Building a small numpy array from individual values: Fast and readable method?

Question

I found that a bottleneck in my program is the creation of numpy arrays from a list of given values, most commonly putting four values into a 2x2 array. There is an obvious, easy-to-read way to do it:

my_array = numpy.array([[1, 3], [2.4, -1]])

which takes 15 us -- very very slow since I'm doing it millions of times.

Then there is a far faster, hard-to-read way:

my_array = numpy.empty((2,2))
my_array[0,0] = 1
my_array[0,1] = 3
my_array[1,0] = 2.4
my_array[1,1] = -1

This is 10 times faster, at just 1 us.

Is there any method that is BOTH fast AND easy-to-read?

What I tried so far: Using asarray instead of array makes no difference; passing dtype=float into array also makes no difference. Finally, I understand that I can do it myself:

def make_array_from_list(the_list, num_rows, num_cols):
    the_array = np.empty((num_rows, num_cols))
    for i in range(num_rows):
        for j in range(num_cols):
            the_array[i,j] = the_list[i][j]
    return the_array

This will create the array in 4us, which is medium readability at medium speed (compared to the two approaches above). But really, I cannot believe that there is not a better approach using built-in methods.

Thank you in advance!!

Answer 1

This is a great question. I can't find anything which will approach the speed of your completely unrolled solution (edit @BiRico was able to come up with something close. See comments and update :). Here are a bunch of different options that I (and others) came up with and associated timings:

import numpy as np

def f1():
    "np.array + nested lists"
    my_array = np.array([[1, 3], [2.4, -1]])

def f2():
    "np.array + nested tuples"
    my_array = np.array(((1, 3), (2.4, -1)))

def f3():
    "Completely unrolled"
    my_array = np.empty((2,2),dtype=float)
    my_array[0,0] = 1
    my_array[0,1] = 3
    my_array[1,0] = 2.4
    my_array[1,1] = -1

def f4():
    "empty + ravel + list"
    my_array = np.empty((2,2),dtype=float)
    my_array.ravel()[:] = [1,3,2.4,-1]

def f5():
    "empty + ravel + tuple"
    my_array = np.empty((2,2),dtype=float)
    my_array.ravel()[:] = (1,3,2.4,-1)

def f6():
    "empty + slice assignment"
    my_array = np.empty((2,2),dtype=float)
    my_array[0,:] = (1,3)
    my_array[1,:] = (2.4,-1)

def f7():
    "empty + index assignment"
    my_array = np.empty((2,2),dtype=float)
    my_array[0] = (1,3)
    my_array[1] = (2.4,-1)

def f8():
    "np.array + flat list + reshape"
    my_array = np.array([1, 3, 2.4, -1]).reshape((2,2))

def f9():
    "np.empty + ndarray.flat  (Pierre GM)"
    my_array = np.empty((2,2), dtype=float)
    my_array.flat = (1,3,2.4,-1)

def f10():
    "np.fromiter (Bi Roco)"
    my_array = np.fromiter((1,3,2.4,-1), dtype=float).reshape((2,2))

import timeit
results = {}
for i in range(1,11):
    func_name = 'f%d'%i
    my_import = 'from __main__ import %s'%func_name
    func_doc = globals()[func_name].__doc__
    results[func_name] = (timeit.timeit(func_name+'()',
                                        my_import,
                                        number=100000),
                          '\t'.join((func_name,func_doc)))

for result in sorted(results.values()):
    print '\t'.join(map(str,result))

And the important timings:

On Ubuntu Linux, Core i7:

0.158674955368  f3  Completely unrolled
0.225094795227  f10 np.fromiter (Bi Roco)
0.737828969955  f8  np.array + flat list + reshape
0.782918930054  f5  empty + ravel + tuple
0.786983013153  f9  np.empty + ndarray.flat  (Pierre GM)
0.814703941345  f4  empty + ravel + list
1.2375421524    f7  empty + index assignment
1.32230591774   f2  np.array + nested tuples
1.3752617836    f6  empty + slice assignment
1.39459013939   f1  np.array + nested lists