log
h= (None) if set, use this bandwidth
scale= (True), if False, don't rescale first
sx2= uncertainty variance of x
OUTPUT:
density (or log)
HISTORY:
2012-11-12 - Written - Bovy (IAS)
"""
if h is None:
thish = self._h
else:
thish = h
x = self._prepare_x(x, scale)
divh = numpy.tile(thish * self._lambda.T, (x.shape[0], 1, 1))
if not sx2 is None:
divh = numpy.sqrt(divh**2. +
numpy.tile(sx2.T, (self._ndata, 1, 1)).T)
thiskernel = self._kernel(numpy.tile(x.T,
(self._ndata, 1, 1)).T / divh,
numpy.tile(self._data.T,
(x.shape[0], 1, 1)) / divh,
log=log)
if log:
return logsumexp(thiskernel+numpy.tile(numpy.log(self._w),(x.shape[0],1))\
# return -self._dim*numpy.log(thish)\
-numpy.sum(numpy.log(divh),axis=1),axis=1) #latter assumes that lambda are spherical
# -self._dim*numpy.log(self._lambda[:,0]),axis=1) #latter assumes that lambda are spherical
else:
return 1./thish**self._dim\
*numpy.sum(numpy.tile(self._w,(x.shape[0],1))\
*thiskernel/self._lambda[:,0]**self._dim,axis=1)
def __call__(self,x,h=None,log=False,scale=True,sx2=None):
"""
NAME:
__call__
PURPOSE:
return the density
INPUT:
x - evaluate at this x [nx,dim]
log= (False) if True, return the log
h= (None) if set, use this bandwidth
scale= (True), if False, don't rescale first
sx2= uncertainty variance of x
OUTPUT:
density (or log)
HISTORY:
2012-11-12 - Written - Bovy (IAS)
"""
if h is None:
thish= self._h
else:
thish= h
x= self._prepare_x(x,scale)
divh= numpy.tile(thish*self._lambda.T,(x.shape[0],1,1))
if not sx2 is None:
divh= numpy.sqrt(divh**2.
+numpy.tile(sx2.T,(self._ndata,1,1)).T)
thiskernel= self._kernel(numpy.tile(x.T,(self._ndata,1,1)).T/divh,
numpy.tile(self._data.T,(x.shape[0],1,1))/divh,
log=log)
if log:
return logsumexp(thiskernel+numpy.tile(numpy.log(self._w),(x.shape[0],1))\
# return -self._dim*numpy.log(thish)\
-numpy.sum(numpy.log(divh),axis=1),axis=1) #latter assumes that lambda are spherical
# -self._dim*numpy.log(self._lambda[:,0]),axis=1) #latter assumes that lambda are spherical
else:
return 1./thish**self._dim\
*numpy.sum(numpy.tile(self._w,(x.shape[0],1))\
*thiskernel/self._lambda[:,0]**self._dim,axis=1)
