1

This is related to several questions that discuss the speed of numpy vs Matlab. However most of them have several matrix operations than a single operation. E.g. Difference on performance between numpy and matlab

For me, the time numpy take just to invert a random matrix is approximately 5 times slower than that of matlab.

Here is the matlab script,

N = 1000;

B = randn(N,N);
h = tic;
T = 40;
for i=1:40

    Rinv = (B)^(-1);
end
toc(h)/40

This gives an average values of 0.08 seconds approximately.

While this python script gives 0.4 seconds (approx). 

import numpy as np 
from numpy import linalg as LA
import time 

N=1000
R = np.random.random((N,N))
T=40

t1 = time.clock()
for i in range(0,T):
    Rinv = LA.inv(R)
t2 = time.clock()
print 'avg time for inverse ',(t2-t1)/T

Is there any reason for this, or anyway to improve python performance ? I have already implemented my work on Python and I am worried whether I will have to port all my code to matlab. I am working on Ubuntu 16.04, Python 2.7, Matlab R2016b.

I have read that the time is not a good module for execution time comparisons, I feel this is something more than that.

Community
  • 1
  • 1
user2939212
  • 405
  • 1
  • 6
  • 14
  • 2
    I myself would also choose 40 random matrices rather than a single one. – President James K. Polk Jan 22 '17 at 17:02
  • 4
    You should use `timeit` to get a reliable benchmark in MATLAB. Taking the inverse of the same variable and storing it in the same variable within a loop is hardly going to be a good estimate of execution time especially considering JIT acceleration – Suever Jan 22 '17 at 17:07
  • 1
    This is a rather poor benchmark - in both cases you should be removing the loop from the timing, and using a more reliable timing mechanism. – miradulo Jan 22 '17 at 17:12
  • @JamesKPolk I thought since generating random numbers is different for two methods. It may not be fair to compare just the inversion time. Will check that anyway. – user2939212 Jan 22 '17 at 17:20
  • @Suever Thanks. I am trying to get the results with timeit module something that I haven't used. – user2939212 Jan 22 '17 at 17:37

1 Answers1

4

On my computer (Windows, python 3.5, numpy 1.11.2) :

In [6]: %timeit inv(a)
10 loops, best of 3: 86 ms per loop

edit:

or, without Ipython:

>>>timeit.timeit('inv(a)','from __main__ import inv,a',number=100)/100

which is similar to Matlab.

to know what code is used in the background, check it :

In [12]: np.__config__.show() 
blas_mkl_info:
include_dirs = ['c:/users/bruno/miniconda3\\Library\\include']
libraries = ['mkl_core_dll', 'mkl_intel_lp64_dll', 'mkl_intel_thread_dll']
...
B. M.
  • 18,243
  • 2
  • 35
  • 54
  • Thanks. I am trying to run it using timeit, can you please post the code ? I wasn't able get it work using timeit module. – user2939212 Jan 22 '17 at 18:26
  • 3
    B.M. used the IPython or Jupyter console that provides a so-called "magic" feature for timing. – Pierre de Buyl Jan 22 '17 at 18:28
  • `t= timeit.Timer("B=np.linalg.inv(R)","import numpy as np;R=np.random.random((1000,1000))");t.timeit(number=100) ` This gave me 9.95. Which is close to 90 ms on average. So I think using time module was the problem. Never though it has such an importance. Thank you ! – user2939212 Jan 22 '17 at 19:06
  • 2
    I edit my post to manual timeit . I encourage you ta adopt Ipython ;) – B. M. Jan 22 '17 at 19:12