def test_a(self):
'''
Poisson equation on a sine curve. In 3D?
It wasn't originally, but the restriciton algorithm was having
trouble with 1D. This might bear investigation.
'''
problemscale = 12
size = problemscale ** 3
gridLevels = 4
u_zeros = np.zeros((size,))
u_actual = np.sin(np.array(range(int(size))) * 3.0 / size).T
A = operators.poisson((size,))
b = tools.flexibleMmult(A, u_actual)
uSmoothed = smooth(A, b, u_zeros, iterations=1)
parameters = {'coarsestLevel': gridLevels - 1,
'problemShape': (problemscale, problemscale, problemscale),
'gridLevels': gridLevels,
'threshold': 8e-3,
}
u_mmg = mgSolve(A, b, parameters)
if self.verbose:
print 'norm is', (np.linalg.norm(tools.getresidual(b, A, uSmoothed.reshape((size, 1)), size)))
residual_norm = np.linalg.norm(tools.flexibleMmult(A, u_mmg) - b)
assert parameters['threshold'] > residual_norm
def mgCycle(A, b, level, R, parameters, initial=None):
"""
Internally used function that shows the actual multi-level solution method,
through a recursive call within the "level < coarsestLevel" block, below.
Parameters
----------
A : list of ndarrays
A list of square arrays; one for each level of resolution. As returned
by operators.coeffecientList().
b : ndarray
top-level RHS vector
level : int
The current multigrid level. This value is 0 at the entry point for standard,
recursive multigrid.
R : list of ndarrays
A list of (generally nonsquare) arrays; one for each transition between
levels of resolution. As returned by operators.restrictionList().
parameters : dict
A dictionary of parameters. See documentation for mgSolve() for details.
Optional Parameters
-------------------
initial=np.zeros((b.size, )) : ndarray
Initial iterate. Defaults to the zero vector, but the last supplied
solution should be used for chained v-cycles.
Returns
-------
uOut : ndarray
solution
infoDict
Dictionary of information about the solution process. Fields:
norm
The final norm of the residual
"""
verbose = parameters["verbose"]
if initial is None:
initial = np.zeros((b.size,))
N = b.size
# The general case is a recursive call. It comprises (1) pre-smoothing,
# (2) finding the coarse residual, (3) solving for the coarse correction
# to account for that residual via recursive call, (4) adding that
# correction to the pre-smoothed solution, (5) then possibly post-smoothing.
if level < parameters["coarsestLevel"]:
# (1) pre-smoothing
uApx = smooth(A[level], b, initial, parameters["preIterations"], verbose=verbose)
# (2) coarse residual
bCoarse = tools.flexibleMmult(R[level], b.reshape((N, 1)))
NH = len(bCoarse)
if verbose:
print level * " " + "calling mgCycle at level %i" % level
residual = tools.getresidual(b, A[level], uApx, N)
coarseResidual = tools.flexibleMmult(R[level], residual.reshape((N, 1))).reshape((NH,))
# (3) correction for residual via recursive call
coarseCorrection = mgCycle(A, coarseResidual, level + 1, R, parameters)[0]
correction = (tools.flexibleMmult(R[level].transpose(), coarseCorrection.reshape((NH, 1)))).reshape((N,))
if parameters["postIterations"] > 0:
# (5) post-smoothing
uOut = smooth(A[level], b, uApx + correction, parameters["postIterations"], verbose=verbose) # (4)
else:
uOut = uApx + correction # (4)
# Save norm, to monitor for convergence at topmost level.
norm = scipy.sparse.base.np.linalg.norm(tools.getresidual(b, A[level], uOut, N))
# The recursive calls only end when we're at the coarsest level, where
# we simply apply the chosen coarse solver.
else:
norm = 0
if verbose:
print level * " " + "direct solving at level %i" % level
uOut = coarseSolve(A[level], b.reshape((N, 1)))
return uOut, {"norm": norm}
def multifreq_1d(self, N, finaliterations, spectralxscale='linear',
solutionname='whitenoise', save=False):
'''Generates 1-dimensional uniform noise with N unknowns, then applies
Gauss-Siedel smoothing iterations to an initial zero-vector guess until
finaliterations is reached, graphing the Fourier transform of the error
each time. Does so pretty inefficiently, by doing 1 iteration, 2 iterations,
3, iterations, etc. until finaliterations, instead of just doing
finaliterations iterations, and graphing after each sweep.
Uses ffmpeg and mplayer.'''
if solutionname == 'summedsine': # Summed sine waves of several frequencies.
data = np.zeros((N,)) # initialize
domainwidthsexponents = range(10)
domainwidths = range(10) # initialize
for i in range(10):
domainwidths[i] = float(N) / (2.0 ** domainwidthsexponents[i])
for domainwidth in domainwidths:
sininput = np.arange(0.0, domainwidth, domainwidth / N)
xdata = (np.sin(sininput) / domainwidth * 10.0)[0:N:]
data = data + xdata
else:
solutionname = 'whitenoise'
data = np.random.random((N,)) # 1D white noise
gif_output_name = 'spectral_solution_rate-%i_unknowns-%s_xscale-%s_solution.gif' % (N, spectralxscale, solutionname)
A = operators.poisson((N,))
b = tools.flexibleMmult(A, data)
if self.saveFig:
from scipy import fftpack
import matplotlib.pyplot as plt
# Prepare for Graphing
fig = plt.figure(figsize=(7.5, 7))
fig.suptitle(gif_output_name)
fig.subplots_adjust(wspace=.2)
fig.subplots_adjust(hspace=.2)
solnax = fig.add_subplot(211) # subplot to hold the solution
solnaxylim = [0, 1]
specax = fig.add_subplot(212) # subplot to hold the spectrum
specaxylim = [10.0 ** (-18), 10.0 ** 3]
solutions = []
spectra = []
filenames = []
loop = 1 # for monitoring progress
iterationslist = range(finaliterations)
# from random import shuffle
# shuffle(iterationslist)
if save:
iterations_to_save = range(1, finaliterations + 1)
str_tosave = []
for x in iterations_to_save:
str_tosave.append(str(x))
from os import system; system("mkdir -p output")
csvfilelabel = 'iterative_spectra-%i_N' % N
csvfile = open('output/%s.csv' % csvfilelabel, 'a')
csvfile.write('frequency,' + ','.join(str_tosave) + '\n')
csvfile.flush()
for iterations in iterationslist:
solutions.append(smooth(A,
b,
np.zeros((N,)),
iterations)
)
if self.saveFig:
spectra.append(fftpack.fft(b - tools.flexibleMmult(A,
solutions[iterations]
)
)
)
if self.saveFig:
filebasename = '%i_unknowns-%i_iterations' % (N, iterations)
filename = filebasename + '-solution.png'
solnax.set_ylim(solnaxylim)
solnax.set_autoscaley_on(False)
solnax.set_title(filebasename + ': Solution')
solnax.plot(solutions[iterations])
specax.set_yscale('log')
specax.set_xscale(spectralxscale)
specax.set_ylim(specaxylim)
specax.set_autoscaley_on(False)
specax.set_title(filebasename + ': Solution Error (frequency domain)')
specax.plot(spectra[iterations])
filename = 'temp/' + filebasename + '-combined.png'
filenames.append(filename)
if self.saveFig:
print "%i of %i: " % (loop, finaliterations) + "saving " + filename
fig.savefig(filename, dpi=80)
solnax.cla()
specax.cla()
loop += 1
frequencies = range(len(solutions[0]))
for i in frequencies:
towrite = '%i,' % frequencies[i]
for spectrum in spectra:
towrite += '%.8f,' % abs(spectrum[i])
if save:
csvfile.write(towrite + '\n')
csvfile.flush()
if save:
csvfile.close()
# make .gif and .mp4 files:
if self.saveFig:
# TODO (Tom) Might use subprocess to suppress super verbose ffmpeg
# import subprocess
# s = subprocess.Popen(['cowsay', 'hello']
# stderr=subprocess.STDOUT
# stdout=subprocess.PIPE).communicate()[0]
# print s
torun = []
torun.extend([
'convert ' + ' '.join(filenames) + ' output/%s' % gif_output_name,
'rm ' + ' '.join(filenames),
'cp %s simple.gif' % ('output/' + gif_output_name),
'mplayer -vo jpeg simple.gif > /dev/null 2> /dev/null',
'ffmpeg -r 12 -i %%08d.jpg -y -an output/%s.avi > /dev/null' % gif_output_name,
'rm simple.gif 0*.jpg',
'ffmpeg -y -i output/%s.avi output/%s.mp4 > /dev/null && rm output/%s.avi' % (gif_output_name, gif_output_name, gif_output_name),
])
for command in torun:
system(command)
def multifreq_2d(self, problemscale, finaliterations):
'''Attempts to generate a pretty 2D solution, by summing several 2D
sine waves of different frequencies. Then, proceeds to apply a GS smoother
to the generated problem, saving pretty pictures after each iteration.
This might work OK if problemscale is big enough.
'''
if self.saveFig:
from scipy import fftpack
import matplotlib.pyplot as plt
from matplotlib import cm
from mpl_toolkits.mplot3d import Axes3D
# Generate test problem. Summed sine waves of several resolutions:
NX = problemscale
floatNX = float(NX)
data = np.zeros((NX, NX))
domainwidths = [floatNX, floatNX / 2.0, 10., 3.1415]
for domainwidth in domainwidths:
sininput = np.arange(0.0, domainwidth, domainwidth / NX)
xdata = (np.sin(sininput) / domainwidth * 10.0)[0:NX:]
data = data + xdata[:, np.newaxis] + xdata[:, np.newaxis].T
X = np.tile(np.array(range(NX)), (NX, 1))
Y = X.T
if self.saveFig:
# Prepare for Graphing:
fig = plt.figure()
ax = Axes3D(fig)
ax.set_zlim3d([-1, 6])
specfig = plt.figure()
specax = specfig.add_subplot(111)
specax.set_ylim([.000001, 10000])
specax.set_autoscaley_on(False)
specax.set_xscale('log')
# Problem Setup:
A = operators.poisson((NX, NX))
solnforb = data.ravel().reshape((NX ** 2, 1))
b = tools.flexibleMmult(A, solnforb)
# Initial "Solution" placeholder:
solutions = range(finaliterations)
spectra = range(finaliterations)
iterations = 0
solutions[iterations] = np.zeros((NX, NX)).ravel()
spectra[iterations] = b - tools.flexibleMmult(A, solutions[iterations])
if self.saveFig:
# Initial graphs:
filebasename = 'output/N_%i-iterations_%i' % (NX ** 2, iterations)
filename = filebasename + '.png'
ax.set_title(filebasename)
# surf = ax.plot_wireframe(X, Y, solutions[iterations].reshape(NX,NX),rstride=stridelength,cstride=stridelength)
surf = ax.plot_surface(X, Y, data, cmap=cm.jet)
# print "saving", filename
if self.saveFig: fig.savefig(filename)
surf.remove()
ax.cla()
# Error Spectrum
specax.plot(spectra[iterations])
specax.set_title(filebasename + ': Error Spectrum')
filename = filebasename + '-error.png'
if self.saveFig: specfig.savefig(filename)
del specax.lines[0]
specax.cla()
verbose = False
for iterations in range(1, finaliterations):
solutions[iterations] = smooth(A,
b,
np.zeros(NX ** 2),
iterations,
verbose=verbose
).reshape(NX, NX)
if self.saveFig:
spectra[iterations] = fftpack.fft2(solutions[iterations])
if self.saveFig:
filebasename = 'output/N_%i-iterations_%i' % (NX ** 2, iterations)
filename = filebasename + '-solution.png'
ax.set_zlim3d([-1, 6])
surf = ax.plot_surface(X, Y, solutions[iterations], cmap=cm.jet)
ax.set_title(filebasename)
surf = ax.plot_wireframe(X, Y, solutions[iterations])
# print "saving", filename
if self.saveFig: fig.savefig(filename)
surf.remove()
ax.cla()
specax.set_yscale('log')
specax.set_ylim([.000001, 10000])
specax.set_autoscaley_on(False)
specax.set_xscale('log')
specax.plot(spectra[iterations])
specax.set_title(filebasename + ': Error Spectrum')
filename = filebasename + '-error.png'
# print "saving", filename
if self.saveFig: specfig.savefig(filename)
del specax.lines[0]
specax.cla()
def mgCycle(A, b, level, R, parameters, initial=None):
"""
Internally used function that shows the actual multi-level solution method,
through a recursive call within the "level < coarsestLevel" block, below.
Parameters
----------
A : list of ndarrays
A list of square arrays; one for each level of resolution. As returned
by operators.coeffecientList().
b : ndarray
top-level RHS vector
level : int
The current multigrid level. This value is 0 at the entry point for standard,
recursive multigrid.
R : list of ndarrays
A list of (generally nonsquare) arrays; one for each transition between
levels of resolution. As returned by operators.restrictionList().
parameters : dict
A dictionary of parameters. See documentation for mgSolve() for details.
Optional Parameters
-------------------
initial=np.zeros((b.size, )) : ndarray
Initial iterate. Defaults to the zero vector, but the last supplied
solution should be used for chained v-cycles.
Returns
-------
uOut : ndarray
solution
infoDict
Dictionary of information about the solution process. Fields:
norm
The final norm of the residual
"""
verbose = parameters['verbose']
if initial is None:
initial = np.zeros((b.size, ))
N = b.size
# The general case is a recursive call. It comprises (1) pre-smoothing,
# (2) finding the coarse residual, (3) solving for the coarse correction
# to account for that residual via recursive call, (4) adding that
# correction to the pre-smoothed solution, (5) then possibly post-smoothing.
if level < parameters['coarsestLevel']:
# (1) pre-smoothing
uApx = smooth(A[level], b, initial,
parameters['preIterations'], verbose=verbose)
# (2) coarse residual
bCoarse = tools.flexibleMmult(R[level], b.reshape((N, 1)))
NH = len(bCoarse)
if verbose: print level * " " + "calling mgCycle at level %i" % level
residual = tools.getresidual(b, A[level], uApx, N)
coarseResidual = tools.flexibleMmult(R[level], residual.reshape((N, 1))).reshape((NH,))
# (3) correction for residual via recursive call
coarseCorrection = mgCycle(A, coarseResidual, level+1, R, parameters)[0]
correction = (tools.flexibleMmult(R[level].transpose(), coarseCorrection.reshape((NH, 1)))).reshape((N, ))
if parameters['postIterations'] > 0:
# (5) post-smoothing
uOut = smooth(A[level],
b,
uApx + correction, # (4)
parameters['postIterations'],
verbose=verbose)
else:
uOut = uApx + correction # (4)
# Save norm, to monitor for convergence at topmost level.
norm = scipy.sparse.base.np.linalg.norm(tools.getresidual(b, A[level], uOut, N))
# The recursive calls only end when we're at the coarsest level, where
# we simply apply the chosen coarse solver.
else:
norm = 0
if verbose: print level * " " + "direct solving at level %i" % level
uOut = coarseSolve(A[level], b.reshape((N, 1)))
return uOut, {'norm': norm}
