def _build_staggered_first_derivative_matrix_part(npoints, order_accuracy, h=1.0, lbc="d", rbc="d"):
# npoints is the number of regular grid points.
if order_accuracy % 2:
raise ValueError("Only even accuracy orders supported.")
# coefficients for the first derivative evaluated in between two regular grid points.
stagger_coeffs = staggered_difference(1, order_accuracy) / h
# Use the old 'stencil_grid' routine.
# Because we do a staggered grid we need to shift the coeffs one entry and the matrix will not be square
incorrect_mtx = stencil_grid(np.insert(stagger_coeffs, 0, 0), (npoints,), format="lil")
# Get rid of the last row which we dont want in our staggered approach
mtx = incorrect_mtx[0:-1, :]
if "n" in lbc or "n" in rbc:
raise ValueError(
"Did not yet implement Neumann boundaries. Perhaps looking at the centered grid implementation would be a good start?"
)
if "g" in lbc or "g" in rbc:
raise ValueError(
"Did not yet implement this boundary condition yet. Perhaps looking at the centered grid implementation would be a good start?"
)
# For dirichlet we don't need to alter the matrix for the first derivative for the boundary nodes as is done in the centered approach
# The reason is that the first staggered point we evaluate at is in the interior of the domain.
return mtx.tocsr()
def build_poisson(n, epsilon, theta, randomize):
data = {}
h = 1. / float(n + 1)
print "Assembling diffusion using Q1 on a regular mesh with epsilon = " + \
str(epsilon) + " and theta = " + str(theta) + " ..."
stencil = diffusion_stencil_2d(type='FE', epsilon=epsilon, theta=theta)
A = stencil_grid(stencil, (n, n), format='csr')
X, Y = meshgrid(linspace(h, 1.0 - h, n), linspace(h, 1.0 - h, n))
data['X'] = X
data['Y'] = Y
if randomize:
print "Random diagonal scaling..."
D = my_rand(A.shape[0], 1)
D[D < 0.] -= 1e-3
D[D >= 0.] += 1e-3
data['D'] = D
D_inv = 1. / D
data['D_inv'] = D_inv
A = scale_rows(A, D)
A = scale_columns(A, D)
if symmetric_scale:
symmetric_rescaling(A, copy=False)
print "Ratio of largest to smallest (in magnitude) diagonal element in A: %1.3e"% \
(abs(A.diagonal()).max() / abs(A.diagonal()).min())
data['A'] = A
print 'Generate initial guess (which is random)...'
data['x0'] = my_rand(data['A'].shape[0], 1)
print 'Generate rhs (which is zero)...'
data['b'] = numpy.zeros((data['A'].shape[0], 1))
return data
def build_linear_interpolation_matrix_part(npoints):
#same logic as in function 'build_staggered_first_derivative_matrix_part
coeffs = np.array([0.5, 0.5])
incorrect_mtx = stencil_grid(np.insert(coeffs, 0, 0), (npoints, ),
format='lil')
mtx = incorrect_mtx[0:-1, :]
return mtx.tocsr()
def _build_staggered_first_derivative_matrix_part(npoints,
order_accuracy,
h=1.0,
lbc='d',
rbc='d'):
#npoints is the number of regular grid points.
if order_accuracy % 2:
raise ValueError('Only even accuracy orders supported.')
#coefficients for the first derivative evaluated in between two regular grid points.
stagger_coeffs = staggered_difference(1, order_accuracy) / h
#Use the old 'stencil_grid' routine.
#Because we do a staggered grid we need to shift the coeffs one entry and the matrix will not be square
incorrect_mtx = stencil_grid(np.insert(stagger_coeffs, 0, 0), (npoints, ),
format='lil')
#Get rid of the last row which we dont want in our staggered approach
mtx = incorrect_mtx[0:-1, :]
if 'n' in lbc or 'n' in rbc:
raise ValueError(
'Did not yet implement Neumann boundaries. Perhaps looking at the centered grid implementation would be a good start?'
)
if 'g' in lbc or 'g' in rbc:
raise ValueError(
'Did not yet implement this boundary condition yet. Perhaps looking at the centered grid implementation would be a good start?'
)
#For dirichlet we don't need to alter the matrix for the first derivative for the boundary nodes as is done in the centered approach
#The reason is that the first staggered point we evaluate at is in the interior of the domain.
return mtx.tocsr()
def _build_derivative_matrix_part(
npoints, derivative, order_accuracy, h=1.0, lbc="d", rbc="d", use_shifted_differences=False
):
if order_accuracy % 2:
raise ValueError("Only even accuracy orders supported.")
centered_coeffs = centered_difference(derivative, order_accuracy) / (h ** derivative)
mtx = stencil_grid(centered_coeffs, (npoints,), format="lil")
max_shift = order_accuracy / 2
if use_shifted_differences:
# Left side
odd_even_offset = 1 - derivative % 2
for i in xrange(0, max_shift):
coeffs = shifted_difference(derivative, order_accuracy, -(max_shift + odd_even_offset) + i)
mtx[i, 0 : len(coeffs)] = coeffs / (h ** derivative)
# Right side
for i in xrange(-1, -max_shift - 1, -1):
coeffs = shifted_difference(derivative, order_accuracy, max_shift + i + odd_even_offset)
mtx[i, slice(-1, -(len(coeffs) + 1), -1)] = coeffs[::-1] / (h ** derivative)
if "d" in lbc: # dirichlet
mtx[0, :] = 0
mtx[0, 0] = 1.0
elif "n" in lbc: # neumann
mtx[0, :] = 0
coeffs = shifted_difference(1, order_accuracy, -max_shift) / h
coeffs /= -1 * coeffs[0]
coeffs[0] = 0.0
mtx[0, 0 : len(coeffs)] = coeffs
elif type(lbc) is tuple and "g" in lbc[0]: # ghost
n_ghost_points = int(lbc[1])
mtx[0:n_ghost_points, :] = 0
for i in xrange(n_ghost_points):
mtx[i, i] = 1.0
if "d" in rbc:
mtx[-1, :] = 0
mtx[-1, -1] = 1.0
elif "n" in rbc:
mtx[-1, :] = 0
coeffs = shifted_difference(1, order_accuracy, max_shift) / h
coeffs /= -1 * coeffs[-1]
coeffs[-1] = 0.0
mtx[-1, slice(-1, -(len(coeffs) + 1), -1)] = coeffs[::-1]
elif type(rbc) is tuple and "g" in rbc[0]:
n_ghost_points = int(rbc[1])
mtx[slice(-1, -(n_ghost_points + 1), -1), :] = 0
for i in xrange(n_ghost_points):
mtx[-i - 1, -i - 1] = 1.0
return mtx.tocsr()
# task1.3
from pyamg.gallery.diffusion import diffusion_stencil_2d
from pyamg.gallery import stencil_grid
sten = diffusion_stencil_2d(type='FD', \
epsilon=0.001, theta=3.1416/3.0)
A = stencil_grid(sten, (100, 100), format='csr')
# task1.4
from pyamg import *
ml = smoothed_aggregation_solver(A)
# task1.5
from numpy import ones
b = ones((A.shape[0], 1))
res = []
x = ml.solve(b, tol=1e-8, \
residuals=res)
from pylab import *
semilogy(res[1:])
xlabel('iteration')
ylabel('residual norm')
title('Residual History')
show()
'iterations': 1
})
smooth = ('energy', {'maxiter': 9, 'degree': 3}) # Prolongation Smoother
classic_theta = 0.0 # Classic Strength Measure
# Drop Tolerance
evolution_theta = 4.0 # evolution Strength Measure
# Drop Tolerance
for n in nlist:
nx = n
ny = n
print "Running Grid = (%d x %d)" % (nx, ny)
# Rotated Anisotropic Diffusion Operator
stencil = diffusion_stencil_2d(type='FE', epsilon=epsilon, theta=theta)
A = stencil_grid(stencil, (nx, ny), format='csr')
# Random initial guess, zero RHS
x0 = scipy.rand(A.shape[0])
b = numpy.zeros((A.shape[0], ))
# Classic SA strength measure
ml = smoothed_aggregation_solver(A,
max_coarse=mcoarse,
coarse_solver='pinv2',
presmoother=prepost,
postsmoother=prepost,
smooth=smooth,
strength=('symmetric', {
'theta': classic_theta
}))
def build_1D_fd(deriv, order, length, delta, lbc=None, rbc=None, limit_boundary=True):
""" Builds the finite difference stencil matrix in 1D that can be kroncker producted to build higher dimensional operators.
None in the BC slot leaves the purest form of the operator.
"""
bulk_npoints = deriv + order - (1 if not deriv%2 else 0)
bulk_center = int(math.floor(bulk_npoints/2))
boundary_npoints = deriv + order
stencil = fd_coeffs(deriv, (bulk_npoints, bulk_center))
stencil[np.abs(stencil) < 1e-12] = 0.0
L = stencil_grid(stencil, (length,), format='lil')
if not limit_boundary:
L /= (delta**deriv)
return L.tocsr()
# left side
for i in range(bulk_center):
boundary_center = i
if i == 0:
if lbc != 'dirichlet':
warnings.warn('Only Dirichlet boundaries are supported in this matrix construction.')
L[i,:] = 0.0
L[0,0]=1.0
# else: #lbc == 'neumann'
# # Not sure that this is correct...neumann likely need to be patched after the time step...
# L[i,:] = 0.0
# coeffs = -fd_coeffs(1, (1+order,boundary_center))
# coeffs /= coeffs[0]
# coeffs[0] = 0.0
# L[i,0:(1+order)] = coeffs
else:
L[i,:] = 0
stencil = fd_coeffs(deriv, (boundary_npoints,boundary_center))
stencil[np.abs(stencil) < 1e-12] = 0.0
L[i,0:boundary_npoints] = stencil
# right side
print(boundary_npoints-bulk_center-1)
for i in range(-1, -(boundary_npoints-bulk_center-deriv+1), -1):
boundary_center = boundary_npoints + i
idx = i
print(i, boundary_center, idx)
if idx == -1:
if lbc != 'dirichlet':
warnings.warn('Only Dirichlet boundaries are supported in this matrix construction.')
L[idx,:] = 0.0
L[-1,-1] = 1.0
# else: #lbc == 'neumann'
# # Not sure that this is correct...neumann likely need to be patched after the time step...
# L[i,:] = 0.0
# coeffs = -fd_coeffs(1, (1+order,boundary_center))
# coeffs /= coeffs[0]
# coeffs[0] = 0.0
# L[i,0:(1+order)] = coeffs
else:
L[idx,:] = 0
stencil = fd_coeffs(deriv, (boundary_npoints,boundary_center))
stencil[np.abs(stencil)<1e-12] = 0.0
L[idx,-boundary_npoints::] = stencil
L /= (delta**deriv)
return L.tocsr()
def build_linear_interpolation_matrix_part(npoints):
#same logic as in function 'build_staggered_first_derivative_matrix_part
coeffs = np.array([0.5, 0.5])
incorrect_mtx = stencil_grid(np.insert(coeffs,0,0), (npoints, ), format='lil')
mtx = incorrect_mtx[0:-1,:]
return mtx.tocsr()
def test_evolution_strength_of_connection(self):
cases = []
# Single near nullspace candidate
stencil = [[0.0, -1.0, 0.0], [-0.001, 2.002, -0.001], [0.0, -1.0, 0.0]]
A = 1.0j*stencil_grid(stencil, (4, 4), format='csr')
B = 1.0j*np.ones((A.shape[0], 1))
B[0] = 1.2 - 12.0j
B[11] = -14.2
cases.append({'A': A.copy(), 'B': B.copy(), 'epsilon': 4.0, 'k': 2,
'proj': 'l2'})
# Multiple near nullspace candidate
B = 1.0j*np.ones((A.shape[0], 2))
B[0:-1:2, 0] = 0.0
B[1:-1:2, 1] = 0.0
B[-1, 0] = 0.0
B[11, 1] = -14.2
B[0, 0] = 1.2 - 12.0j
cases.append({'A': A.copy(), 'B': B.copy(), 'epsilon': 4.0, 'k': 2,
'proj': 'l2'})
Absr = A.tobsr(blocksize=(2, 2))
cases.append({'A': Absr.copy(), 'B': B.copy(), 'epsilon': 4.0, 'k': 2,
'proj': 'l2'})
for ca in cases:
scipy.random.seed(0) # make results deterministic
result = evolution_soc(ca['A'], ca['B'], epsilon=ca['epsilon'],
k=ca['k'], proj_type=ca['proj'],
symmetrize_measure=False)
scipy.random.seed(0) # make results deterministic
expected = reference_evolution_soc(ca['A'], ca['B'],
epsilon=ca['epsilon'],
k=ca['k'], proj_type=ca['proj'])
assert_array_almost_equal(result.todense(), expected.todense())
# Test Scale Invariance for a single candidate
A = 1.0j*poisson((5, 5), format='csr')
B = 1.0j*arange(1, A.shape[0]+1, dtype=float).reshape(-1, 1)
scipy.random.seed(0) # make results deterministic
result_unscaled = evolution_soc(A, B, epsilon=4.0, k=2,
proj_type="D_A",
symmetrize_measure=False)
# create scaled A
D = spdiags([arange(A.shape[0], 2*A.shape[0], dtype=float)],
[0], A.shape[0], A.shape[0], format='csr')
Dinv = spdiags([1.0/arange(A.shape[0], 2*A.shape[0], dtype=float)],
[0], A.shape[0], A.shape[0], format='csr')
scipy.random.seed(0) # make results deterministic
result_scaled = evolution_soc(D*A*D, Dinv*B, epsilon=4.0, k=2,
proj_type="D_A",
symmetrize_measure=False)
assert_array_almost_equal(result_scaled.todense(),
result_unscaled.todense(), decimal=2)
# Test that the l2 and D_A are the same for the 1 candidate case
scipy.random.seed(0) # make results deterministic
resultDA = evolution_soc(D*A*D, Dinv*B, epsilon=4.0,
k=2, proj_type="D_A",
symmetrize_measure=False)
scipy.random.seed(0) # make results deterministic
resultl2 = evolution_soc(D*A*D, Dinv*B, epsilon=4.0,
k=2, proj_type="l2",
symmetrize_measure=False)
assert_array_almost_equal(resultDA.todense(), resultl2.todense())
# Test Scale Invariance for multiple candidates
(A, B) = linear_elasticity((5, 5), format='bsr')
A = 1.0j*A
B = 1.0j*B
scipy.random.seed(0) # make results deterministic
result_unscaled = evolution_soc(A, B, epsilon=4.0, k=2,
proj_type="D_A",
symmetrize_measure=False)
# create scaled A
D = spdiags([arange(A.shape[0], 2*A.shape[0], dtype=float)],
[0], A.shape[0], A.shape[0], format='csr')
Dinv = spdiags([1.0/arange(A.shape[0], 2*A.shape[0], dtype=float)],
[0], A.shape[0], A.shape[0], format='csr')
scipy.random.seed(0) # make results deterministic
result_scaled = evolution_soc((D*A*D).tobsr(blocksize=(2, 2)), Dinv*B,
epsilon=4.0, k=2, proj_type="D_A",
symmetrize_measure=False)
assert_array_almost_equal(result_scaled.todense(),
result_unscaled.todense(), decimal=2)
# task1.3
from pyamg.gallery.diffusion import diffusion_stencil_2d
from pyamg.gallery import stencil_grid
sten = diffusion_stencil_2d(type='FD', \
epsilon=0.001, theta=3.1416/3.0)
A = stencil_grid(sten, (100,100), format='csr')
# task1.4
from pyamg import *
ml = smoothed_aggregation_solver(A)
# task1.5
from numpy import ones
b = ones((A.shape[0],1))
res = []
x = ml.solve(b, tol=1e-8, \
residuals=res)
from pylab import *
semilogy(res[1:])
xlabel('iteration')
ylabel('residual norm')
title('Residual History')
show()
if iter == maxiter:
return (postprocess(x), iter)
if __name__ == '__main__':
# from numpy import diag
# A = random((4,4))
# A = A*A.transpose() + diag([10,10,10,10])
# b = random((4,1))
# x0 = random((4,1))
from pyamg.gallery import stencil_grid
from numpy.random import random
import time
from pyamg.krylov._gmres import gmres
A = stencil_grid([[0, -1, 0], [-1, 4, -1], [0, -1, 0]], (100, 100),
dtype=float, format='csr')
b = random((A.shape[0],))
x0 = random((A.shape[0],))
print '\n\nTesting CR with %d x %d 2D Laplace Matrix' % \
(A.shape[0], A.shape[0])
t1 = time.time()
r = []
(x, flag) = cr(A, b, x0, tol=1e-8, maxiter=100, residuals=r)
t2 = time.time()
print '%s took %0.3f ms' % ('cr', (t2-t1)*1000.0)
print 'norm = %g' % (norm(b - A*x))
print 'info flag = %d' % (flag)
t1 = time.time()
r2 = []
import scipy
from pyamg.gallery import stencil_grid
from pyamg.gallery.diffusion import diffusion_stencil_2d
from pyamg.strength import classical_strength_of_connection
from pyamg.classical.classical import ruge_stuben_solver
from convergence_tools import print_cycle_history
if __name__ == '__main__':
n = 100
nx = n
ny = n
# Rotated Anisotropic Diffusion
stencil = diffusion_stencil_2d(type='FE',epsilon=0.001,theta=scipy.pi/3)
A = stencil_grid(stencil, (nx,ny), format='csr')
S = classical_strength_of_connection(A, 0.0)
numpy.random.seed(625)
x = scipy.rand(A.shape[0])
b = A*scipy.rand(A.shape[0])
ml = ruge_stuben_solver(A, max_coarse=10)
resvec = []
x = ml.solve(b, x0=x, maxiter=20, tol=1e-14, residuals=resvec)
print_cycle_history(resvec, ml, verbose=True, plotting=True)
def _build_derivative_matrix_part(npoints,
derivative,
order_accuracy,
h=1.0,
lbc='d',
rbc='d',
use_shifted_differences=False):
if order_accuracy % 2:
raise ValueError('Only even accuracy orders supported.')
centered_coeffs = centered_difference(derivative,
order_accuracy) / (h**derivative)
mtx = stencil_grid(centered_coeffs, (npoints, ), format='lil')
max_shift = order_accuracy // 2
if use_shifted_differences:
# Left side
odd_even_offset = 1 - derivative % 2
for i in range(0, max_shift):
coeffs = shifted_difference(derivative, order_accuracy,
-(max_shift + odd_even_offset) + i)
mtx[i, 0:len(coeffs)] = coeffs / (h**derivative)
# Right side
for i in range(-1, -max_shift - 1, -1):
coeffs = shifted_difference(derivative, order_accuracy,
max_shift + i + odd_even_offset)
mtx[i, slice(-1, -(len(coeffs) +
1), -1)] = coeffs[::-1] / (h**derivative)
if 'd' in lbc: #dirichlet
mtx[0, :] = 0
mtx[0, 0] = 1.0
elif 'n' in lbc: #neumann
mtx[0, :] = 0
coeffs = shifted_difference(1, order_accuracy, -max_shift) / h
coeffs /= (-1 * coeffs[0])
coeffs[0] = 0.0
mtx[0, 0:len(coeffs)] = coeffs
elif type(lbc) is tuple and 'g' in lbc[0]: #ghost
n_ghost_points = int(lbc[1])
mtx[0:n_ghost_points, :] = 0
for i in range(n_ghost_points):
mtx[i, i] = 1.0
if 'd' in rbc:
mtx[-1, :] = 0
mtx[-1, -1] = 1.0
elif 'n' in rbc:
mtx[-1, :] = 0
coeffs = shifted_difference(1, order_accuracy, max_shift) / h
coeffs /= (-1 * coeffs[-1])
coeffs[-1] = 0.0
mtx[-1, slice(-1, -(len(coeffs) + 1), -1)] = coeffs[::-1]
elif type(rbc) is tuple and 'g' in rbc[0]:
n_ghost_points = int(rbc[1])
mtx[slice(-1, -(n_ghost_points + 1), -1), :] = 0
for i in range(n_ghost_points):
mtx[-i - 1, -i - 1] = 1.0
return mtx.tocsr()
from pyamg.gallery.diffusion import diffusion_stencil_2d from pyamg.gallery import stencil_grid from numpy import set_printoptions set_printoptions(precision=2) sten = diffusion_stencil_2d(type="FD", epsilon=0.001, theta=3.1416 / 3.0) A = stencil_grid(sten, (100, 100), format="csr") # print the matrix stencil print(A[5050, :].data) print(sten)
def test_evolution_strength_of_connection(self):
cases = []
# Single near nullspace candidate
stencil = [[0.0, -1.0, 0.0], [-0.001, 2.002, -0.001], [0.0, -1.0, 0.0]]
A = 1.0j * stencil_grid(stencil, (4, 4), format='csr')
B = 1.0j * np.ones((A.shape[0], 1))
B[0] = 1.2 - 12.0j
B[11] = -14.2
cases.append({
'A': A.copy(),
'B': B.copy(),
'epsilon': 4.0,
'k': 2,
'proj': 'l2'
})
# Multiple near nullspace candidate
B = 1.0j * np.ones((A.shape[0], 2))
B[0:-1:2, 0] = 0.0
B[1:-1:2, 1] = 0.0
B[-1, 0] = 0.0
B[11, 1] = -14.2
B[0, 0] = 1.2 - 12.0j
cases.append({
'A': A.copy(),
'B': B.copy(),
'epsilon': 4.0,
'k': 2,
'proj': 'l2'
})
Absr = A.tobsr(blocksize=(2, 2))
cases.append({
'A': Absr.copy(),
'B': B.copy(),
'epsilon': 4.0,
'k': 2,
'proj': 'l2'
})
for ca in cases:
scipy.random.seed(0) # make results deterministic
result = evolution_soc(ca['A'],
ca['B'],
epsilon=ca['epsilon'],
k=ca['k'],
proj_type=ca['proj'],
symmetrize_measure=False)
scipy.random.seed(0) # make results deterministic
expected = reference_evolution_soc(ca['A'],
ca['B'],
epsilon=ca['epsilon'],
k=ca['k'],
proj_type=ca['proj'])
assert_array_almost_equal(result.todense(), expected.todense())
scipy.random.seed(0) # make results deterministic
result = evolution_soc(ca['A'],
ca['B'],
epsilon=ca['epsilon'],
k=ca['k'],
proj_type=ca['proj'],
symmetrize_measure=False,
weighting='local')
scipy.random.seed(0) # make results deterministic
expected = reference_evolution_soc(ca['A'],
ca['B'],
epsilon=ca['epsilon'],
k=ca['k'],
proj_type=ca['proj'],
weighting='local')
assert_array_almost_equal(result.todense(), expected.todense())
# Test Scale Invariance for a single candidate
A = 1.0j * poisson((5, 5), format='csr')
B = 1.0j * arange(1, A.shape[0] + 1, dtype=float).reshape(-1, 1)
scipy.random.seed(0) # make results deterministic
result_unscaled = evolution_soc(A,
B,
epsilon=4.0,
k=2,
proj_type="D_A",
symmetrize_measure=False)
# create scaled A
D = spdiags([arange(A.shape[0], 2 * A.shape[0], dtype=float)], [0],
A.shape[0],
A.shape[0],
format='csr')
Dinv = spdiags([1.0 / arange(A.shape[0], 2 * A.shape[0], dtype=float)],
[0],
A.shape[0],
A.shape[0],
format='csr')
scipy.random.seed(0) # make results deterministic
result_scaled = evolution_soc(D * A * D,
Dinv * B,
epsilon=4.0,
k=2,
proj_type="D_A",
symmetrize_measure=False)
assert_array_almost_equal(result_scaled.todense(),
result_unscaled.todense(),
decimal=2)
# Test that the l2 and D_A are the same for the 1 candidate case
scipy.random.seed(0) # make results deterministic
resultDA = evolution_soc(D * A * D,
Dinv * B,
epsilon=4.0,
k=2,
proj_type="D_A",
symmetrize_measure=False)
scipy.random.seed(0) # make results deterministic
resultl2 = evolution_soc(D * A * D,
Dinv * B,
epsilon=4.0,
k=2,
proj_type="l2",
symmetrize_measure=False)
assert_array_almost_equal(resultDA.todense(), resultl2.todense())
# Test Scale Invariance for multiple candidates
(A, B) = linear_elasticity((5, 5), format='bsr')
A = 1.0j * A
B = 1.0j * B
scipy.random.seed(0) # make results deterministic
result_unscaled = evolution_soc(A,
B,
epsilon=4.0,
k=2,
proj_type="D_A",
symmetrize_measure=False)
# create scaled A
D = spdiags([arange(A.shape[0], 2 * A.shape[0], dtype=float)], [0],
A.shape[0],
A.shape[0],
format='csr')
Dinv = spdiags([1.0 / arange(A.shape[0], 2 * A.shape[0], dtype=float)],
[0],
A.shape[0],
A.shape[0],
format='csr')
scipy.random.seed(0) # make results deterministic
result_scaled = evolution_soc((D * A * D).tobsr(blocksize=(2, 2)),
Dinv * B,
epsilon=4.0,
k=2,
proj_type="D_A",
symmetrize_measure=False)
assert_array_almost_equal(result_scaled.todense(),
result_unscaled.todense(),
decimal=2)
CSR matrix for a nonsymmetric recirculating flow problem
Many more solver parameters may be specified than outlined in the below
examples. Only the most basic are shown.
Run with
>>> python demo.py
and examine the on-screen output and file output.
'''
from scipy import pi
from pyamg import gallery
from pyamg.gallery import diffusion
stencil = diffusion.diffusion_stencil_2d(type='FE',
epsilon=0.001,
theta=2 * pi / 16.0)
A = gallery.stencil_grid(stencil, (50, 50), format='csr')
from pyamg import gallery
from solver_diagnostics import solver_diagnostics
from scipy import pi
from pyamg.gallery import diffusion
choice = input('\nThere are four different test problems. Enter \n' + \
'1: Isotropic diffusion example\n' + \
'2: Anisotropic diffusion example\n' + \
'3: Elasticity example\n' + \
'4: Nonsymmetric flow example\n\n ')
if choice == 1:
##
# Try a basic isotropic diffusion problem from finite differences
if iter == maxiter:
return (postprocess(x), iter)
if __name__ == '__main__':
# from numpy import diag
# A = random((4,4))
# A = A*A.transpose() + diag([10,10,10,10])
# b = random((4,1))
# x0 = random((4,1))
from pyamg.gallery import stencil_grid
from numpy.random import random
import time
from pyamg.krylov._gmres import gmres
A = stencil_grid([[0, -1, 0], [-1, 4, -1], [0, -1, 0]], (100, 100),
dtype=float, format='csr')
b = random((A.shape[0],))
x0 = random((A.shape[0],))
print '\n\nTesting CR with %d x %d 2D Laplace Matrix' % \
(A.shape[0], A.shape[0])
t1 = time.time()
r = []
(x, flag) = cr(A, b, x0, tol=1e-8, maxiter=100, residuals=r)
t2 = time.time()
print '%s took %0.3f ms' % ('cr', (t2-t1)*1000.0)
print 'norm = %g' % (norm(b - A*x))
print 'info flag = %d' % (flag)
t1 = time.time()
r2 = []
# Step 1: import scipy and pyamg packages #------------------------------------------------------------------ from numpy import meshgrid, linspace from scipy import rand, pi from scipy.linalg import norm from pyamg import * from pyamg.gallery import stencil_grid from pyamg.gallery.diffusion import diffusion_stencil_2d #------------------------------------------------------------------ # Step 2: setup up the system using pyamg.gallery #------------------------------------------------------------------ n=200 X,Y = meshgrid(linspace(0,1,n),linspace(0,1,n)) stencil = diffusion_stencil_2d(type='FE',epsilon=0.001,theta=pi/3) A = stencil_grid(stencil, (n,n), format='csr') b = rand(A.shape[0]) # pick a random right hand side #------------------------------------------------------------------ # Step 3: setup of the multigrid hierarchy #------------------------------------------------------------------ ml = smoothed_aggregation_solver(A) # construct the multigrid hierarchy #------------------------------------------------------------------ # Step 4: solve the system #------------------------------------------------------------------ res1 = [] x = ml.solve(b, tol=1e-12, residuals=res1)# solve Ax=b to a tolerance of 1e-12 #------------------------------------------------------------------ # Step 5: print details
# Step 1: import scipy and pyamg packages #------------------------------------------------------------------ from numpy import meshgrid, linspace from scipy import rand, pi from scipy.linalg import norm from pyamg import * from pyamg.gallery import stencil_grid from pyamg.gallery.diffusion import diffusion_stencil_2d #------------------------------------------------------------------ # Step 2: setup up the system using pyamg.gallery #------------------------------------------------------------------ n=200 X,Y = meshgrid(linspace(0,1,n),linspace(0,1,n)) stencil = diffusion_stencil_2d(type='FE',epsilon=0.001,theta=pi/3) A = stencil_grid(stencil, (n,n), format='csr') b = rand(A.shape[0]) # pick a random right hand side #------------------------------------------------------------------ # Step 3: setup of the multigrid hierarchy #------------------------------------------------------------------ ml = smoothed_aggregation_solver(A) # construct the multigrid hierarchy #------------------------------------------------------------------ # Step 4: solve the system #------------------------------------------------------------------ res1 = [] x = ml.solve(b, tol=1e-12, residuals=res1)# solve Ax=b to a tolerance of 1e-12 #------------------------------------------------------------------ # Step 5: print details
mcoarse = 10 # Max coarse grid size
prepost = ("gauss_seidel", {"sweep": "symmetric", "iterations": 1}) # pre/post smoother
smooth = ("energy", {"maxiter": 9, "degree": 3}) # Prolongation Smoother
classic_theta = 0.0 # Classic Strength Measure
# Drop Tolerance
evolution_theta = 4.0 # evolution Strength Measure
# Drop Tolerance
for n in nlist:
nx = n
ny = n
print "Running Grid = (%d x %d)" % (nx, ny)
# Rotated Anisotropic Diffusion Operator
stencil = diffusion_stencil_2d(type="FE", epsilon=epsilon, theta=theta)
A = stencil_grid(stencil, (nx, ny), format="csr")
# Random initial guess, zero RHS
x0 = scipy.rand(A.shape[0])
b = numpy.zeros((A.shape[0],))
# Classic SA strength measure
ml = smoothed_aggregation_solver(
A,
max_coarse=mcoarse,
coarse_solver="pinv2",
presmoother=prepost,
postsmoother=prepost,
smooth=smooth,
strength=("symmetric", {"theta": classic_theta}),
)
