[SOLVED] Combine several NumPy "where" statements to one to improve performance


I am trying to speed up a code that is using Numpy’s where() function. There are two calls to where(), which return an array of indices for where the statement is evaluated as True, which are then compared for overlap with numpy’s intersect1d() function, of which the length of the intersection is returned.

import numpy as np

def find_match(x,y,z):

    A = np.where(x == z)
    B = np.where(y == z)
    #A = True
    #B = True
    return len(np.intersect1d(A,B))

N = np.power(10, 8)
M = 10

X = np.random.randint(M, size=N)
Y = np.random.randint(M, size=N)
Z = np.random.randint(M, size=N)



  • This code takes about 8 seconds on my laptop. If I replace both the np.where() with A=True and B=True, then it takes about 5 seconds. If I replace only one of the np.where() then it takes about 6 seconds.

  • Scaling up, by switching to N = np.power(10, 9), the code takes 87 seconds. Replacing both the np.where() statements results in the code takes 51 seconds. Replacing just one of the np.where() takes about 61 seconds.

How can I merge the two np.where statements that can speed up the code?

This is already an improved version of the code where the speed was increased ~4x by replacing a slower lookup with for-loops. Multiprocessing will be used at a higher level in this code, so I can’t apply it also here.

For the record, the actual data are strings, so doing integer math won’t be helpful.

Version info:

Python 3.9.1 (default, Jan  8 2021, 17:17:43) 
[Clang 12.0.0 (clang-1200.0.32.28)] on darwin
>>> import numpy
>>> print(numpy.__version__)


Does this help?

def find_match2(x, y, z):
    return len(np.nonzero(np.logical_and(x == z, y == z))[0])

Sample run:

In [227]: print(find_match(X,Y,Z))

In [228]: print(find_match2(X,Y,Z))

In [229]: %timeit find_match(X,Y,Z)
2.37 s ± 70.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [230]: %timeit find_match2(X,Y,Z)
272 ms ± 9.64 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

I’ve added np.random.seed(210) before creating the arrays for the sake of reproducibility.

Answered By – Tonechas

Answer Checked By – Robin (BugsFixing Admin)

Leave a Reply

Your email address will not be published. Required fields are marked *