I want to use vectorization to do some computation on numpy.ndarray. Suppose I have the following vectorized function:
import numpy as np
fun = lambda x:x[0]+x[1]
fun = np.vectorize(fun)
and the following numpy.ndarray
a = range(10)
b = range(10)
c = np.array([a,b])
When I apply
result = fun(c)
I obtain the following error
IndexError: invalid index to scalar variable.
\Why is this the case and how should I fix it?
np.vectorize feeds scalar values to your function. It iterates on the input arrays, broadcasting if needed, and feeds func scalars, not arrays or lists. It then collects the values in a new array of shape and dtype that it deduces.
For example:
In [108]: fun = lambda x,y: x+y
...: fun = np.vectorize(fun)
In [110]: a=np.arange(10); b=np.arange(10)
In [111]: fun(a,b)
Out[111]: array([ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18])
It is not 'vectorize' in the sense of turning your function into fast compiled code. It's a convenience, saving you some work in setting up an interation.
I'm sure your fun is just a example, but as written it is already 'vectorized'
In [112]: (lambda x,y: x+y)(a,b)
Out[112]: array([ 0, 2, 4, 6, 8, 10, 12, 14, 16, 18])
Expressing your calculation with numpy primitives, without iteration, is the true 'vectorization'. That isn't always possible, but if you feel you must fall back on np.vectorize remember that
otypes if possible
np.vectorizeif you are looking for performance. Look intoufuncsrather.funto do? A vectorized function will apply to every element of an array, so how isx[0] + x[1]suppose to work whenxis a number? Also, I second the call not to usenp.vectorize, it's essentially a thin wrapper around a vanilla Python loop.