在 MATLAB 上改进 Numpy 中复数矩阵求和元素的更好方法

Question

我试图将下面的 MATLAB 代码重写为 Python，发现我的 Python 代码（2.7 秒）比 MATLAB（1.2 秒）慢。我尝试了许多不同的方法，包括模块 numba，但还没有运气。如何使 Python 代码更快？

MATLAB 代码：

szA=[1024,1280]; HfszA=[512,640];

[aPx,aPy]=meshgrid(-HfszA(2):HfszA(2)-1,-HfszA(1):HfszA(1)-1);

img=randi(255,1024,1280);

fx=rand(); fy=rand();

tic

for i=1:20

    F=abs(sum(sum(img.*exp(-1i*2*pi*(fx*aPx+fy*aPy)))));

end

toc

蟒蛇代码：

import numpy as np

import time

szA=[1024,1280]; HfszA=[512,640]

aPx,aPy=np.meshgrid(np.arange(-HfszA[1],HfszA[1]),np.arange(-HfszA[0],HfszA[0]))

img=np.array(np.random.randint(256,size=(1024,1280)))

fx=np.random.rand()

fy=np.random.rand()

start = time.time()

for i in range(20):

    F=abs(np.sum(img*np.exp(-1j*2*np.pi*(fx*aPx+fy*aPy))))

end = time.time()

print("Elapsed (after compilation) = %s" % (end - start))

print(F)

Answer 1

请始终发布您迄今为止尝试过的内容。关于您的 Numba 版本，我认为您做了一些导致性能不佳的事情。例子import numpy as npimport numba as nbimport [e-mail](fastmath=True)def your_function(fx,fy,aPx,aPy,img):  pi=np.pi  sum=0.  for i in range(aPx.shape[0]):    for j in range(aPx.shape[1]):      sum+=img[i,j]*np.exp(-1j*2*pi*(fx*aPx[i,j]+fy*aPy[i,j]))  return np.abs(sum)@nb.njit(fastmath=True,parallel=True)def your_function_p(fx,fy,aPx,aPy,img):  pi=np.pi  sum=0.  for i in nb.prange(aPx.shape[0]):    for j in range(aPx.shape[1]):      sum+=img[i,j]*np.exp(-1j*2*pi*(fx*aPx[i,j]+fy*aPy[i,j]))  return np.abs(sum)#The function gets compiled at the first call#you may also use cache=True, which only works in single threaded codeF=your_function(fx,fy,aPx,aPy,img)start = time.time()for i in range(20):    F=your_function(fx,fy,aPx,aPy,img)end = time.time()print("Elapsed (after compilation) = %s" % (end - start))print(F)F=your_function_p(fx,fy,aPx,aPy,img)start = time.time()for i in range(20):    F=your_function_p(fx,fy,aPx,aPy,img)end = time.time()print("Elapsed (after compilation) = %s" % (end - start))print(F)计时（4C/8T）your_version: 2.45sNumba single threaded: 0.17sNumba parallel: 0.07s