# Improve performance of math function applied to arrays

This block of code is already somewhat optimized thanks to some answers given over at Stack Overflow. In the last question I made (Improve performance of function without parallelization) the code was marginally improved but I later had to re-write a little bit of it and the changes proposed didn't apply anymore.

I'm looking at optimizing this block of code without resorting to parallelization, other than that I'm pretty much open to using any package out there.

I think this is perhaps the correct site to post this since it's really more of a question about how can I improve my code rather than something not working with it.

Here's the MWE (minimum working example):

import numpy as np
import timeit

def random_data(N):
# Generate some random data.
return np.random.uniform(0., 10., N)

# Data lists.
array1 = np.array([random_data(4) for _ in range(1000)])
array2 = np.array([random_data(4) for _ in range(2000)])

def func():
lst = []

for elem in array1:

# Define factors.
a_01, a_11 = max(elem, 1e-10), max(elem, 1e-10)
a_02, a_12 = array2[:,1], array2[:,3]
# Combine the factors defined above.
a_0 = a_01**2 + a_02**2
a_1 = a_11**2 + a_12**2
Bs, Cs = -0.5/a_0, -0.5/a_1

# Perform operations.
A = 1./(np.sqrt(a_0*a_1))
B = Bs*(elem-array2[:,0])**2
C = Cs*(elem-array2[:,2])**2
ABC = A*np.exp(B+C)

lst.append(max(ABC.sum(), 1e-10))

return lst

# Time function.
func_time = timeit.timeit(func, number=10000)
print func_time

• What is this code supposed to do? What problem are you solving? – Gareth Rees Feb 18 '14 at 18:55
• @GarethRees it's a bit hard to explain since this is a tiny part of a much larger code I'm working on. In short, it calculates the probability that each element of array1 "belongs" (or "is related") to array2. – Gabriel Feb 18 '14 at 18:58

I couldn't make much improvement on this, sorry. Since you didn't explain what the code is supposed to do, there was no opportunity for algorithmic insight, so the only changes I could make were purely mechanical transformations with no understanding.

One obvious thing to do is to replace:

def random_data(N):
# Generate some random data.
return np.random.uniform(0., 10., N)

array1 = np.array([random_data(4) for _ in range(1000)])
array2 = np.array([random_data(4) for _ in range(2000)])


with

array1 = np.random.uniform(0, 10, (1000, 4))
array2 = np.random.uniform(0, 10, (2000, 4))


which is about 50 times faster. (Note that you don't have to write 0. and 10. here: if you read the documentation for numpy.random.uniform you'll see that the parameters low and high have type "float", so which whatever values you provide will be coerced to floats.)

Other observations are:

• Operations on array2 don't depend on elem and so can be hoisted out of the loop.
• It's quicker to make one call to numpy.split than to slice array2 four times.
• It's quicker to call numpy.square than use the expression A ** 2.
• The multiplication by -0.5 can be saved until after B+C, reducing the number of operations.
• A single division X / Y is cheaper than calculating a reciprocal A = 1 / Y and multiplying A * X.

Applying all of these, my best effort is as follows:

def func2():
lst = []

epsilon = 1e-10

P = np.split(array1, 4, axis=1)
P = np.square(np.maximum(P, epsilon))
P = np.square(np.maximum(P, epsilon))
P = np.hstack(P)

Q = np.split(array2, 4, axis=1)
Q = np.square(Q)
Q = np.square(Q)

for elem in P:
a_0 = elem + Q
a_1 = elem + Q

B = np.square(elem - Q) / a_0
C = np.square(elem - Q) / a_1
ABC = np.exp(-0.5 * (B + C)) / np.sqrt(a_0 * a_1)

lst.append(max(ABC.sum(), epsilon))

return lst


>>> timeit(func2, number=100) / timeit(func, number=100)
0.7512780299658295


For comparison, here's the fully-vectorized version of your code:

def func3():
epsilon = 1e-10
P = array1.reshape(-1, 1, 4)
Q = array2
R = np.square(np.maximum(P[...,(1,3)], epsilon))
S = np.square(Q[:,(1,3)])
T = R + S
U = np.square(P[...,(0,2)] - Q[:,(0,2)]) / T
V = np.exp(-0.5 * U.sum(axis=2)) / np.sqrt(T.prod(axis=2))
return np.maximum(V.sum(axis=1), epsilon)


This is about 25% slower than your code because it's O(n2) in space as well as time, but it's substantially simpler, so it might be a better place to start when making algorithmic improvements.

• Thank you very much for your answer @Gareth, I'll try it right now and report back the results. I'm sorry I can't go into much detail but my actual code is really large, I did comment on what this block does in any case (right below your comment in the original question) – Gabriel Feb 19 '14 at 13:42
• @Gabriel: I was hoping for a more technical explanation like "this computes the radial basis function kernel" or whatever. – Gareth Rees Feb 19 '14 at 13:50
• Gareth this article goo.gl/7krj3L can give you an idea of what I'm after, particularly eq. 4. Sorry I can't be more succinct but it really is quite a large issue to be able to discuss it here. – Gabriel Feb 19 '14 at 14:12
• @Gabiel: Thank you, that's more like it. I was wondering why the input arrays seemed to consist of multiple kinds of interleaved data. – Gareth Rees Feb 19 '14 at 14:14

Well, some things I noticed:

    # Define factors.
a_01, a_11 = max(elem, 1e-10), max(elem, 1e-10)
a_02, a_12 = array2[:,1], array2[:,3]


a_02 and a_12 are static across the loop, no need to redefine them every time. a_01 and a_12 could be defined array-wise using np.maximum:

a_01 = np.maximum(array1[:,1], np.array(1e-10 for _ in xrange(len(array1))
a_11 = np.maximum(array1[:,3], np.array(1e-10 for _ in xrange(len(array1))


However, these are trivial improvements; function call overhead is often blamed for much in the Python optimisation world, but there is no way that removing 2n calls to max will improve things here. With that said, following the static data path further, we come across an instance of

a_0 = a_01**2 + a_02**2
a_1 = a_11**2 + a_12**2


a_022 = a_02**2
a_122 = a_12**2


at the beginning of the function, we can avoid n multiplications by replacing later occurrences of a_02**2 with a_022 (and likewise for a_12), which reduces the running time of the program by about 5% on my machine. That isn't too impressive, though I'm sure you should be able to knock that down a fair bit by doing more thing using numpy functions. However, my numpy skills evaporate when it comes to 2d arrays, so I can't help.

It is worth mentioning that the algorithm you are using is O(n^2) at least, so wouldn't get your hopes up too high. Otherwise, I'd recommend looking at Cython as this seems like a problem it would excel at, though you've mentioned that parallelism isn't an option, so I assume using C extensions won't be either. However, if it is, I think you will find great time savings there.

This was the final code I came up with:

def func2():
lst = []
# Define factors.
a_02, a_12 = array2[:,1], array2[:,3]
a_022 = a_02 ** 2
a_122 = a_12 ** 2
e10 = np.array([1e-10 for _ in xrange(len(array1))])
a_01 = np.maximum(array1[:,1], e10)
a_11 = np.maximum(array1[:,3], e10)
for i, elem in enumerate(array1):
# Combine the factors defined above.
a_0 = a_01[i]**2 + a_022
a_1 = a_11[i]**2 + a_122
Bs, Cs = -0.5/a_0, -0.5/a_1
# Perform operations.
A = 1./(np.sqrt(a_0*a_1))
B = Bs*(elem-array2[:,0])**2
C = Cs*(elem-array2[:,2])**2
ABC = A*np.exp(B+C)
lst.append(max(ABC.sum(), 1e-10))
return lst

N = 100
func_time = timeit.timeit(func, number=N)
print "Original\t", func_time / N
print "-" * 50

func2_time = timeit.timeit(func2, number=N)
print "Revision 1\t", func2_time / N
print "Improvement\t", str(int((func_time - func2_time) / func_time * 100)) + "%"
print "-" * 50

• I'm getting a result of Improvement -90015% with this new function. Can you confirm this? – Gabriel Feb 18 '14 at 21:21
• I think that might be because the indentation of the return lst in the original function is incorrect; I presumed that was a typo as most people don't want to exit a loop after only one iteration. – BluePeppers Feb 18 '14 at 21:27
• This was the exact code I tested: gist.github.com/bluepeppers/9080540 – BluePeppers Feb 18 '14 at 21:32
• You are right @BluePeppers, the indent was incorrect in my original code (now fixed) With the correct code I get a ~4% improvement. Thank you very much! I'll wait a bit to see if someone else can top your code otherwise I'll mark your answer as accepted. Cheers. – Gabriel Feb 18 '14 at 21:34
• Take advantage of broadcasting and write np.maximum(X, 1e-10) instead of np.maximum(X, np.array([1e-10 for _ in range(len(X))])) – Gareth Rees Feb 19 '14 at 11:31