def _fg(x):
return func(x) if grad is None else func(x), grad(x)
def getbundle(func, x0, grad=None, g0=None, samprad=1e-4, n=None):
"""
Get bundle of n-1 gradients at points near x, in addition to g,
which is gradient at x and goes in first column
intended to be called by gradsampfixed
Parameters
----------
func: callable function on 1D arrays of length nvar
function being optimized
grad: callable function
gradient of func
x0: 1D array of len nvar, optional (default None)
intial points, one per column
g0: 1D array of len nvar, optional (default None)
gradient at point x0
samprad: float, optional (default 1e-4)
sampling radius; this should be a small positive float
n: int, optional (default min(100, 2 * nvar, nvar + 10))
number of points and gradients to sample
Returns
-------
xbundle: 2D array of shape (nvar, n)
bundle of n points sampled in the samprand-ball around x0
xbundle: 2D array of shape (nvar, n)
bundle of n gradients sampled in the samprand-ball around x0
Raises
------
RuntimeError
"""
def _fg(x):
return func(x) if grad is None else (func(x), grad(x))
x0 = np.ravel(x0)
nvar = len(x0)
n = min(100, min(2 * nvar, nvar + 10)) if n is None else n
xbundle = np.ndarray((nvar, n))
gbundle = np.ndarray((nvar, n))
xbundle[..., 0] = x0
gbundle[..., 0] = g0 if not g0 is None else _fg(x0)[1]
for k in xrange(1, n): # note the 1
xpert = x0 + samprad * (np.random.rand(nvar) - 0.5
) # uniform distribution
f, g = _fg(xpert)
count = 0
# in particular, disallow infinite function values
while np.isnan(f) or np.isinf(f) or np.any(np.isnan(g)) or np.any(
np.isinf(g)):
xpert = (x0 + xpert) / 2. # contract back until feasible
f, g = func(xpert), grad(xpert)
count = count + 1
if count > 100: # should never happen, but just in case
raise RuntimeError(
'getbundle: too many contractions needed to find finite'
' func and grad values')
xbundle[..., k] = xpert
gbundle[..., k] = g
return xbundle, gbundle
def _fg(x):
return func(x) if grad is None else func(x), grad(x)
def getbundle(func, x0, grad=None, g0=None, samprad=1e-4, n=None):
"""
Get bundle of n-1 gradients at points near x, in addition to g,
which is gradient at x and goes in first column
intended to be called by gradsampfixed
Parameters
----------
func: callable function on 1D arrays of length nvar
function being optimized
grad: callable function
gradient of func
x0: 1D array of len nvar, optional (default None)
intial points, one per column
g0: 1D array of len nvar, optional (default None)
gradient at point x0
samprad: float, optional (default 1e-4)
sampling radius; this should be a small positive float
n: int, optional (default min(100, 2 * nvar, nvar + 10))
number of points and gradients to sample
Returns
-------
xbundle: 2D array of shape (nvar, n)
bundle of n points sampled in the samprand-ball around x0
xbundle: 2D array of shape (nvar, n)
bundle of n gradients sampled in the samprand-ball around x0
Raises
------
RuntimeError
"""
def _fg(x):
return func(x) if grad is None else (func(x), grad(x))
x0 = np.ravel(x0)
nvar = len(x0)
n = min(100, min(2 * nvar, nvar + 10)) if n is None else n
xbundle = np.ndarray((nvar, n))
gbundle = np.ndarray((nvar, n))
xbundle[..., 0] = x0
gbundle[..., 0] = g0 if not g0 is None else _fg(x0)[1]
for k in xrange(1, n): # note the 1
xpert = x0 + samprad * (np.random.rand(nvar) - 0.5
) # uniform distribution
f, g = _fg(xpert)
count = 0
# in particular, disallow infinite function values
while np.isnan(f) or np.isinf(f) or np.any(
np.isnan(g)) or np.any(np.isinf(g)):
xpert = (x0 + xpert) / 2. # contract back until feasible
f, g = func(xpert), grad(xpert)
count = count + 1
if count > 100: # should never happen, but just in case
raise RuntimeError(
'getbundle: too many contractions needed to find finite'
' func and grad values')
xbundle[..., k] = xpert
gbundle[..., k] = g
return xbundle, gbundle