def iteration(points, v, sig, tvector):
x = hstack((v, tvector))
J = jacobian(v, sig, tvector)
r = residuals(points[:, 0], points[:, 1], v, sig, tvector)
dell = dot(J.T, r)
#Solving this with a conjugate gradient method,
#Hopefully the zero entries lead to some savings, but I haven't really checked.
#there may also be a way to get savings by not multiplying out J'J,
#but idk how to do that.
pk = linalg.cg(dot(J.T, J), -dell)[0]
#Using backtracking to get alpha
#Using this method because it's what I know.
#http://en.wikipedia.org/wiki/Wolfe_conditions
alpha = 1.0 #"For Newton and quasi-Newton methods, the step \(\alpha_0 = 1\) should always be used as the initial trial step length."--pg. 59
rho = 0.7 #between 0 and 1
c = 0.1**6 #The "C_1" used in the first wolfe condish. between 0 and 1, typically "quite small"
xnew = x + alpha * pk
rnew = residuals(points[:, 0], points[:, 1], xnew[0:9], sig,
xnew[9:x.shape[0]])
#this is to prevent sticking when the solution is outside the boundary.
#I don't think it works very well.
tired = False
while (dot(rnew, rnew) > dot(r, r) + c * alpha * dot(dell, pk)
or not (discr(xnew[0], xnew[1], xnew[2], xnew[3], xnew[4],
xnew[9:x.shape[0]]) >= 0).all()) and not tired:
alpha = rho * alpha
xnew = x + alpha * pk
rnew = residuals(points[:, 0], points[:, 1], xnew[0:9], sig,
xnew[9:x.shape[0]])
if (rho < 0.1**2) and (discr(xnew[0], xnew[1], xnew[2], xnew[3],
xnew[4], xnew[9:x.shape[0]]) >= 0).all():
print "warning: Boundary condition issues"
tired = True
return xnew[0:9], xnew[9:x.shape[0]]
def iteration(points,v,sig,tvector):
x=hstack((v,tvector))
J=jacobian(v,sig,tvector)
r=residuals(points[:,0],points[:,1],v,sig,tvector)
dell=dot(J.T,r)
#Solving this with a conjugate gradient method,
#Hopefully the zero entries lead to some savings, but I haven't really checked.
#there may also be a way to get savings by not multiplying out J'J,
#but idk how to do that.
pk=linalg.cg(dot(J.T,J),-dell)[0]
#Using backtracking to get alpha
#Using this method because it's what I know.
#http://en.wikipedia.org/wiki/Wolfe_conditions
alpha=1.0 #"For Newton and quasi-Newton methods, the step \(\alpha_0 = 1\) should always be used as the initial trial step length."--pg. 59
rho=0.7 #between 0 and 1
c=0.1**6 #The "C_1" used in the first wolfe condish. between 0 and 1, typically "quite small"
xnew=x+alpha*pk
rnew=residuals(points[:,0],points[:,1],xnew[0:9],sig,xnew[9:x.shape[0]])
#this is to prevent sticking when the solution is outside the boundary.
#I don't think it works very well.
tired=False
while (dot(rnew,rnew) > dot(r,r)+c*alpha*dot(dell,pk) or not (discr(xnew[0],xnew[1],xnew[2],xnew[3],xnew[4],xnew[9:x.shape[0]])>=0).all()) and not tired:
alpha=rho*alpha
xnew=x+alpha*pk
rnew=residuals(points[:,0],points[:,1],xnew[0:9],sig,xnew[9:x.shape[0]])
if (rho < 0.1**2) and (discr(xnew[0],xnew[1],xnew[2],xnew[3],xnew[4],xnew[9:x.shape[0]])>=0).all():
print "warning: Boundary condition issues"
tired=True
return xnew[0:9],xnew[9:x.shape[0]]
def _compute_grad_solve_iterative(self, G, H, grad_loss_params, tol=1e-6):
from scipy.sparse import linalg
if self._warm_start is None:
v, convergence = linalg.cg(
H.tondarray(), grad_loss_params.tondarray(), tol=tol)
else:
v, convergence = linalg.cg(
H.tondarray(), grad_loss_params.tondarray(), tol=tol,
x0=self._warm_start.tondarray())
if convergence != 0:
warnings.warn('Convergence of poisoning algorithm not reached!')
v = CArray(v.ravel())
# store v to be used as warm start at the next iteration
self._warm_start = v
gt = -G.dot(v.T)
return gt.ravel()
import scipy.io.mmio
matrix = scipy.io.mmio.mmread(filename).tocsr()
print >>sys.stderr, "done"
b = numpy.ones(matrix.shape[0])
calc_time = 0
if options.scipy:
if scipy.version.version == '0.6.0':
import scipy.linalg as linalg
else:
import scipy.sparse.linalg as linalg
for i in xrange(5):
cg_time = time.time()
x, info = linalg.cg(matrix, b, maxiter=options.m)
cg_time = time.time() - cg_time
calc_time += cg_time
print("time = " + str(cg_time))
for i in xrange(5):
cg_time = time.time()
for j in xrange(options.m):
dummy = matrix * x
cg_time = time.time() - cg_time
calc_time += cg_time
print("mul_time = " + str(cg_time))
else:
print >>sys.stderr, "Initializing akx...",
import akxconfig
if options.tb_num:
akxconfig.threadcounts = [options.tb_num]
import scipy.io.mmio
matrix = scipy.io.mmio.mmread(filename).tocsr()
print >> sys.stderr, "done"
b = numpy.ones(matrix.shape[0])
calc_time = 0
if options.scipy:
if scipy.version.version == '0.6.0':
import scipy.linalg as linalg
else:
import scipy.sparse.linalg as linalg
for i in xrange(5):
cg_time = time.time()
x, info = linalg.cg(matrix, b, maxiter=options.m)
cg_time = time.time() - cg_time
calc_time += cg_time
print("time = " + str(cg_time))
for i in xrange(5):
cg_time = time.time()
for j in xrange(options.m):
dummy = matrix * x
cg_time = time.time() - cg_time
calc_time += cg_time
print("mul_time = " + str(cg_time))
else:
print >> sys.stderr, "Initializing akx...",
import akxconfig
if options.tb_num:
akxconfig.threadcounts = [options.tb_num]