我正在Linux Anaconda 2.5.0环境中的Jupyter笔记本中运行scipy.sparse.linalg.eigs函数,但每次运行它时,特征值似乎都会改变。当我在Mac OSX环境中运行完全相同的笔记本时,无论我多少次运行笔记本,它只给我一个特征值的解决方案,应该是这种情况。这可能是Linux系统上的scipy包的错误吗?我正在运行的笔记本如下:
import numpy as np
from scipy import fftpack as fft
from scipy.sparse import lil_matrix, csc_matrix
from scipy.sparse.linalg import eigs, inv
nz = 100
zin = np.arange(nz+1, dtype=np.float64)/nz
N2 = np.full(nz, 1.)
f0 = 1.
beta = 0.
Nx = int(1e2)
Ny = int(1e2)
dx = 1e-1
dy = 1e-1
vbar = np.zeros(nz+1)
ubar = zin
etax = np.zeros(2)
etay = np.zeros(2)
k = fft.fftfreq(Nx, dx)[:Nx/2]
l = fft.fftfreq(Ny, dy)[:Ny/2]
zc = np.hstack(0.5*(zin[1:] + zin[:-1]))
dzc = np.hstack(np.diff(zc))
depth = 1.
dztop = zc[0]
dzbot = depth - zc[-1]
zf = zin
dzf = np.diff(zf)
j = 0
for i in range(len(k)):
L = lil_matrix((nz+1, nz+1), dtype=np.float64)
G = L.copy()
################
# n = 0 (surface)
################
R = k[i] * .5*(ubar[0]+ubar[1]) + l[j] * .5*(vbar[0]+vbar[1])
D = dzf[0]**-1
S = .5 * ( k[i] * ( (ubar[0]-ubar[1])*D - N2[0]/f0 * etay[0] )
+ l[j] * ( (vbar[0]-vbar[1])*D + N2[0]/f0 * etax[0] ) )
L[0, 0] = R * D - S
L[0, 1] = R * (-D) - S
G[0, 0] = D
G[0, 1] = - D
################
# n = nz (bottom)
################
R = k[i] * .5*(ubar[nz-1]+ubar[nz]) + l[j] * .5*(vbar[nz-1]+vbar[nz])
D = dzf[nz-1]**-1
S = .5 * ( k[i] * ( (ubar[nz-1]-ubar[nz])*D - N2[nz-1]/f0 * etay[1] )
+ l[j] * ( (vbar[nz-1]-vbar[nz])*D + N2[nz-1]/f0 * etax[1] ) )
L[nz, nz-1] = R * D - S
L[nz, nz] = R * (-D) - S
G[nz, nz-1] = D
G[nz, nz] = - D
################
# 0 < n < nz (interior)
################
for n in range(1,nz):
R = k[i] * ubar[n] + l[j] * vbar[n]
K2 = k[i]**2 + l[j]**2
bf = f0**2 * dzc[n-1]**-1
b_1 = N2[n-1] * dzf[n-1]
b = N2[n] * dzf[n]
B_1 = bf * b_1**-1
B = - (bf * ( b**-1 + b_1**-1 ) + K2)
Bt1 = bf * b**-1
N2Z = (N2[n]*dzf[n])**-1
N2Z_1 = (N2[n-1]*dzf[n-1])**-1
P = ( k[i] * ( beta - bf * ( ubar[n+1] * N2Z
- (N2Z + N2Z_1) * ubar[n]
+ N2Z_1 * ubar[n-1] ) )
- l[j] * bf * ( vbar[n+1] * N2Z
- (N2Z + N2Z_1) * vbar[n]
+ N2Z_1 * vbar[n-1] )
)
L[n, n-1] = R * B_1
L[n, n] = R * B + P
L[n, n+1] = R * Bt1
G[n, n-1] = B_1
G[n, n] = B
G[n, n+1] = Bt1
val, func = eigs( csc_matrix(inv(csc_matrix(G)).dot(csc_matrix(L))),
k=2, which='LI', ncv=100, maxiter=1000 ) # default returns 6 eigenvectors
if i == 0 and j == 0:
omega = np.zeros( (len(val), len(k)), dtype=complex )
psi = np.zeros( (nz+1, len(val), len(k)), dtype=complex )
omega[:, i] = val
psi[:, :, i] = func # Each column is the eigenfunction
omega_imag = omega.imag.max(axis=1)
p = np.argsort(omega_imag)[::-1]
omega = omega[p]
psi = psi[:, p]
print(omega_imag[p])
我第一次在Linux环境中运行此代码时,特征值(即omega_imag)
[ 0.30862832 0.28027858]
但第二次给出了
[ 0.3097162 0.26593829]
我正在运行的环境是:
# packages in environment at /home/takaya/.conda/envs/oceanmodes:
#
backports 1.0 py27_0 defaults
backports-abc 0.4 <pip>
backports.shutil-get-terminal-size 1.0.0 <pip>
backports.ssl-match-hostname 3.4.0.2 <pip>
backports_abc 0.4 py27_0 defaults
cairo 1.12.18 6 defaults
configparser 3.5.0b2 py27_1 defaults
cycler 0.10.0 py27_0 defaults
decorator 4.0.9 py27_0 defaults
entrypoints 0.2 py27_1 defaults
fontconfig 2.11.1 5 defaults
freetype 2.5.5 0 defaults
functools32 3.2.3.2 py27_0 defaults
get_terminal_size 1.0.0 py27_0 defaults
ipykernel 4.3.1 py27_0 defaults
ipython 4.2.0 py27_0 defaults
ipython-genutils 0.1.0 <pip>
ipython_genutils 0.1.0 py27_0 defaults
ipywidgets 4.1.1 py27_0 defaults
jinja2 2.8 py27_0 defaults
jsonschema 2.5.1 py27_0 defaults
jupyter 1.0.0 py27_2 defaults
jupyter-client 4.2.2 <pip>
jupyter-console 4.1.1 <pip>
jupyter-core 4.1.0 <pip>
jupyter_client 4.2.2 py27_0 defaults
jupyter_console 4.1.1 py27_0 defaults
jupyter_core 4.1.0 py27_0 defaults
libgcc 5.2.0 0 defaults
libgfortran 3.0.0 1 defaults
libpng 1.6.17 0 defaults
libsodium 1.0.3 0 defaults
libxml2 2.9.2 0 defaults
markupsafe 0.23 py27_0 <unknown>
matplotlib 1.5.1 np111py27_0 defaults
mistune 0.7.2 py27_0 defaults
mkl 11.3.1 0 defaults
nbconvert 4.2.0 py27_0 defaults
nbformat 4.0.1 py27_0 defaults
notebook 4.2.0 py27_0 defaults
numpy 1.11.0 py27_0 defaults
openssl 1.0.2h 0 defaults
path.py 8.2.1 py27_0 defaults
pexpect 4.0.1 py27_0 defaults
pickleshare 0.5 py27_0 defaults
pip 8.1.1 py27_1 defaults
pixman 0.32.6 0 defaults
ptyprocess 0.5 py27_0 defaults
pycairo 1.10.0 py27_0 defaults
pygments 2.1.3 py27_0 defaults
pyparsing 2.1.1 py27_0 defaults
pyqt 4.11.4 py27_1 defaults
python 2.7.11 0 defaults
python-dateutil 2.5.2 py27_0 defaults
pytz 2016.3 py27_0 defaults
pyzmq 15.2.0 py27_0 defaults
qt 4.8.7 1 defaults
qtconsole 4.2.1 py27_0 defaults
readline 6.2 2 <unknown>
scipy 0.17.0 np111py27_3 defaults
setuptools 20.7.0 py27_0 defaults
simplegeneric 0.8.1 py27_0 defaults
singledispatch 3.4.0.3 py27_0 defaults
sip 4.16.9 py27_0 defaults
six 1.10.0 py27_0 defaults
sqlite 3.9.2 0 defaults
ssl_match_hostname 3.4.0.2 py27_1 defaults
terminado 0.5 py27_1 defaults
tk 8.5.18 0 http://repo.continuum.io/pkgs/free/linux-64/tk-8.5.18-0.tar.bz2
tornado 4.3 py27_0 defaults
traitlets 4.2.1 py27_0 defaults
wheel 0.29.0 py27_0 defaults
zeromq 4.1.3 0 defaults
zlib 1.2.8 0 http://repo.continuum.io/pkgs/free/linux-64/zlib-1.2.8-0.tar.bz2
答案 0 :(得分:1)
ARPACK中的起始向量(如果未指定)是随机的。您可以尝试指定v0
参数以使用固定的参数。
此外,如果计算对初始值敏感(病态条件),则可能会在不同的运行中获得不同的结果,因为浮点舍入误差可能会受到可能变化的外部因素的影响(内存对齐,BLAS中的线程等)。
换句话说,现在确保确定性结果不仅仅需要编写程序并运行它 - 您可能需要使用额外的C / Fortran编译器选项编译所有内容(Numpy,scipy等),这些选项可以产生更严格的确定性浮动 - 点算术机器码。您可能还需要在优化的线性algera库(MKL,Openblas)中控制线程等的使用。最简单的选择通常是确保计算不会受到条件限制,因此小扰动并不重要 - 但在ARPACK的情况下,可能并非直接在您的控制范围内,除非在意义上可能有可能以这种变化无关紧要的方式构建您正在尝试进行的计算(或者您根本不计算特征值)。