I have these two arrays:
import numpy as np
a = np.array([0, 10, 20])
b = np.array([20, 30, 40, 50])
I´d like to add both in the following way while avoiding slow Python for loops:
c = array([20, 30, 40, 50, 30, 40, 50, 60, 40, 50, 60, 70])
It means, for the first element of "a" add all the elements of b,then for the second element and so on.
I know it seems quite easy, but for huge arrays Python loops are too slow.
Thanks.
Here you go:
In [17]: a = np.array([0, 10, 20])
In [18]: b = np.array([20, 30, 40, 50])
In [19]: (a.reshape(-1, 1) + b).ravel()
Out[19]: array([20, 30, 40, 50, 30, 40, 50, 60, 40, 50, 60, 70])
Here are the details.
a.reshape(-1, 1) converts a to an array with shape (3, 1):
In [20]: a.reshape(-1, 1)
Out[20]:
array([[ 0],
[10],
[20]])
When b is added to that, broadcasting applies, which in effect does an "outer sum" (i.e. adds all the pairwise combinations), forming an array with shape (3, 4):
In [21]: a.reshape(-1, 1) + b
Out[21]:
array([[20, 30, 40, 50],
[30, 40, 50, 60],
[40, 50, 60, 70]])
The ravel() method flattens the result into a one-dimensional array:
In [22]: (a.reshape(-1, 1) + b).ravel()
Out[22]: array([20, 30, 40, 50, 30, 40, 50, 60, 40, 50, 60, 70])
See @HYRY's answer for an even more concise version.
you can use outer method of ufunc:
np.add.outer(a, b).ravel()
One possible and only memory not time intensive solution is using np.repeat and np.resize to repeat and resize the arrays a and b to the size of the resulting shape first and then simply add those two arrays.
Code:
import numpy as np
a = np.array([0, 10, 20])
b = np.array([20, 30, 40, 50])
def extend_and_add(a, b):
return np.repeat(a, len(b)) + np.resize(b, len(a)*len(b))
So extend_and_add(a, b) returns:
extend_and_add(a, b)
> array([20, 30, 40, 50, 30, 40, 50, 60, 40, 50, 60, 70])
Explanation:
Basically np.repeat(a, len(b)) repeats:
a
> array([ 0, 10, 20])
to
np.repeat(a, len(b))
> array([ 0, 0, 0, 0, 10, 10, 10, 10, 20, 20, 20, 20])
after this you need the second array resized with np.resize:
b
> array([20, 30, 40, 50])
is resized to:
np.resize(b, len(a)*len(b))
> array([20, 30, 40, 50, 20, 30, 40, 50, 20, 30, 40, 50])
Now we can simply add the arrays:
array([ 0, 0, 0, 0, 10, 10, 10, 10, 20, 20, 20, 20])
+
array([20, 30, 40, 50, 20, 30, 40, 50, 20, 30, 40, 50])
returns:
array([20, 30, 40, 50, 30, 40, 50, 60, 40, 50, 60, 70])
First you need to specify the array type, if you use a constructor like that. For instance for integers, use that:
a = array("i",[0, 10, 20]) # signed int type
b = array("i",[20, 30, 40, 50])
Them you might want to use while loops with counters, it is more complex than for, but avoids the for loop.
from array import array
a = array("i",[0, 10, 20]) # signed int
b = array("i",[20, 30, 40, 50])
c = array("i",[])
count1 = 0
while count1 < len(a):
count2 = 0
while count2 < len(b):
c.append(a[count1]+b[count2])
count2 += 1
count1 += 1
print(c)
numpy; no need to switch to the other Python array module.
forloops?for's in list comprehensions count? ;)