def time_solve(self, n, solver):
if solver == 'dense':
linalg.solve(self.P_dense, self.b)
elif solver == 'cg':
cg(self.P_sparse, self.b)
elif solver == 'minres':
minres(self.P_sparse, self.b)
elif solver == 'spsolve':
spsolve(self.P_sparse, self.b)
else:
raise ValueError('Unknown solver: %r' % solver)
def time_solve(self, n, solver):
if solver == 'dense':
linalg.solve(self.P_dense, self.b)
elif solver == 'cg':
cg(self.P_sparse, self.b)
elif solver == 'minres':
minres(self.P_sparse, self.b)
elif solver == 'spsolve':
spsolve(self.P_sparse, self.b)
else:
raise ValueError('Unknown solver: %r' % solver)
def solve(self, b, maxiter=1000, tol=1.e-10):
if self.method == 'Dense':
from scipy.linalg import solve
return solve(self.mat, b)
else:
from scipy.sparse.linalg import gmres, aslinearoperator, minres
P = self.P
Aop = aslinearoperator(self)
residual = Residual()
if P != None:
x, info = gmres(Aop,
b,
tol=tol,
restart=30,
maxiter=maxiter,
callback=residual,
M=P)
else:
x, info = minres(Aop,
b,
tol=tol,
maxiter=maxiter,
callback=residual)
self.solvmatvecs += residual.itercount()
if self.verbose:
print("Number of iterations is %g and status is %g" %
(residual.itercount(), info))
return x
def _solve_sym_nonposdef(self, Q, q):
# since Q is indefinite, i.e., the function is linear along the eigenvectors
# correspondent to the null eigenvalues, the system has not solutions, so we
# will choose the one that minimizes the residue, i.e. the least-squares solution
# see more @ https://docs.scipy.org/doc/scipy/reference/sparse.linalg.html#solving-linear-problems
# bad numerical solution: does not exploit the symmetricity of Q, waiting for `symmlq` in scipy
if self.sym_nonposdef_solver == 'lsqr':
x = lsqr(Q, -q)[0]
else:
# `min ||Ax - b||` is formally equivalent to solve the linear system:
# A^T A x = A^T b
Q, q = np.inner(Q, Q), Q.T.dot(q)
if self.sym_nonposdef_solver == 'minres':
x = minres(Q, -q)[0]
else:
raise TypeError(
f'{self.sym_nonposdef_solver} is not an allowed solver, '
f'choose one of `minres` or `lsqr`')
return x
def gradient_descent(self, x, y, w, l, hv):
# fix number of steps
for i in range(100):
max_l = np.fmax(0, l)
# compute objective function value
obj = np.sum(max_l**2) / 2 + self.lmbda * w.dot(w) / 2
# compute gradient
grad = self.lmbda * w - np.append([np.dot(max_l * y, x)],
[np.sum(max_l * y)])
# perform line search for optimal w (normal to hyperplane)
sv = np.where(l > 0)[0]
# vector where optimal solution is located
vec, info = linalg.minres(hv, -grad)
x_d = x.dot(vec[0:-1]) + vec[-1]
w_d = self.lmbda * w[0:-1].dot(vec[0:-1])
d_d = self.lmbda * vec[0:-1].dot(vec[0:-1])
t = 0
l_0 = l
for i in range(1000):
l_0 = l - t * (y * x_d)
sv = np.where(l_0 > 0)[0]
g = w_d + t * d_d - (l_0[sv] * y[sv]).dot(x_d[sv])
h = d_d + x_d[sv].dot(x_d[sv])
t -= g / h
w += t * vec
if -vec.dot(grad) < 0.0000001 * obj:
break
def _solve_lin(self):
'Auxilliary function: solve the linear system exactly'
# Linear equation coefficients
Lhs, rhs, x0 = self._get_lineq()
# Set x[I], y, czl, czu by solving linear equation
# xy = cholmod.splinsolve(Lhs,rhs)
self.CG_r0 = Lhs * x0 - rhs
# Convert to numpy/scipy matrix
npLhs = cvxopt_to_numpy_matrix(Lhs)
nprhs = cvxopt_to_numpy_matrix(matrix(rhs))
npx0 = cvxopt_to_numpy_matrix(x0)
# Solve the linear system
collector = Iter_collect(tol=self.option['ResTol'])
# collector = Iter_collect()
cg = minres(npLhs, nprhs, npx0, callback=collector)
xy = numpy_to_cvxopt_matrix(cg[0])
self.cgiter = collector.iter
self.CG_r = Lhs * numpy_to_cvxopt_matrix(collector.x) - rhs
nI = len(self.I)
ny = self.QP.numeq
ncL = len(self.cAL)
ncU = len(self.cAU)
self.x[self.I] = xy[0:nI]
if xy[nI:nI + ny].size[1] == 1:
self.y = xy[nI:nI + ny]
if self.czl.size[0] > 0:
self.czl[self.cAL] = xy[nI + ny:nI + ny + ncL]
self.czu[self.cAU] = xy[nI + ny + ncL:]
def rvs(self, b, omega, tau, random_state):
"""Generate a random draw from this distribution."""
eps = random_state.standard_normal(self._n_plus_k)
rnorm1 = np.sqrt(omega) * eps[:self._n]
rnorm2 = self._eigen @ (sqrt(tau) * eps[self._n:])
out = b + rnorm1 + rnorm2
block_prec = self._block_Q.copy()
block_prec.data = tau * block_prec.data
diag_vec = np.tile(omega, 2)
block_prec.setdiag(block_prec.diagonal() + diag_vec)
self._rhs[:self._n] = out
# Iterative solvers are efficient at solving sparse large systems.
xz, fail = minres(block_prec, self._rhs, x0=self._guess)
# update starting guess for next call to the function
self._guess = xz
if fail:
raise RuntimeError('MINRES solver did not converge!')
x = xz[:self._n]
z = xz[self._n:]
ensure_sums_to_zero(x, z, out)
return out
def estimate_rhoB(adjacency_matrix):
print("Tuning rhoB estimation")
degrees = np.asarray(adjacency_matrix.sum(axis=1), dtype=np.float64).flatten()
guessForFirstEigen = (degrees ** 2).mean() / degrees.mean() - 1
errtol = 1e-2
maxIter = 10
err = 1
iteration = 0
rhoB = guessForFirstEigen
print("Initial guess of rhoB is %f" % rhoB)
while err > errtol and iteration < maxIter:
iteration += 1
print("Building matrices")
BH = build_weighted_bethe_hessian(adjacency_matrix, rhoB)
BHprime = build_weighted_bethe_hessian_derivative(adjacency_matrix, rhoB)
sigma = 0
op_inverse = lambda v: minres(BH, v, tol=1e-5)[0]
OPinv = LinearOperator(matvec=op_inverse, shape=adjacency_matrix.shape, dtype=np.float64)
print("Solving the eigenproblem")
mu, x = eigsh(A=BH, M=BHprime, k=1, which='LM', sigma=sigma, OPinv=OPinv)
mu = mu[0]
print("mu is %f" % mu)
err = abs(mu) / rhoB
rhoB -= mu
print("Iteration %d, updated value of rhoB %f, relative error %f" % (iteration, rhoB, err))
return rhoB
def izracun_najbrzeg_silaska(A, b, x0):
x = ln.minres(A, b, x0, 0.0, 1e-8, None, None, neka_vrijednost, 1)
x = x[0]
print("\nRezultat je\n")
print(x, '\n')
return x
def eigPsi_0(H, Dim, L, psi0_flag, sigma=None):
"""Compute the initial state as the groundstate of the Hamiltonian or a middle eigenstate
Args:
H(2d sparse matrix) = Hamiltonian matrix in sparse form
Dim(int) = Hilbert space dimension
L(int) = Size of the chain
psi0_flag(int) = Flag for initial state (0 ---> GS, 1 ---> Middle state)
sigma(float) = Target energy for shift-invert method
"""
if psi0_flag == 1:
if sigma is None:
print("No target energy given. Check psi0_flag and/or sigma")
else:
if L < 14:
e, v = np.linalg.eig(H.todense())
idx = e.argsort()[::1]
e_0 = e[idx]
v_0 = v[:, idx][Dim // 2]
elif L == 14 or L == 16:
e_0, v_0 = eigsh(H, 1, sigma=sigma, maxiter=1E4)
v_0 = np.asarray(v_0).ravel()
else:
OP = H - sigma * eye(Dim)
OPinv = LinearOperator(
matvec=lambda v: minres(OP, v, tol=1e-5)[0],
shape=H.shape,
dtype=H.dtype)
e_0, v_0 = eigsh(H, sigma=sigma, k=1, tol=1e-9, OPinv=OPinv)
v_0 = np.asarray(v_0).ravel()
if psi0_flag == 0:
e_0, v_0 = eigsh(H, k=1)
return e_0, v_0
def _solve_lin(self):
'Auxilliary function: solve the linear system exactly'
# Linear equation coefficients
Lhs, rhs, x0 = self._get_lineq()
# Set x[I], y, czl, czu by solving linear equation
# xy = cholmod.splinsolve(Lhs,rhs)
self.CG_r0 = Lhs*x0 - rhs
# Convert to numpy/scipy matrix
npLhs = cvxopt_to_numpy_matrix(Lhs)
nprhs = cvxopt_to_numpy_matrix(matrix(rhs))
npx0 = cvxopt_to_numpy_matrix(x0)
# Solve the linear system
collector = Iter_collect(tol=self.option['ResTol'])
# collector = Iter_collect()
cg = minres(npLhs,nprhs,npx0,callback=collector)
xy = numpy_to_cvxopt_matrix(cg[0])
self.cgiter = collector.iter
self.CG_r = Lhs*numpy_to_cvxopt_matrix(collector.x) - rhs
nI = len(self.I)
ny = self.QP.numeq
ncL = len(self.cAL)
ncU = len(self.cAU)
self.x[self.I] = xy[0:nI]
if xy[nI:nI + ny].size[1] == 1:
self.y = xy[nI:nI + ny]
if self.czl.size[0] > 0:
self.czl[self.cAL] = xy[nI+ny:nI+ny+ncL]
self.czu[self.cAU] = xy[nI+ny+ncL:]
def _update_g(self, z, lamb, D2g_ext, **kwargs):
"""
Evaluate the g-update proximal map for ADMM.
input:
z : current z-value (2*N_pad)-shaped
lamb : lagrange multiplier enforcing D.dot(g) = z
(same shape as z)
D2g_ext : D.T.dot(D.dot(g_ext)) the fourth-derivative of the
loop field
kwargs are passed to spl.minres. tol and maxiter control how hard it
tries
output:
updated g : (N_pad)-shaped
"""
self._oldg = self._oldg if self._oldg is not None else np.zeros(
self.N_pad)
self._oldAg = self.A.dot(self._oldg)
self._c = (self._Mtphi - self.D.T.dot(lamb - self.rho * z) -
self.rho * D2g_ext) - self._oldAg
maxiter = kwargs.get('maxiter', 200)
tol = kwargs.get('tol', 1E-6)
self._gminsol = spl.minres(self.A, self._c, maxiter=maxiter, tol=tol)
self._newg = self._gminsol[0] + self._oldg
self._oldg = self._newg.copy()
return self._newg
def dfs_trunk(sim, A, alpha=0.99, QUERYKNN=10, maxiter=8, K=100, tol=1e-3):
qsim = sim_kernel(sim).T
sortidxs = np.argsort(-qsim, axis=1)
for i in range(len(qsim)):
qsim[i, sortidxs[i, QUERYKNN:]] = 0
qsims = sim_kernel(qsim)
W = sim_kernel(A)
W = csr_matrix(topK_W(W, K))
out_ranks = []
t = time.time()
for i in range(qsims.shape[0]):
qs = qsims[i, :]
tt = time.time()
w_idxs, W_trunk = find_trunc_graph(qs, W, 2)
Wn = normalize_connection_graph(W_trunk)
Wnn = eye(Wn.shape[0]) - alpha * Wn
f, inf = s_linalg.minres(Wnn, qs[w_idxs], tol=tol, maxiter=maxiter)
ranks = w_idxs[np.argsort(-f.reshape(-1))]
missing = np.setdiff1d(np.arange(A.shape[1]), ranks)
out_ranks.append(
np.concatenate([ranks.reshape(-1, 1),
missing.reshape(-1, 1)],
axis=0))
print(time.time() - t, 'qtime')
out_ranks = np.concatenate(out_ranks, axis=1)
return out_ranks
def solve(A, b, method, tol=1e-3):
""" General sparse solver interface.
method can be one of
- spsolve_umfpack_mmd_ata
- spsolve_umfpack_colamd
- spsolve_superlu_mmd_ata
- spsolve_superlu_colamd
- bicg
- bicgstab
- cg
- cgs
- gmres
- lgmres
- minres
- qmr
- lsqr
- lsmr
"""
if method == 'spsolve_umfpack_mmd_ata':
return spla.spsolve(A, b, use_umfpack=True, permc_spec='MMD_ATA')
elif method == 'spsolve_umfpack_colamd':
return spla.spsolve(A, b, use_umfpack=True, permc_spec='COLAMD')
elif method == 'spsolve_superlu_mmd_ata':
return spla.spsolve(A, b, use_umfpack=False, permc_spec='MMD_ATA')
elif method == 'spsolve_superlu_colamd':
return spla.spsolve(A, b, use_umfpack=False, permc_spec='COLAMD')
elif method == 'bicg':
res = spla.bicg(A, b, tol=tol)
return res[0]
elif method == 'bicgstab':
res = spla.bicgstab(A, b, tol=tol)
return res[0]
elif method == 'cg':
res = spla.cg(A, b, tol=tol)
return res[0]
elif method == 'cgs':
res = spla.cgs(A, b, tol=tol)
return res[0]
elif method == 'gmres':
res = spla.gmres(A, b, tol=tol)
return res[0]
elif method == 'lgmres':
res = spla.lgmres(A, b, tol=tol)
return res[0]
elif method == 'minres':
res = spla.minres(A, b, tol=tol)
return res[0]
elif method == 'qmr':
res = spla.qmr(A, b, tol=tol)
return res[0]
elif method == 'lsqr':
res = spla.lsqr(A, b, atol=tol, btol=tol)
return res[0]
elif method == 'lsmr':
res = spla.lsmr(A, b, atol=tol, btol=tol)
return res[0]
else:
raise Exception('UnknownSolverType')
def disc_proj(u1, u2, n, h, V, P,type): proj = Function(V) u = TrialFunction(V) v = TestFunction(V) if type=='Uh': dof_coords = V.tabulate_dof_coordinates().reshape(-1, 2) #print(len(dof_coords)) #u_sol = Function(V) for j in range(len(dof_coords)): proj.vector()[j] = u1(dof_coords[j])+u2(dof_coords[j]) elif type=='L2': m = u*v*dx Me = assemble(m) M_mat = as_backend_type(Me).mat() M = sp.csc_matrix(sp.csr_matrix(M_mat.getValuesCSR()[::-1], shape = M_mat.size)) rhs = assemble(u1*v*dx) + assemble(u2*v*dx) val,auuu = spl.minres(M,rhs,tol=1e-10) proj.vector()[:] = val elif type=='H1': m = u*v*dx + inner(grad(u),grad(v))*dx Me = assemble(m) M_mat = as_backend_type(Me).mat() M = sp.csc_matrix(sp.csr_matrix(M_mat.getValuesCSR()[::-1], shape = M_mat.size)) rhs = assemble(u1*v*dx) + assemble(u2*v*dx) + assemble(inner(grad(u1),grad(v))*dx) + assemble(inner(grad(u2),grad(v))*dx) val,auuu = spl.cg(M,rhs,tol=1e-7) proj.vector()[:] = val elif type=='Vh': Me = gh(u,v,Constant(1.)) M_mat = as_backend_type(Me).mat() M = sp.csc_matrix(sp.csr_matrix(M_mat.getValuesCSR()[::-1], shape = M_mat.size)) rhs = gh(u1,v,Constant(1.)) + gh(u2,v,Constant(1.)) val,auuu = spl.cg(M,rhs,M=P,tol=1e-7) proj.vector()[:] = val return proj
def JDLoop(A, m, perp, V, tol, verbose = False): assert(type(A) == type(np.zeros(1))) assert(type(perp) == type(np.zeros(1))) n = A.shape[0] assert(perp.shape[0]== n and perp.shape[1]== 1) assert(V.shape[0] == n and V.shape[1] == m-1) t = copy.copy(perp) k = 0 while k < .001: perp = copy.copy(t) #print "\tT0:",t for i in xrange(m-1): t[:,0] -= (np.dot(V[:,i].T,t[:,0]) * V[:,i]) #print "\tTF:", t #print "Norm v0, iteration",m,":",la.norm(v0) #print "Norm t:",la.norm(t) k = la.norm(t) / la.norm(perp) if verbose: print "\t\tOrthogonalization rescale kappa: ", k perp= copy.copy(t) perp /= la.norm(t) vm = perp for i in xrange(m-1): assert(np.dot(t[:,0].T,V[:,i]) < 1e-10) V = np.append(V,perp,1) M = np.dot(V.T, np.dot(A, V)) th, s = la.eigh(M, eigvals=(m-1,m-1)) u = np.dot(V,s) r = np.dot(A, np.dot(V, s)) - th * u if la.norm(r) < tol: print "\t\t\t residual", la.norm(r) print "\t\t\t SUCCESS!" return la.norm(r), th, u, t, vm diag = A.diagonal() diag.shape = (n,1) I = np.eye(n) P = I - np.dot(u, u.T) A2 = A - (th * I) MAT = np.dot(P, np.dot(A2,P)) #print MAT.shape, r.shape x, info = sla.minres(MAT,-r) t = x t /= la.norm(t) t.shape = (n,1) if verbose: print "\t\t\t residual", la.norm(r) print "\t\t\t t dot u: ", np.dot(t.T,u) #assert(np.dot(t.T,u) <= 1e-5) return la.norm(r), th, u, t, vm
def cg_diffusion(qsims, Wn, alpha=0.99, maxiter=10, tol=1e-3):
Wnn = eye(Wn.shape[0]) - alpha * Wn
out_sims = []
for i in range(qsims.shape[0]):
#f,inf = s_linalg.cg(Wnn, qsims[i,:], tol=tol, maxiter=maxiter)
f, inf = s_linalg.minres(Wnn, qsims[i, :], tol=tol, maxiter=maxiter)
out_sims.append(f.reshape(-1, 1))
out_sims = np.concatenate(out_sims, axis=1)
ranks = np.argsort(-out_sims, axis=0)
return ranks
def calculate(self,alldata_par, alldata_obs, alldata_wei):
no_snapshots, no_parameters = alldata_par.shape
TEMP = np.zeros([no_snapshots,no_snapshots])
for i in range(no_parameters):
v = alldata_par[:,i]
Q = npm.repmat(v,no_snapshots,1)
T = np.transpose(Q)
TEMP = TEMP + np.power(Q-T,2)
np.sqrt(TEMP,out=TEMP) # distances between all points
if self.no_keep>0:
MAX = 2*np.max(TEMP)
M = TEMP+MAX*np.eye(no_snapshots)
to_keep = np.ones(no_snapshots,dtype=bool)
for i in range(no_snapshots - self.no_keep):
argmin = np.argmin(M)
xx = argmin // no_snapshots
if self.expensive>0:
S=sum(M)
yy = argmin % no_snapshots
M[xx,yy]=MAX
M[yy,xx]=MAX
if S[yy]<S[xx]:
yy=xx
M[xx,:]=MAX
M[:,xx]=MAX
to_keep[xx]=False
TEMP = TEMP[to_keep,:]
TEMP = TEMP[:,to_keep]
alldata_par = alldata_par[to_keep,:]
alldata_obs = alldata_obs[to_keep,:]
try:
to_keep=np.append(to_keep,np.ones(no_parameters+1,dtype=bool))
self.initial_iteration = self.initial_iteration[to_keep]
except:
print("Exception!")
kernel.kernel(TEMP,self.type_kernel)
P = np.ones([no_snapshots,1]) # only constant polynomials
# print(P.shape, no_evaluations, alldata_par.shape, TEMP.shape)
P = np.append(alldata_par,P,axis=1) # plus linear polynomials
no_polynomials = P.shape[1]
TEMP = np.append(TEMP,P,axis=1)
TEMP2 = np.append(np.transpose(P),np.zeros([no_polynomials,no_polynomials]),axis=1)
TEMP2 = np.vstack([TEMP,TEMP2])
RHS = np.vstack([ alldata_obs, np.zeros([no_polynomials,1]) ])
# print("Condition_number:",np.linalg.cond(TEMP2))
# print("DEBUG2:", TEMP2.shape, np.linalg.matrix_rank(TEMP2), RHS.shape)
if self.type_solver == 0:
SOL=np.linalg.solve(TEMP2,RHS)
else:
SOL_=splin.minres(TEMP2,RHS,x0=self.initial_iteration,tol=pow(10,-self.type_solver),show=False)
SOL=SOL_[0]
# RES=RHS-np.reshape(np.matmul(TEMP2,SOL),(no_evaluations+no_polynomials,1))
# print("Residual norm:",np.linalg.norm(RES), "RHSnorm:", np.linalg.norm(RHS))
return SOL, no_snapshots, alldata_par, alldata_obs, alldata_wei, TEMP2, RHS
def solve_factorized_aug(z, Fval, LU, G, A):
M, N = G.shape
P, N = A.shape
"""Total number of inequality constraints"""
m = M
"""Primal variable"""
x = z[0:N]
"""Multiplier for equality constraints"""
nu = z[N:N + P]
"""Multiplier for inequality constraints"""
l = z[N + P:N + P + M]
"""Slacks"""
s = z[N + P + M:]
"""Dual infeasibility"""
rd = Fval[0:N]
"""Primal infeasibility"""
rp1 = Fval[N:N + P]
rp2 = Fval[N + P:N + P + M]
"""Centrality"""
rc = Fval[N + P + M:]
"""Sigma matrix"""
SIG = diags(l / s, 0)
"""LU is actually the augmented system W"""
W = LU
b1 = -rd - mydot(G.T, mydot(SIG, rp2)) + mydot(G.T, rc / s)
b2 = -rp1
b = np.hstack((b1, b2))
"""Prepare iterative solve via MINRES"""
sign = np.zeros(N + P)
sign[0:N // 2] = 1.0
sign[N // 2:] = -1.0
T = diags(sign, 0)
"""Change rhs"""
b_new = mydot(T, b)
dW = np.abs(W.diagonal())
dPc = np.ones(W.shape[0])
ind = (dW > 0.0)
dPc[ind] = 1.0 / dW[ind]
Pc = diags(dPc, 0)
dxnu, info = minres(W, b_new, tol=1e-10, M=Pc)
# dxnu = solve(J, b)
dx = dxnu[0:N]
dnu = dxnu[N:]
"""Obtain search directions for l and s"""
ds = -rp2 - mydot(G, dx)
# ds = s*ds
# SIG = np.diag(l/s)
dl = -mydot(SIG, ds) - rc / s
dz = np.hstack((dx, dnu, dl, ds))
return dz
def test_minres_scipy():
H = sprandsym(10)
v = sp_rand(10,3,0.8)
A = sparse([[H],[v]])
vrow = sparse([[v.T],[spmatrix([],[],[],(3,3))]])
A = sparse([A,vrow])
b = sp_rand(13,1,0.8)
As = cvxopt_to_numpy_matrix(A)
bs = cvxopt_to_numpy_matrix(matrix(b))
result = minres(As,bs,)
x = numpy_to_cvxopt_matrix(result[0])
print nrm2(A*x-b)
def solve(self, b, maxiter=1000, tol=1.e-10):
"""
Compute Q^{-1}b
Parameters:
-----------
b: (n,) ndarray
given right hand side
maxiter: int, optional. default = 1000
Maximum number of iterations for the linear solver
tol: float, optional. default = 1.e-10
Residual stoppingtolerance for the iterative solver
Notes:
------
If 'Dense' then inverts using LU factorization otherwise uses iterative solver
- MINRES if used without preconditioner or GMRES with preconditioner.
Preconditioner is not guaranteed to be positive definite.
"""
if self.method == 'Dense':
from scipy.linalg import solve
x = solve(self.mat, b)
else:
from scipy.sparse.linalg import gmres, aslinearoperator, minres
P = self.P
Aop = aslinearoperator(self)
residual = _Residual()
if P != None:
x, info = gmres(Aop,
b,
tol=tol,
restart=30,
maxiter=1000,
callback=residual,
M=P)
else:
x, info = minres(Aop,
b,
tol=tol,
maxiter=maxiter,
callback=residual)
self.solvmatvecs += residual.itercount()
if self.verbose:
print
"Number of iterations is %g and status is %g" % (
residual.itercount(), info)
return x
def x_star(self):
if not hasattr(self, 'x_opt'):
try:
# use the Cholesky factorization to solve the linear system if Q is
# symmetric and positive definite, i.e., the function is strictly convex
self.x_opt = cho_solve(cho_factor(self.Q), -self.q)
except np.linalg.LinAlgError:
# since Q is is not strictly psd, i.e., the function is linear along the
# eigenvectors correspondent to the null eigenvalues, the system has infinite
# solutions, so we will choose the one that minimizes the residue
self.x_opt = minres(self.Q, -self.q)[0]
return self.x_opt
def test(mesh, normalize=False, scipy_solver='lu'):
V = FunctionSpace(mesh, 'Lagrange', 1)
bc= DirichletBC(V, Constant(0.0), DomainBoundary())
u =TrialFunction(V)
v = TestFunction(V)
f = Constant(1.0)
a = inner(grad(u), grad(v))*dx
L = f*v*dx
A, b = assemble_system(a, L, bc)
if normalize:
max = A.array().max()
A /= max
b /= max
u = Function(V)
solve(A, u.vector(), b)
plot(u, interactive=True, title='dolfin')
x = A.array()
print "Value of A are [%g, %g]" %(x.min(), x.max())
print "Num of entries in A larger theb 100", np.where(x > 1E2)[0].sum()
print "Number of mesh cells", mesh.num_cells()
# see about scipy
rows, cols, values = A.data()
B = csr_matrix((values, cols, rows))
d = b.array().T
if scipy_solver == 'cg':
v_, info = cg(B, d)
elif scipy_solver == 'bicg':
v_, info = bicg(B, d)
elif scipy_solver == 'bicgstab':
v_, info = bicgstab(B, d)
elif scipy_solver == 'gmres':
v_, info = gmres(B, d)
elif scipy_solver == 'minres':
v_, info = minres(B, d)
else:
v_ = spsolve(B, d)
v = Function(V)
v.vector()[:] = v_
plot(v, interactive=True, title='scipy')
try:
print "info", info, 'v_max', v_.max()
except:
pass
def test(mesh, normalize=False, scipy_solver='lu'):
V = FunctionSpace(mesh, 'Lagrange', 1)
bc = DirichletBC(V, Constant(0.0), DomainBoundary())
u = TrialFunction(V)
v = TestFunction(V)
f = Constant(1.0)
a = inner(grad(u), grad(v)) * dx
L = f * v * dx
A, b = assemble_system(a, L, bc)
if normalize:
max = A.array().max()
A /= max
b /= max
u = Function(V)
solve(A, u.vector(), b)
plot(u, interactive=True, title='dolfin')
x = A.array()
print "Value of A are [%g, %g]" % (x.min(), x.max())
print "Num of entries in A larger theb 100", np.where(x > 1E2)[0].sum()
print "Number of mesh cells", mesh.num_cells()
# see about scipy
rows, cols, values = A.data()
B = csr_matrix((values, cols, rows))
d = b.array().T
if scipy_solver == 'cg':
v_, info = cg(B, d)
elif scipy_solver == 'bicg':
v_, info = bicg(B, d)
elif scipy_solver == 'bicgstab':
v_, info = bicgstab(B, d)
elif scipy_solver == 'gmres':
v_, info = gmres(B, d)
elif scipy_solver == 'minres':
v_, info = minres(B, d)
else:
v_ = spsolve(B, d)
v = Function(V)
v.vector()[:] = v_
plot(v, interactive=True, title='scipy')
try:
print "info", info, 'v_max', v_.max()
except:
pass
def iterative(self, subtype):
assert subtype in ['cg', 'gmres', 'minres']
A = sparse.csr_matrix(self.matrix_a)
b = self.vector_b
counter = method_counter()
if subtype == 'gmres':
(x, info) = linalg.gmres(A, b, callback=counter)
elif subtype == 'minres':
(x, info) = linalg.minres(A, b, callback=counter)
elif subtype == 'cg':
(x, info) = linalg.cg(A, b, callback=counter)
print(counter.niter)
return x
def minres_solve(A, u, b, tol=1e-08):
print("Solving system using MINRES solver")
""" Solves the linear system A*u = b
Input
A: numpy array of NxN components (LHS)
b: numpy array of Nx1 components (RHS)
u: numpy array of Nx1 components (Solution)
"""
# Change RHS representation
A = sp.csr_matrix(A)
# Solve system
u[:] = spla.minres(A, b, tol=tol)[0]
def solveSchurDiag(self, res ):
'''Solves diagonal of Schur matrix
Useful for Block Jacobi'''
res = res.reshape([self.n, self.m])
out = zeros_like(res)
for i in xrange(self.n):
diag_op = splinalg.LinearOperator( (self.m, self.m) ,matvec = lambda x: self.schurDiag(i,x), dtype = float )
x0 = out[i-1] if i>0 else None
ans, err = splinalg.minres(diag_op, res[i], x0=x0)
out[i] = ans
return out.ravel()
def _solve_CG(self, X, Y):
"""
Solve the primal SVM problem with Newton method without computing hessina matrix explicit,
good for big sparse matrix
:param X:
:param Y:
"""
[n, d] = X.shape
# we add one last component, which is b (bias)
self.w = np.zeros(d + 1)
# helper variable for storing 1-Y*(np.dot(X,w))
self.out = np.ones(n)
l = self.l2reg
# the number of alg. iteration
iter = 0
sv = np.where(self.out > 0)[0]
# create linear operator, acts as matrix vector multiplication, without storing full matrix(hessian)
#hess_vec = linalg.LinearOperator((d + 1, d + 1), matvec=self._matvec_mull)
# This is a hack in order to pass additional parameters to linear matvec function
mv2 = lambda v: self._matvec_mull(v, sv)
# create linear operator, acts as matrix vector multiplication, without storing full matrix(hessian)
hess_vec = linalg.LinearOperator((d + 1, d + 1), matvec=mv2)
while True:
iter = iter + 1
if iter > self.newton_iter:
print("Maximum {0} of Newton steps reached, change newton_iter parameter or try larger lambda".format(
iter))
break
obj, grad = self._obj_func(self.w, X, Y, self.out)
# np.where returns a tuple, we take the first dim
sv = np.where(self.out > 0)[0]
step, info = linalg.minres(hess_vec, -grad)
t, self.out = self._line_search(self.w, step, self.out)
self.w += t * step
if -step.dot(grad) < self._prec * obj:
break
def task_dfs_loop(i, qs, alpha, W, tol, maxiter, top_k):
w_idxs, W_trunk = find_trunc_graph(qs, W, 2)
Wn = normalize_connection_graph(W_trunk)
Wnn = eye(Wn.shape[0]) - alpha * Wn
f, inf = s_linalg.minres(Wnn,
qs[w_idxs].toarray(),
tol=tol,
maxiter=maxiter)
ranks = w_idxs[np.argsort(-f.reshape(-1))]
missing = np.setdiff1d(np.arange(W.shape[1]), ranks)
cur_rank = np.concatenate(
[ranks.reshape(-1, 1), missing.reshape(-1, 1)], axis=0)
if top_k is not None:
cur_rank = cur_rank[:top_k]
return i, cur_rank
def test_minres_scipy():
H = sprandsym(10)
v = sp_rand(10, 3, 0.8)
A = sparse([[H], [v]])
vrow = sparse([[v.T], [spmatrix([], [], [], (3, 3))]])
A = sparse([A, vrow])
b = sp_rand(13, 1, 0.8)
As = cvxopt_to_numpy_matrix(A)
bs = cvxopt_to_numpy_matrix(matrix(b))
result = minres(
As,
bs,
)
x = numpy_to_cvxopt_matrix(result[0])
print nrm2(A * x - b)
def _arrlist_lsq_solve(fA, b, iter=10):
'''
Min residual method on lists of numpy arrays.
fA is a function that takes an array list and outputs an array list.
b is an array list. x = fA^-1 b
https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.linalg.minres.html#scipy.sparse.linalg.minres
'''
x = [np.zeros_like(bb) for bb in b]
#shape_arr = [bb.shape for bb in b]
#for shapea in shape_arr: assert shapea==shape_arr[0]
fAOp = LinearOperator((-1, -1), matvec=fA)
x, info = minres(
fAOp, b, x0=x, shift=1e-8, tol=1e-05
) #, maxiter=iter, xtype=None, M=None, callback=None, show=False, check=False)
#r=(0.0,1e3)[info==0]
return x, np.abs(info)
def cg_wrapper(operator, rhs, x0 = None):
"""
do reshapes, and call cg; is that all there is to it?
if rhs and x0 are in boundified subspace, we should not need to reapply this constraint
"""
if not x0 is None: x0=np.ravel(x0)
shape = rhs.shape
from scipy.sparse.linalg import cg, LinearOperator, minres, gmres
def flat_operator(x):
return np.ravel( operator(x.reshape(shape)))
N = np.prod(shape)
wrapped_operator = LinearOperator((N,N), flat_operator, dtype=np.float)
y = minres(wrapped_operator, np.ravel(rhs), x0, maxiter = 10)[0]
return y.reshape(shape)
def rqi(A=None, x=None, k=None):
from numpy.linalg import norm, solve
from numpy import dot, eye
from tracemin_fiedler import cg
from scipy.sparse.linalg import minres
for j in range(k):
u = x / norm(x) # normalize
lam = dot(u, A * u) # Rayleigh quotient
#B = A - lam * eye(A.shape[0], A.shape[1])
#x = solve(B,u) # inverse power iteration
#D = scipy.sparse.dia_matrix((1.0/(A.diagonal()-lam), 0), shape=A.shape)
x, flag = minres(A, u, tol=1e-5, maxiter=30, shift=lam)
x = x / norm(x)
lam = dot(x, A * x) # Rayleigh quotient
return [lam, x]
def rqi(A=None, x=None, k=None):
from numpy.linalg import norm, solve
from numpy import dot, eye
from tracemin_fiedler import cg
from scipy.sparse.linalg import minres
for j in range(k):
u = x/norm(x) # normalize
lam = dot(u,A*u) # Rayleigh quotient
#B = A - lam * eye(A.shape[0], A.shape[1])
#x = solve(B,u) # inverse power iteration
#D = scipy.sparse.dia_matrix((1.0/(A.diagonal()-lam), 0), shape=A.shape)
x,flag = minres(A,u,tol=1e-5,maxiter=30,shift=lam)
x = x/norm(x)
lam = dot(x,A*x) # Rayleigh quotient
return [lam,x]
def in_sample_kfoldcv(self, folds, maxiter = None):
"""
Computes the in-sample k-fold cross-validation predictions. By in-sample we denote the
setting, where we leave a set of arbitrary entries of Y out at a time.
Returns
-------
F : array, shape = [n_samples1*n_samples2]
Training set labels. Label for (X1[i], X2[j]) maps to
F[i + j*n_samples1] (column order).
"""
if not self.kernelmode:
X1, X2 = self.X1, self.X2
P = X1 @ self.W @ X2.T
R1 = la.inv(X1.T @ X1 + self.regparam1 * np.eye(X1.shape[1])) @ X1.T
R2 = la.inv(X2.T @ X2 + self.regparam2 * np.eye(X2.shape[1])) @ X2.T
else:
P = self.K1 @ self.A @ self.K2.T
H1 = self.K1 @ la.inv(self.K1 + self.regparam1 * np.eye(self.K1.shape[0]))
H2 = self.K2 @ la.inv(self.K2 + self.regparam2 * np.eye(self.K2.shape[0]))
allhopreds = np.zeros(self.Y.shape)
for fold in folds:
row_inds_K1, row_inds_K2 = fold
if not self.kernelmode:
u_inds_1, i_inds_1 = np.unique(row_inds_K1, return_inverse = True)
r_inds_1 = np.arange(len(u_inds_1))[i_inds_1]
H1_ho = X1[u_inds_1] @ R1[:, u_inds_1]
u_inds_2, i_inds_2 = np.unique(row_inds_K2, return_inverse = True)
r_inds_2 = np.arange(len(u_inds_2))[i_inds_2]
H2_ho = X2[u_inds_2] @ R2[:, u_inds_2]
pko = PairwiseKernelOperator(H1_ho, H2_ho, r_inds_1, r_inds_2, r_inds_1, r_inds_2)
else:
pko = PairwiseKernelOperator(H1, H2, row_inds_K1, row_inds_K2, row_inds_K1, row_inds_K2)
temp = P[row_inds_K1, row_inds_K2]
temp -= np.array(pko.matvec(np.array(self.Y)[row_inds_K1, row_inds_K2].squeeze())).squeeze()
def mv(v):
return v - pko.matvec(v)
G = LinearOperator((len(row_inds_K1), len(row_inds_K1)), matvec = mv, dtype = np.float64)
hopred = minres(G, temp.T, tol=1e-20, maxiter = maxiter)[0]
allhopreds[row_inds_K1, row_inds_K2] = hopred
return allhopreds.ravel(order = 'F')
def solve(self, b, maxiter = 1000, tol = 1.e-10):
"""
Compute Q^{-1}b
Parameters:
-----------
b: (n,) ndarray
given right hand side
maxiter: int, optional. default = 1000
Maximum number of iterations for the linear solver
tol: float, optional. default = 1.e-10
Residual stoppingtolerance for the iterative solver
Notes:
------
If 'Dense' then inverts using LU factorization otherwise uses iterative solver
- MINRES if used without preconditioner or GMRES with preconditioner.
Preconditioner is not guaranteed to be positive definite.
"""
if self.method == 'Dense':
from scipy.linalg import solve
x = solve(self.mat, b)
else:
from scipy.sparse.linalg import gmres, aslinearoperator, minres
P = self.P
Aop = aslinearoperator(self)
residual = _Residual()
if P != None:
x, info = gmres(Aop, b, tol = tol, restart = 30, maxiter = 1000, callback = residual, M = P)
else:
x, info = minres(Aop, b, tol = tol, maxiter = maxiter, callback = residual )
self.solvmatvecs += residual.itercount()
if self.verbose:
print "Number of iterations is %g and status is %g"% (residual.itercount(), info)
return x
def _update_g(self, h, lamb, **kwargs):
"""
Evaluate the g-update proximal map for ADMM.
input:
h : g-like array
lamb : lagrange multiplier enforcing D.dot(g) = z
(same shape as z)
kwargs are passed to spl.minres. tol and maxiter control how hard it
tries
output:
updated g : (N_pad)-shaped
"""
self._oldg = self._oldg if self._oldg is not None else np.zeros(self.N_pad)
oldAg = self.A.dot(self._oldg)
self._c = self._Mtphi/self.sigma**2 + lamb + self.rho * h - oldAg
maxiter = kwargs.get('maxiter', 200)
tol = kwargs.get('tol', 1E-12)
self._gminsol = spl.minres(self.A, self._c, maxiter = maxiter, tol = tol)
self._newg = self._gminsol[0] + self._oldg
self._oldg = self._newg.copy()
return self._newg
def solve_factorized_aug(z, Fval, LU, G, A):
M, N=G.shape
P, N=A.shape
"""Total number of inequality constraints"""
m=M
"""Primal variable"""
x=z[0:N]
"""Multiplier for equality constraints"""
nu=z[N:N+P]
"""Multiplier for inequality constraints"""
l=z[N+P:N+P+M]
"""Slacks"""
s=z[N+P+M:]
"""Dual infeasibility"""
rd = Fval[0:N]
"""Primal infeasibility"""
rp1 = Fval[N:N+P]
rp2 = Fval[N+P:N+P+M]
"""Centrality"""
rc = Fval[N+P+M:]
"""Sigma matrix"""
SIG = diags(l/s, 0)
"""LU is actually the augmented system W"""
W = LU
b1 = -rd - mydot(G.T, mydot(SIG, rp2)) + mydot(G.T, rc/s)
b2 = -rp1
b = np.hstack((b1, b2))
"""Prepare iterative solve via MINRES"""
sign = np.zeros(N+P)
sign[0:N/2] = 1.0
sign[N/2:] = -1.0
T = diags(sign, 0)
"""Change rhs"""
b_new = mydot(T, b)
dW = np.abs(W.diagonal())
dPc = np.ones(W.shape[0])
ind = (dW > 0.0)
dPc[ind] = 1.0/dW[ind]
Pc = diags(dPc, 0)
dxnu, info = minres(W, b_new, tol=1e-10, M=Pc)
# dxnu = solve(J, b)
dx = dxnu[0:N]
dnu = dxnu[N:]
"""Obtain search directions for l and s"""
ds = -rp2 - mydot(G, dx)
# ds = s*ds
# SIG = np.diag(l/s)
dl = -mydot(SIG, ds) - rc/s
dz = np.hstack((dx, dnu, dl, ds))
return dz
def __init__(self, **kwargs):
self.Y = kwargs["Y"]
#self.Y = array_tools.as_2d_array(Y)
self.trained = False
if "regparam" in kwargs:
self.regparam = kwargs["regparam"]
else:
self.regparam = 0.
regparam = self.regparam
if CALLBACK_FUNCTION in kwargs:
self.callbackfun = kwargs[CALLBACK_FUNCTION]
else:
self.callbackfun = None
if "compute_risk" in kwargs:
self.compute_risk = kwargs["compute_risk"]
else:
self.compute_risk = False
if 'K1' in kwargs or 'pko' in kwargs:
if 'pko' in kwargs:
pko = kwargs['pko']
else:
self.input1_inds = np.array(kwargs["label_row_inds"], dtype = np.int32)
self.input2_inds = np.array(kwargs["label_col_inds"], dtype = np.int32)
K1 = kwargs['K1']
K2 = kwargs['K2']
if 'weights' in kwargs: weights = kwargs['weights']
else: weights = None
pko = pairwise_kernel_operator.PairwiseKernelOperator(K1, K2, self.input1_inds, self.input2_inds, self.input1_inds, self.input2_inds, weights)
self.pko = pko
if 'maxiter' in kwargs: maxiter = int(kwargs['maxiter'])
else: maxiter = None
Y = np.array(self.Y).ravel(order = 'F')
self.bestloss = float("inf")
def mv(v):
return pko.matvec(v) + regparam * v
def mvr(v):
raise Exception('This function should not be called!')
def cgcb(v):
if self.compute_risk:
P = sampled_kronecker_products.sampled_vec_trick(v, K2, K1, self.input2_inds, self.input1_inds, self.input2_inds, self.input1_inds)
z = (Y - P)
Ka = sampled_kronecker_products.sampled_vec_trick(v, K2, K1, self.input2_inds, self.input1_inds, self.input2_inds, self.input1_inds)
loss = (np.dot(z, z) + regparam * np.dot(v, Ka))
print("loss", 0.5 * loss)
if loss < self.bestloss:
self.A = v.copy()
self.bestloss = loss
else:
self.A = v
if not self.callbackfun is None:
#self.predictor = KernelPairwisePredictor(self.A, self.input1_inds, self.input2_inds, self.pko.weights)
self.callbackfun.callback(self)
G = LinearOperator((self.Y.shape[0], self.Y.shape[0]), matvec = mv, rmatvec = mvr, dtype = np.float64)
self.A = minres(G, self.Y, maxiter = maxiter, callback = cgcb, tol=1e-20)[0]
self.predictor = KernelPairwisePredictor(self.A, self.pko.original_col_inds_K1, self.pko.original_col_inds_K2, self.pko.weights)
if not self.callbackfun is None:
self.callbackfun.finished(self)
else:
self.input1_inds = np.array(kwargs["label_row_inds"], dtype = np.int32)
self.input2_inds = np.array(kwargs["label_col_inds"], dtype = np.int32)
X1 = kwargs['X1']
X2 = kwargs['X2']
self.X1, self.X2 = X1, X2
if 'maxiter' in kwargs: maxiter = int(kwargs['maxiter'])
else: maxiter = None
if 'weights' in kwargs: weights = kwargs['weights']
else: weights = None
Y = np.array(self.Y).ravel(order = 'F')
self.bestloss = float("inf")
def mv(v):
v_after = pko.matvec(v)
v_after = pko.rmatvec(v_after) + regparam * v
return v_after
def cgcb(v):
if self.compute_risk:
P = sampled_kronecker_products.sampled_vec_trick(v, X2, X1, self.input2_inds, self.input1_inds)
z = (Y - P)
loss = (np.dot(z,z)+regparam*np.dot(v,v))
if loss < self.bestloss:
self.W = v.copy().reshape(pko.shape, order = 'F')
self.bestloss = loss
else:
self.W = v
if not self.callbackfun is None:
self.predictor = LinearPairwisePredictor(self.W)
self.callbackfun.callback(self)
v_init = np.array(self.Y).reshape(self.Y.shape[0])
pko = pairwise_kernel_operator.PairwiseKernelOperator(X1, X2, self.input1_inds, self.input2_inds, None, None, weights)
G = LinearOperator((pko.shape[1], pko.shape[1]), matvec = mv, dtype = np.float64)
v_init = pko.rmatvec(v_init)
'''if 'warm_start' in kwargs:
x0 = np.array(kwargs['warm_start']).reshape(kronfcount, order = 'F')
else:
x0 = None'''
minres(G, v_init, maxiter = maxiter, callback = cgcb, tol=1e-20)#[0].reshape((pko_T.shape[0], pko.shape[1]), order='F')
self.predictor = LinearPairwisePredictor(self.W, self.input1_inds, self.input2_inds, weights)
if not self.callbackfun is None:
self.callbackfun.finished(self)
self.pde = pde
self.hist = []
def __call__(self, x):
self.n += 1
if self.n == 1 or self.n % 10 == 0:
resnorm = norm(self.pde.matvec(x, 1))
gradient = kuramoto.c_gradAdj()
print 'iter ', self.n, resnorm, gradient
self.hist.append([self.n, resnorm, gradient])
sys.stdout.flush()
# --- solve with minres (if cg converges this should converge -#
callback = Callback(pde)
callback(rhs * 0)
vw, info = splinalg.minres(oper, rhs, maxiter=100, tol=1E-6,
callback=callback)
pde.matvec(vw, 1)
u, v, w, v0, uEnd = [], [], [], [], []
for i in range(kuramoto.cvar.N_CHUNK):
for j in range(kuramoto.cvar.N_STEP):
u.append(kuramoto.c_u(i, j))
v.append(kuramoto.c_v(i, j))
w.append(kuramoto.c_w(i, j))
v0.append(v[-1].copy())
kuramoto.c_project_ddt(i, j, v0[-1])
uEnd.append(kuramoto.c_u(i, kuramoto.cvar.N_STEP))
u, v, w, v0, uEnd = array(u), array(v), array(w), array(v0), array(uEnd)
def __init__(self, **kwargs):
self.resource_pool = kwargs
Y = kwargs["Y"]
self.input1_inds = np.array(kwargs["label_row_inds"], dtype=np.int32)
self.input2_inds = np.array(kwargs["label_col_inds"], dtype=np.int32)
Y = array_tools.as_2d_array(Y)
self.Y = np.mat(Y)
self.trained = False
if kwargs.has_key("regparam"):
self.regparam = kwargs["regparam"]
else:
self.regparam = 0.0
if kwargs.has_key(CALLBACK_FUNCTION):
self.callbackfun = kwargs[CALLBACK_FUNCTION]
else:
self.callbackfun = None
if kwargs.has_key("compute_risk"):
self.compute_risk = kwargs["compute_risk"]
else:
self.compute_risk = False
regparam = self.regparam
if self.resource_pool.has_key("K1"):
K1 = self.resource_pool["K1"]
K2 = self.resource_pool["K2"]
if "maxiter" in self.resource_pool:
maxiter = int(self.resource_pool["maxiter"])
else:
maxiter = None
Y = np.array(self.Y).ravel(order="F")
self.bestloss = float("inf")
def mv(v):
return (
sampled_kronecker_products.sampled_vec_trick(
v, K2, K1, self.input2_inds, self.input1_inds, self.input2_inds, self.input1_inds
)
+ regparam * v
)
def mvr(v):
raise Exception("You should not be here!")
def cgcb(v):
if self.compute_risk:
P = sampled_kronecker_products.sampled_vec_trick(
v, K2, K1, self.input2_inds, self.input1_inds, self.input2_inds, self.input1_inds
)
z = Y - P
Ka = sampled_kronecker_products.sampled_vec_trick(
v, K2, K1, self.input2_inds, self.input1_inds, self.input2_inds, self.input1_inds
)
loss = np.dot(z, z) + regparam * np.dot(v, Ka)
print "loss", 0.5 * loss
if loss < self.bestloss:
self.A = v.copy()
self.bestloss = loss
else:
self.A = v
if not self.callbackfun is None:
self.predictor = KernelPairwisePredictor(self.A, self.input1_inds, self.input2_inds)
self.callbackfun.callback(self)
G = LinearOperator((len(self.input1_inds), len(self.input1_inds)), matvec=mv, rmatvec=mvr, dtype=np.float64)
self.A = minres(G, self.Y, maxiter=maxiter, callback=cgcb, tol=1e-20)[0]
self.predictor = KernelPairwisePredictor(self.A, self.input1_inds, self.input2_inds)
else:
X1 = self.resource_pool["X1"]
X2 = self.resource_pool["X2"]
self.X1, self.X2 = X1, X2
if "maxiter" in self.resource_pool:
maxiter = int(self.resource_pool["maxiter"])
else:
maxiter = None
x1tsize, x1fsize = X1.shape # m, d
x2tsize, x2fsize = X2.shape # q, r
kronfcount = x1fsize * x2fsize
Y = np.array(self.Y).ravel(order="F")
self.bestloss = float("inf")
def mv(v):
v_after = sampled_kronecker_products.sampled_vec_trick(v, X2, X1, self.input2_inds, self.input1_inds)
v_after = (
sampled_kronecker_products.sampled_vec_trick(
v_after, X2.T, X1.T, None, None, self.input2_inds, self.input1_inds
)
+ regparam * v
)
return v_after
def mvr(v):
raise Exception("You should not be here!")
return None
def cgcb(v):
if self.compute_risk:
P = sampled_kronecker_products.sampled_vec_trick(v, X2, X1, self.input2_inds, self.input1_inds)
z = Y - P
loss = np.dot(z, z) + regparam * np.dot(v, v)
if loss < self.bestloss:
self.W = v.copy().reshape((x1fsize, x2fsize), order="F")
self.bestloss = loss
else:
self.W = v.reshape((x1fsize, x2fsize), order="F")
if not self.callbackfun is None:
self.predictor = LinearPairwisePredictor(self.W)
self.callbackfun.callback(self)
G = LinearOperator((kronfcount, kronfcount), matvec=mv, rmatvec=mvr, dtype=np.float64)
v_init = np.array(self.Y).reshape(self.Y.shape[0])
v_init = sampled_kronecker_products.sampled_vec_trick(
v_init, X2.T, X1.T, None, None, self.input2_inds, self.input1_inds
)
v_init = np.array(v_init).reshape(kronfcount)
if self.resource_pool.has_key("warm_start"):
x0 = np.array(self.resource_pool["warm_start"]).reshape(kronfcount, order="F")
else:
x0 = None
minres(G, v_init, x0=x0, maxiter=maxiter, callback=cgcb, tol=1e-20)[0].reshape(
(x1fsize, x2fsize), order="F"
)
self.predictor = LinearPairwisePredictor(self.W)
if not self.callbackfun is None:
self.callbackfun.finished(self)
from pdas.prob import randQP
from pdas.convert import numpy_to_cvxopt_matrix, cvxopt_to_numpy_matrix
import ctypes
import inspect
class LS(object):
'Class holder for linear equation object.'
def __init__(self):
qp = randQP(10)
self.pdas = PDAS(qp)
self.Lhs, self.rhs, self.x0 = self.pdas._get_lineq()
self.it = 0
def __call__(self,x):
# Access value of itn in minres
frame = inspect.currentframe().f_back
self.it = frame.f_locals['itn']
# Change value of
# ctype.pythonapi.Pycell_Set(id(inner.func_closure[0]), id(x))
print('solution updated to: ', self.it)
if __name__=='__main__':
a = LS()
Lhs = cvxopt_to_numpy_matrix(a.Lhs)
rhs = cvxopt_to_numpy_matrix(matrix(a.rhs))
x0 = cvxopt_to_numpy_matrix(matrix(a.x0))
minres(Lhs,rhs,x0,callback = a)
def solve_factorized(KKTval, z, LU, A0, Ak, Ak0, G):
r"""Compute the Newton increment for a DKKT system with prefactored
LHS.
"""
"""Dimensions of subproblems"""
n = A0.shape[1] # Primal point
p_E0 = A0.shape[0] # Equality constraints alpha=0
p_E = Ak.shape[0] # Equality constraints alpha>0
m = G.shape[0] # Inequality constraints
"""Total number of subproblems, alpha=0...K-1"""
K = (z.shape[0] - (n+p_E0+m+m))/(n+p_E+m+m) + 1
"""Total dimensions of subvectors"""
N = K*n # Primal
P = p_E0 + (K-1)*p_E # Multipliers equality
M = K*m # Multipliers/slacks inequality
"""Multiplier for inequality constraints"""
l = z[N+P:N+P+M]
"""Slacks for inequality constraints"""
s = z[N+P+M:]
"""Dual residuum"""
rd = KKTval[0:N]
"""Primal residuum (equalities)"""
rp1 = KKTval[N:N+P]
"""Primal residuum (inequalities)"""
rp2 = KKTval[N+P:N+P+M]
"""Complementarity"""
rc = KKTval[N+P+M:]
"""Unpack decomposition"""
S, LU_W = LU
"""Set up extended block containing coupling to augmented system
at alpha=0"""
B = bmat([[None, csr_matrix((n, p_E0))], [Ak0, None]], format='csr')
"""Diagonal of Sigma matrix"""
sig = l/s
"""Compute RHS of augmented system for all alpha and assemble RHS
of condensed system, store intermediate results W_k^{-1}b_k, k>0, for
later use"""
Winvb = []
for k in range(K):
"""Extract subvectors"""
sk = s[k*m:(k+1)*m]
sigk = sig[k*m:(k+1)*m]
rdk = rd[k*n:(k+1)*n]
rp2k = rp2[k*m:(k+1)*m]
rck = rc[k*m:(k+1)*m]
if k==0:
rp1k = rp1[0:p_E0]
else:
rp1k = rp1[p_E0+(k-1)*p_E:p_E0+k*p_E]
"""RHS of augmented system"""
b1 = rdk + mydot(G.T, sigk*rp2k) - mydot(G.T, rck/sk)
b2 = rp1k
b = np.hstack((b1, b2))
if k==0:
"""RHS of condensed system"""
rhs0 = b
else:
"""Compute W_k^{-1}b_k"""
Winvbk = mysolve(LU_W[k-1], b)
"""Update RHS of condensed system"""
rhs0 -= mydot(B.T, Winvbk)
"""Store W_k^{-1}b_k"""
Winvb.append(Winvbk)
"""Increment subvectors for return"""
dx = np.zeros(N)
dnu = np.zeros(P)
dl = np.zeros(M)
ds = np.zeros(M)
"""Set up the symmetrization matrix"""
T0 = diags(np.hstack((np.ones(n/2), -1.0*np.ones(n/2+p_E0))), 0)
rhs0sym = T0.dot(rhs0)
"""Set up symmetric Schur complement as LinearOperator"""
def matvec(x):
return T0.dot(S.dot(x))
Ssym = LinearOperator(S.shape, matvec=matvec)
"""Set up Jacobi preconditioning for Ssym"""
dSsym = np.abs(probe(Ssym, k=1))
dPc = np.ones(Ssym.shape[0])
ind = (dSsym > 0.0)
dPc[ind] = 1.0/dSsym[ind]
Pc = diags(dPc, 0)
mycounter = MyCounter()
"""Compute increment subvectors and assign"""
for k in range(K):
if k==0:
dy0, info = minres(Ssym, -rhs0sym, tol=1e-13, M=Pc, callback=mycounter.count)
# print mycounter.N
dy = dy0
dnu[0:p_E0] = dy0[n:]
else:
dy = -Winvb[k-1] - mysolve(LU_W[k-1], mydot(B, dy0))
dnu[p_E0+(k-1)*p_E:p_E0+k*p_E] = dy[n:]
dx[k*n:(k+1)*n] = dy[0:n]
ds[k*m:(k+1)*m] = -rp2[k*m:(k+1)*m] - mydot(G, dx[k*n:(k+1)*n])
dl[k*m:(k+1)*m] = -sig[k*m:(k+1)*m]*ds[k*m:(k+1)*m] - rc[k*m:(k+1)*m]/s[k*m:(k+1)*m]
return np.hstack((dx, dnu, dl, ds))
def solve_full(z, Fval, DPhival, G, A):
M, N=G.shape
P, N=A.shape
"""Total number of inequality constraints"""
m=M
"""Primal variable"""
x=z[0:N]
"""Multiplier for equality constraints"""
nu=z[N:N+P]
"""Multiplier for inequality constraints"""
l=z[N+P:N+P+M]
"""Slacks"""
s=z[N+P+M:]
"""Dual infeasibility"""
rd = Fval[0:N]
"""Primal infeasibility"""
rp1 = Fval[N:N+P]
rp2 = Fval[N+P:N+P+M]
"""Centrality"""
rc = Fval[N+P+M:]
"""Sigma matrix"""
SIG = np.diag(l/s)
"""Condensed system"""
if issparse(DPhival):
if not issparse(A):
A = csr_matrix(A)
H = DPhival + mydot(G.T, mydot(SIG, G))
J = bmat([[H, A.T], [A, None]])
else:
if issparse(A):
A = A.toarray()
J = np.zeros((N+P, N+P))
J[0:N, 0:N] = DPhival + mydot(G.T, mydot(SIG, G))
J[0:N, N:] = A.T
J[N:, 0:N] = A
b1 = -rd - mydot(G.T, mydot(SIG, rp2)) + mydot(G.T, rc/s)
b2 = -rp1
b = np.hstack((b1, b2))
"""Prepare iterative solve via MINRES"""
sign = np.zeros(N+P)
sign[0:N/2] = 1.0
sign[N/2:] = -1.0
S = diags(sign, 0)
J_new = mydot(S, csr_matrix(J))
b_new = mydot(S, b)
dJ_new = np.abs(J_new.diagonal())
dPc = np.ones(J_new.shape[0])
ind = (dJ_new > 0.0)
dPc[ind] = 1.0/dJ_new[ind]
Pc = diags(dPc, 0)
dxnu, info = minres(J_new, b_new, tol=1e-8, M=Pc)
# dxnu = solve(J, b)
dx = dxnu[0:N]
dnu = dxnu[N:]
"""Obtain search directions for l and s"""
ds = -rp2 - mydot(G, dx)
dl = -mydot(SIG, ds) - rc/s
dz = np.hstack((dx, dnu, dl, ds))
return dz
mua = np.ones ((1,nlen)) * 0.025
mus = np.ones ((1,nlen)) * 2.0
ref = np.ones ((1,nlen)) * 1.4
freq = 100
# Set up the linear system
smat = mesh.Sysmat (mua, mus, ref, freq)
qvec = mesh.Qvec ()
mvec = mesh.Mvec ()
# Solve the linear system
nq = qvec.shape[1]
phi = np.empty(qvec.shape,dtype='complex128')
for q in range(nq):
qq = qvec[:,q].todense()
res = linalg.minres(smat,qq,tol=1e-12)
phi[:,q] = res[0]
# Project to boundary
y = mvec.transpose() * phi
logy = np.log(y)
# Display as sinogram
plt.figure(1)
im = plt.imshow(logy.real,interpolation='none')
plt.title('log amplitude')
plt.xlabel('detector index')
plt.ylabel('source index')
plt.colorbar()
plt.draw()
#plt.show()
def solveDiagBlock(self, i, rhs, tol=1e-4, x0=None):
Mi = self.diagBlockOperator(i)
ans, err = splinalg.minres(Mi, rhs, tol=tol, x0=x0)
return ans
def _action(self, act=1):
if QSMODE == MODE_MPMATH:
args = {}
if act==0:
self.typeStr = "mpmath_qr_solve_dps"+getArgDesc(mpmath.qr_solve, args)+" DPS"+str(DPS)
self.matrixType = 1
self.indexType = 1
else:
self.coeffVec = mpmath.qr_solve(self.sysMat, self.resVec, **args)
self.printCalStr()
else:
if PYTYPE_COEFF_SOLVE_METHOD == "numpy_solve":
args = {}
if act==0:
self.typeStr = "numpy_solve"+getArgDesc(np.linalg.solve, args)
self.matrixType = 0
self.indexType = 0
else:
self.coeffVec = np.linalg.solve(self.sysMat, self.resVec, **args)
self.printCalStr()
elif PYTYPE_COEFF_SOLVE_METHOD == "numpy_lstsq":
args = {}
if act==0:
self.typeStr = "numpy_lstsq"+getArgDesc(np.linalg.lstsq, args)
self.matrixType = 0
self.indexType = 0
else:
self.coeffVec = np.linalg.lstsq(self.sysMat, self.resVec, **args)[0]
self.printCalStr()
elif PYTYPE_COEFF_SOLVE_METHOD == "numpy_sparse_bicg":
args = {}
if act==0:
self.typeStr = "numpy_sparse_bicg"+getArgDesc(sp_sparse_linalg.bicg, args)
self.matrixType = 0
self.indexType = 1
else:
self.coeffVec = self._sparseRet(sp_sparse_linalg.bicg(self.sysMat, self.resVec, **args))#, tol=1e-05, maxiter=10*len(self.resVec)
elif PYTYPE_COEFF_SOLVE_METHOD == "numpy_sparse_bicgstab":
args = {}
if act==0:
self.typeStr = "numpy_sparse_bicgstab"+getArgDesc(sp_sparse_linalg.bicgstab, args)
self.matrixType = 0
self.indexType = 1
else:
self.coeffVec = self._sparseRet(sp_sparse_linalg.bicgstab(self.sysMat, self.resVec, **args))#, tol=1e-05, maxiter=10*len(self.resVec)
elif PYTYPE_COEFF_SOLVE_METHOD == "numpy_sparse_lgmres":
args = {}
if act==0:
self.typeStr = "numpy_sparse_lgmres"+getArgDesc(sp_sparse_linalg.lgmres, args)
self.matrixType = 0
self.indexType = 1
else:
self.coeffVec = self._sparseRet(sp_sparse_linalg.lgmres(self.sysMat, self.resVec, **args))#, tol=1e-05, maxiter=1000
elif PYTYPE_COEFF_SOLVE_METHOD == "numpy_sparse_minres":
args = {}
if act==0:
self.typeStr = "numpy_sparse_minres"+getArgDesc(sp_sparse_linalg.minres, args)
self.matrixType = 0
self.indexType = 1
else:
self.coeffVec = self._sparseRet(sp_sparse_linalg.minres(self.sysMat, self.resVec, **args))#, tol=1e-05, maxiter=5*self.sysMat.shape[0]
elif PYTYPE_COEFF_SOLVE_METHOD == "numpy_sparse_qmr":
args = {}
if act==0:
self.typeStr = "numpy_sparse_qmr"+getArgDesc(sp_sparse_linalg.qmr, args)
self.matrixType = 0
self.indexType = 1
else:
self.coeffVec = self._sparseRet(sp_sparse_linalg.qmr(self.sysMat, self.resVec, **args))#, tol=1e-05, maxiter=10*len(self.resVec)
elif PYTYPE_COEFF_SOLVE_METHOD == "numpy_qr":
args_qr = {}
args_s = {}
if act==0:
self.typeStr = "numpy_qr"+getArgDesc(np.linalg.qr, args_qr)+",numpy_solve"+getArgDesc(np.linalg.solve, args_s)
self.matrixType = 0
self.indexType = 0
else:
Q,R = np.linalg.qr(self.sysMat, **args_qr)
y = np.dot(Q.T,self.resVec)
self.coeffVec = np.linalg.solve(R,y, **args_s)
self.printCalStr()
def main(write_output=True):
from hedge.data import GivenFunction, ConstantGivenFunction
from hedge.backends import guess_run_context
rcon = guess_run_context()
dim = 2
def boundary_tagger(fvi, el, fn, points):
from math import atan2, pi
normal = el.face_normals[fn]
if -90/180*pi < atan2(normal[1], normal[0]) < 90/180*pi:
return ["neumann"]
else:
return ["dirichlet"]
def dirichlet_boundary_tagger(fvi, el, fn, points):
return ["dirichlet"]
if dim == 2:
if rcon.is_head_rank:
from hedge.mesh.generator import make_disk_mesh
mesh = make_disk_mesh(r=0.5,
boundary_tagger=dirichlet_boundary_tagger,
max_area=1e-3)
elif dim == 3:
if rcon.is_head_rank:
from hedge.mesh.generator import make_ball_mesh
mesh = make_ball_mesh(max_volume=0.0001,
boundary_tagger=lambda fvi, el, fn, points:
["dirichlet"])
else:
raise RuntimeError, "bad number of dimensions"
if rcon.is_head_rank:
print "%d elements" % len(mesh.elements)
mesh_data = rcon.distribute_mesh(mesh)
else:
mesh_data = rcon.receive_mesh()
discr = rcon.make_discretization(mesh_data, order=5,
debug=[])
def dirichlet_bc(x, el):
from math import sin
return sin(10*x[0])
def rhs_c(x, el):
if la.norm(x) < 0.1:
return 1000
else:
return 0
def my_diff_tensor():
result = numpy.eye(dim)
result[0,0] = 0.1
return result
try:
from hedge.models.poisson import (
PoissonOperator,
HelmholtzOperator)
from hedge.second_order import \
IPDGSecondDerivative, LDGSecondDerivative, \
StabilizedCentralSecondDerivative
k = 1
from hedge.mesh import TAG_NONE, TAG_ALL
op = HelmholtzOperator(k, discr.dimensions,
#diffusion_tensor=my_diff_tensor(),
#dirichlet_tag="dirichlet",
#neumann_tag="neumann",
dirichlet_tag=TAG_ALL,
neumann_tag=TAG_NONE,
#dirichlet_tag=TAG_ALL,
#neumann_tag=TAG_NONE,
#dirichlet_bc=GivenFunction(dirichlet_bc),
dirichlet_bc=ConstantGivenFunction(0),
neumann_bc=ConstantGivenFunction(-10),
scheme=StabilizedCentralSecondDerivative(),
#scheme=LDGSecondDerivative(),
#scheme=IPDGSecondDerivative(),
)
bound_op = op.bind(discr)
if False:
from hedge.iterative import parallel_cg
u = -parallel_cg(rcon, -bound_op,
bound_op.prepare_rhs(discr.interpolate_volume_function(rhs_c)),
debug=20, tol=5e-4,
dot=discr.nodewise_dot_product,
x=discr.volume_zeros())
else:
rhs = bound_op.prepare_rhs(discr.interpolate_volume_function(rhs_c))
def compute_resid(x):
return bound_op(x)-rhs
from scipy.sparse.linalg import minres, LinearOperator
u, info = minres(
LinearOperator(
(len(discr), len(discr)),
matvec=bound_op, dtype=bound_op.dtype),
rhs,
callback=ResidualPrinter(compute_resid),
tol=1e-5)
print
if info != 0:
raise RuntimeError("gmres reported error %d" % info)
print "finished gmres"
print la.norm(bound_op(u)-rhs)/la.norm(rhs)
if write_output:
from hedge.visualization import SiloVisualizer, VtkVisualizer
vis = VtkVisualizer(discr, rcon)
visf = vis.make_file("fld")
vis.add_data(visf, [ ("sol", discr.convert_volume(u, kind="numpy")), ])
visf.close()
finally:
discr.close()
