Skip to content

Instantly share code, notes, and snippets.

@jakevdp

jakevdp/banded_tools.py

Created December 23, 2011 02:26
Show Gist options
  • Save jakevdp/1512840 to your computer and use it in GitHub Desktop.
Save jakevdp/1512840 to your computer and use it in GitHub Desktop.
Download ZIP
Benchmarks for eigenvalue decomposition
Raw
banded_tools.py
from time import time
import numpy as np
from scipy.sparse import spdiags, issparse, dia_matrix
from scipy.sparse.linalg import factorized
from scipy import linalg as splinalg
class BandedMatrix(object):
def __init__(self, data, lu=None):
if issparse(data):
if lu:
raise ValueError("Data type not understood")
M = data.todia()
Nd_u = max(M.offsets)
Nd_l = -min(M.offsets)
self.data_ = np.zeros((1 + Nd_u + Nd_l, min(M.shape)))
self.data_[Nd_u - M.offsets, :M.data.shape[1]] = M.data
self.l_ = Nd_l
self.u_ = Nd_u
else:
self.data_ = data
self.l_ = lu[0]
self.u_ = lu[1]
assert self.data_.shape[0] == 1 + self.l_ + self.u_
assert self.u_ < self.data_.shape[1]
assert self.l_ < self.data_.shape[1]
shape = property(lambda self: (self.data_.shape[1], self.data_.shape[1]))
lu = property(lambda self: (self.l_, self.u_))
data = property(lambda self: self.data_)
def tocsr(self):
return spdiags(self.data,
range(-self.l_, self.u_ + 1)[::-1],
*self.shape, format='csr')
def tocsc(self):
return spdiags(self.data,
range(-self.l_, self.u_ + 1)[::-1],
*self.shape, format='csc')
def todia(self):
return spdiags(self.data,
range(-self.l_, self.u_ + 1)[::-1],
*self.shape, format='dia')
def toarray(self):
return self.todia().toarray()
def todense(self):
return self.todia().todense()
def to_symmetric(self, lower=False):
if lower:
return SymmetricBandedMatrix(self.data_[self.u_:], lower=True)
else:
return SymmetricBandedMatrix(self.data_[:self.u_ + 1], lower=False)
def diag(self, d=0):
"""
return a view of the d-th diagonal
"""
assert d == 0
return self.data_[self.u_]
class SymmetricBandedMatrix(object):
def __init__(self, data, lower=False):
if issparse(data):
M = data.todia()
if lower:
i_lower = np.where(M.offsets <= 0)
Nd = -min(M.offsets)
self.data_ = np.zeros((Nd + 1, M.shape[1]))
self.data_[-M.offsets[i_lower],
:M.data.shape[1]] = M.data[i_lower]
else:
i_upper = np.where(M.offsets >= 0)
Nd = max(M.offsets)
self.data_ = np.zeros((Nd + 1, M.shape[1]))
self.data_[Nd - M.offsets[i_upper],
:M.data.shape[1]] = M.data[i_upper]
else:
self.data_ = data
self.lower_ = lower
data = property(lambda self: self.data_)
lower = property(lambda self: self.lower_)
shape = property(lambda self: (self.data_.shape[1], self.data_.shape[1]))
def _nonsym_data(self):
Nd = self.data_.shape[0] - 1
newdata = np.zeros((1 + 2 * Nd, self.data_.shape[1]))
if self.lower_:
newdata[-Nd - 1:] = self.data_
for i in range(Nd):
newdata[Nd - i - 1, i + 1:] = newdata[Nd + 1 + i, : -1 - i]
else:
newdata[:Nd + 1] = self.data_
for i in range(Nd):
newdata[Nd + 1 + i, :-1 - i] = newdata[Nd - i - 1, i + 1:]
return newdata
def tocsr(self):
Nd = self.data_.shape[0] - 1
return spdiags(self._nonsym_data(),
np.arange(-Nd, Nd + 1)[::-1],
*self.shape, format='csr')
def tocsc(self):
Nd = self.data_.shape[0] - 1
return spdiags(self._nonsym_data(),
np.arange(-Nd, Nd + 1)[::-1],
*self.shape, format='csc')
def todia(self):
Nd = self.data_.shape[0] - 1
return spdiags(self._nonsym_data(),
np.arange(-Nd, Nd + 1)[::-1],
*self.shape, format='dia')
def toarray(self):
return self.todia().toarray()
def todense(self):
return self.todia().todense()
def togeneral(self):
Nd = self.data_.shape[0] - 1
return BandedMatrix(self._nonsym_data(), (Nd, Nd))
def diag(self, d=0):
"""
return a view of the d-th diagonal
"""
assert d == 0
if self.lower_:
return self.data_[0]
else:
return self.data_[-1]
def solve_banded(M, b, overwrite_M=False, overwrite_b=False):
assert M.__class__ == BandedMatrix
return splinalg.solve_banded(M.lu, M.data, b,
overwrite_ab=overwrite_M,
overwrite_b=overwrite_b)
def solveh_banded(M, b, overwrite_M=False, overwrite_b=False):
assert M.__class__ == SymmetricBandedMatrix
return splinalg.solveh_banded(M.data, b,
overwrite_ab=overwrite_M,
overwrite_b=overwrite_b)
def test_tools():
N = 6
nnz = 6
X = np.zeros((N, N))
ind = np.arange(nnz, dtype=int)
X[ind, ind] = 1
X[ind[1:], ind[:-1]] = 2 + 0.1 * np.arange(nnz - 1)
X[ind[:-1], ind[1:]] = 3 + 0.1 * np.arange(nnz - 1)
X[ind[:-2], ind[2:]] = 4 + 0.1 * np.arange(nnz - 2)
from scipy.sparse import dia_matrix
print X
X_dia = dia_matrix(X)
X_banded = BandedMatrix(X_dia)
X_up = SymmetricBandedMatrix(X_dia)
X_lo = SymmetricBandedMatrix(X_dia, lower=True)
print np.all(X_up.toarray() == X_up.togeneral().toarray())
print np.all(X_lo.toarray() == X_lo.togeneral().toarray())
print np.all(X_banded.toarray() == X_dia.toarray())
print
print X_banded.to_symmetric(lower=True).toarray()
print
print X_banded.to_symmetric(lower=False).toarray()
def test_solvers(dim=128 ** 2):
Nd = 2
N = 1 + 2 * Nd
data = np.ones((N, dim))
for i in range(1, Nd + 1):
data[i: N - i] *= 1.633
X_dia = spdiags(data, 2 * np.arange(-Nd, Nd + 1), dim, dim)
b = np.random.random(dim)
# subtract 1 from the diagonal
print "super-lu"
t_1 = time()
X_csc = X_dia.tocsc()
t0 = time()
fac = factorized(X_csc)
t1 = time()
x0 = fac(b)
t2 = time()
print ' - convert to csc: %.2g sec' % (t0 - t_1)
print ' - factorize: %.2g sec' % (t1 - t0)
print ' - solve: %.2g sec' % (t2 - t1)
print "banded general:"
t_1 = time()
X_gbn = BandedMatrix(X_dia)
t0 = time()
x1 = solve_banded(X_gbn, b)
t1 = time()
print ' - convert to banded: %.2g' % (t0 - t_1)
print ' - solve: %.2g sec' % (t1 - t0)
print "banded symmetric:"
t_1 = time()
X_sbn = SymmetricBandedMatrix(X_dia)
t0 = time()
x2 = solveh_banded(X_sbn, b)
t1 = time()
print ' - convert to banded: %.2g' % (t0 - t_1)
print ' - solve: %.2g sec' % (t1 - t0)
print
print "Results match:",
print np.allclose(x0, x1),
print np.allclose(x1, x2)
if __name__ == '__main__':
test_solvers()
Raw
bench_laplacian_solve.py
"""
benchmarks for eigensolution of the Laplacian matrix.
Note:
The banded version is not as efficient as it could be: it
should be upgraded to scipy.linalg.solveh_banded.
Better, the LAPACK routine should be used to first preprocess the banded
matrix by converting it into a tridiagonal matrix, and performing the
solution using this. These routines are not currently wrapped in scipy,
so we'd have to interface with LAPACK directly.
"""
from time import time
import numpy as np
import scipy as sp
from scipy.linalg import solve_banded, solve, eig_banded
from scipy.sparse import spdiags
from scipy.sparse.linalg import LinearOperator, factorized
from sklearn.utils.arpack import eigsh
from sklearn.feature_extraction import image
from sklearn.utils.graph import graph_laplacian
import banded_tools
def get_Laplacian_matrix(num_downsamples = 2):
"""
Construct the laplacian matrix used in the plot_lena_segmentation example.
Some of the code in this function is from
"""
lena = sp.misc.lena()
# downsample pixels by factors of two
for i in range(num_downsamples):
lena = (lena[::2, ::2] + lena[1::2, ::2]
+ lena[::2, 1::2] + lena[1::2, 1::2])
# Convert the image into a graph with the value of the gradient on the
# edges.
adjacency = image.img_to_graph(lena)
# Take a decreasing function of the gradient: an exponential
# The smaller beta is, the more independant the segmentation is of the
# actual image. For beta=1, the segmentation is close to a voronoi
beta = 5
eps = 1e-6
adjacency.data = np.exp(-beta*adjacency.data/lena.std()) + eps
# Compute the negative laplacian of the adjacency (this is from
# sklearn.cluster.spectral_embedding)
laplacian, dd = graph_laplacian(adjacency,
normed=True, return_diag=True)
# Set diagonal of Laplacian to zero
laplacian = laplacian.tocoo()
diag_idx = (laplacian.row == laplacian.col)
laplacian.data[diag_idx] = 0
# Convert to diagonal matrix: this gives the most efficient
# matrix/vector products
return laplacian.todia()
def construct_OPinv_banded(M_banded):
"""
construct LinearOperator OPinv such that if
(M - I) * x = b
then
x = OPinv * b
for vectors x and b, where I is the identity matrix
"""
Nd = (M_banded.shape[0] - 1) / 2
N = M_banded.shape[1]
OP_banded = M_banded.copy()
OP_banded[Nd] -= 1
matvec = lambda b: solve_banded((Nd, Nd), OP_banded, b)
return LinearOperator((N, N), matvec = matvec)
def bench_sparse_solvers(num_downsamples=2, rseed=None,
super_lu=True,
banded_gen=True,
banded_sym=True):
print "creating Laplacian matrix:"
M_dia = -get_Laplacian_matrix(num_downsamples)
print " - shape of matrix:", M_dia.shape
# subtract 1 from the diagonal
i = np.where(M_dia.offsets == 0)
M_dia.data[i] -= 1
if rseed is not None:
np.random.seed(rseed)
b = np.zeros(M_dia.shape[1])
res = []
if super_lu:
print "super-lu"
t_1 = time()
M_csc = M_dia.tocsc()
t0 = time()
fac = factorized(M_csc)
t1 = time()
x = fac(b)
t2 = time()
print ' - convert to csc: %.2g sec' % (t0 - t_1)
print ' - factorize: %.2g sec' % (t1 - t0)
print ' - solve: %.2g sec' % (t2 - t1)
res.append(x)
if banded_gen:
print "banded general:"
t_1 = time()
M_gbn = banded_tools.BandedMatrix(M_dia)
t0 = time()
x = banded_tools.solve_banded(M_gbn, b)
t1 = time()
print ' - convert to banded: %.2g' % (t0 - t_1)
print ' - solve: %.2g sec' % (t1 - t0)
res.append(x)
if banded_sym:
print "banded symmetric:"
t_1 = time()
M_sbn = banded_tools.SymmetricBandedMatrix(-M_dia)
d = M_sbn.diag()
d += 1E-8
t0 = time()
x = banded_tools.solveh_banded(M_sbn, b)
t1 = time()
print ' - convert to banded: %.2g' % (t0 - t_1)
print ' - solve: %.2g sec' % (t1 - t0)
res.append(x)
for r in res:
print r[:6]
print
print "Results match:",
for i in range(len(res)):
print np.allclose(res[i - 1], res[i])
print
def bench_eigendecompositions(num_downsamples=2,
nev=11, tol=1E-6,
eigsh_direct=True,
eigsh_invert=True,
eigsh_invert_banded=True,
eigsh_invert_buckling=True,
eigsh_invert_cayley=True,
check_equivalence=True):
"""
Run benchmark tests of different eigenvalue decomposition methods
"""
print "creating Laplacian matrix:"
M_dia = -get_Laplacian_matrix(num_downsamples)
print " - shape of matrix:", M_dia.shape
print
print "constructing banded representation:"
M_banded = banded_tools.BandedMatrix(M_dia)
print " - banded storage shape:", M_banded.data.shape
if check_equivalence:
print
print "checking that conversion was successful:"
M_csr = M_banded.tocsr()
v = np.random.random(M_banded.shape[1])
t0 = time()
x1 = M_dia * v
t1 = time()
x2 = M_csr * v
t2 = time()
print " - matrix products match:", np.allclose(x1, x2)
print " [dia_matrix mat-vec product: %.2g sec]" % (t1 - t0)
print " [csr_matrix mat-vec product: %.2g sec]" % (t2 - t1)
print 75 * '_'
print "Computing %i eigenvalues with tol=%.1g" % (nev, tol)
if eigsh_direct:
print
print "eigsh: direct mode"
t0 = time()
evals, evecs = eigsh(M_dia, nev,
which='LA', tol=tol)
t1 = time()
print " - finished in %.2g sec" % (t1 - t0)
print "eigenvalues:"
print evals
if eigsh_invert:
print
print "eigsh: shift-invert mode"
t0 = time()
evals, evecs = eigsh(M_dia, nev, sigma=1,
which='LM', tol=tol)
t1 = time()
print " - finished in %.2g sec" % (t1 - t0)
print "eigenvalues:"
print evals
if eigsh_invert_banded:
print
print "eigsh: shift-invert mode (banded representation)"
t0 = time()
evals, evecs = eigsh(M_dia, nev, sigma=1,
OPinv=construct_OPinv_banded(M_banded.data),
which='LM', tol=tol)
t1 = time()
print " - finished in %.2g sec" % (t1 - t0)
print "eigenvalues:"
print evals
if eigsh_invert_buckling:
print
print "eigsh: shift-invert/buckling mode"
t0 = time()
evals, evecs = eigsh(M_dia, nev, sigma=1,
which='LM', tol=tol,
mode='buckling')
t1 = time()
print " - finished in %.2g sec" % (t1 - t0)
print "eigenvalues:"
print evals
if eigsh_invert_cayley:
print
print "eigsh: shift-invert/cayley mode"
t0 = time()
evals, evecs = eigsh(M_dia, nev, sigma=1,
which='LM', tol=tol,
mode='cayley')
t1 = time()
print " - finished in %.2g sec" % (t1 - t0)
print "eigenvalues:"
print evals
if __name__ == '__main__':
#bench_sparse_solvers()
bench_eigendecompositions()
Raw
gistfile1.txt
Related to https://github.com/scikit-learn/scikit-learn/pull/50
Summary:
- shift-invert mode is fast, especially with newer versions of scipy (probably due to improvements in SuperLU?)
- using the LAPACK banded functionality doesn't help much. Wrapping LAPACK's DSBTRD and DPTSV may improve this a bit, but I'm not sure it's worth it given the performance of later scipy versions
- on the 16384 x 16384 array used in the plot_lena_segmentation.py example, newer versions of scipy with a simple shift-invert converge to the solution in < 1 sec with tol=1E-12. This should work fine in practice.
Raw
results_1e-12.txt
Scipy 0.10
==========
In [3]: scipy.__version__
Out[3]: '0.10.0.dev'
In [4]: bench_eigendecompositions(tol=1E-12)
creating Laplacian matrix:
- shape of matrix: (16384, 16384)
constructing banded representation:
- banded storage shape: (257, 16384)
checking that conversion was successful:
- matrix products match: True
[dia_matrix mat-vec product: 0.00063 sec]
[csr_matrix mat-vec product: 0.0005 sec]
___________________________________________________________________________
Computing 11 eigenvalues with tol=1e-12
eigsh: direct mode
- finished in 55 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode
- finished in 1 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode (banded representation)
- finished in 41 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/buckling mode
- finished in 1 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/cayley mode
- finished in 1 sec
eigenvalues:
[ 0.99948885 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
Scipy 0.9
=========
In [3]: scipy.__version__
Out[3]: '0.9.1.dev'
In [4]: bench_eigendecompositions(tol=1E-12)
creating Laplacian matrix:
- shape of matrix: (16384, 16384)
constructing banded representation:
- banded storage shape: (257, 16384)
checking that conversion was successful:
- matrix products match: True
[dia_matrix mat-vec product: 0.00062 sec]
[csr_matrix mat-vec product: 0.00051 sec]
___________________________________________________________________________
Computing 11 eigenvalues with tol=1e-12
eigsh: direct mode
- finished in 56 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode
- finished in 1.1 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode (banded representation)
- finished in 42 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/buckling mode
- finished in 1 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/cayley mode
- finished in 1 sec
eigenvalues:
[ 0.99948885 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
Scipy 0.8
=========
In [5]: scipy.__version__
Out[5]: '0.8.0.dev'
In [6]: bench_eigendecompositions(tol=1E-12)
creating Laplacian matrix:
- shape of matrix: (16384, 16384)
constructing banded representation:
- banded storage shape: (257, 16384)
checking that conversion was successful:
- matrix products match: True
[dia_matrix mat-vec product: 0.00066 sec]
[csr_matrix mat-vec product: 0.0005 sec]
___________________________________________________________________________
Computing 11 eigenvalues with tol=1e-12
eigsh: direct mode
- finished in 55 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode
- finished in 1 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode (banded representation)
- finished in 50 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/buckling mode
- finished in 1.5 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/cayley mode
- finished in 1.6 sec
eigenvalues:
[ 0.99948885 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
Raw
results_1e-6.txt
Scipy 0.9
=========
In [11]: scipy.__version__
Out[11]: '0.9.1.dev'
In [12]: bench_eigendecompositions()
creating Laplacian matrix:
- shape of matrix: (16384, 16384)
constructing banded representation:
- banded storage shape: (257, 16384)
checking that conversion was successful:
- matrix products match: True
[dia_matrix mat-vec product: 0.00065 sec]
[csr_matrix mat-vec product: 0.00051 sec]
___________________________________________________________________________
Computing 11 eigenvalues with tol=1e-06
eigsh: direct mode
- finished in 37 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode
- finished in 0.8 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode (banded representation)
- finished in 39 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/buckling mode
- finished in 1.2 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/cayley mode
- finished in 1.3 sec
eigenvalues:
[ 0.99948885 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
Scipy 0.8
=========
In [3]: scipy.__version__
Out[3]: '0.8.0.dev'
In [4]: bench_eigendecompositions()
creating Laplacian matrix:
- shape of matrix: (16384, 16384)
constructing banded representation:
- banded storage shape: (257, 16384)
checking that conversion was successful:
- matrix products match: True
[dia_matrix mat-vec product: 0.00064 sec]
[csr_matrix mat-vec product: 0.00051 sec]
___________________________________________________________________________
Computing 11 eigenvalues with tol=1e-06
eigsh: direct mode
- finished in 37 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode
- finished in 0.81 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode (banded representation)
- finished in 30 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/buckling mode
- finished in 0.88 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/cayley mode
- finished in 0.96 sec
eigenvalues:
[ 0.99948885 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
Scipy 0.7
=========
In [4]: scipy.__version__
Out[4]: '0.7.0'
In [5]: bench_eigendecompositions()
creating Laplacian matrix:
- shape of matrix: (16384, 16384)
constructing banded representation:
- banded storage shape: (257, 16384)
checking that conversion was successful:
- matrix products match: True
[dia_matrix mat-vec product: 0.00068 sec]
[csr_matrix mat-vec product: 0.00059 sec]
___________________________________________________________________________
Computing 11 eigenvalues with tol=1e-06
eigsh: direct mode
- finished in 74 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode
- finished in 19 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert mode (banded representation)
- finished in 31 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/buckling mode
- finished in 20 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
eigsh: shift-invert/cayley mode
- finished in 20 sec
eigenvalues:
[ 0.99948884 0.99958284 0.99966101 0.99970649 0.99973406 0.99977508
0.99983923 0.99989525 0.99993584 0.9999775 1. ]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment