0

I was looking at some code of a neural network written in python with numpy and I noticed that when passing the results of the neurons through the activation function:

def sigmoid(x):
    return 1 / ( 1 + np.exp(-x))

Instead of passing the weights one by one that person called the function with an np.array as the argument!

Like so:

import numpy as np

def sigmoid(x):
    return 1.0/(1.0 + np.exp(-x))

x = np.zeros((2,2))

print "Before\n",x
x = sigmoid(x)
print "After\n",x

To my surprise this worked. I have always called the activation function with a loop for each weight individually, since the function takes a single number as input.

output:

Before
[[ 0.  0.]
 [ 0.  0.]]
After
[[ 0.5  0.5]
 [ 0.5  0.5]]

I also tried some other operations and they seem to work fine.

Is this 'correct'? Are there cases where this doesn't work or are numpy arrays designed to do that? What I really want to know is, can I use this or am I better off doing what I was doing before?

Giorgos Diamantopoulos
  • 13
  • 5
  • Yes, it's correct, and it's called vectorization. – user3483203 Jun 06 '19 at 20:16
  • Numpy is especially good at broadcasting functions to vectorize operations like this – G. Anderson Jun 06 '19 at 20:17
  • Check this out: https://stackoverflow.com/questions/47755442/what-is-vectorization – Sheldore Jun 06 '19 at 20:18
  • Individually the operations in the expression work with `numpy` arrays, `np.exp`, `+` (`np.add`) and `/`, `np.divide`. So jointly they also work with an array. There are things that don't work with arrays, for example `if/else` expressions. Look at the basic and universal functions in this turtorial, https://docs.scipy.org/doc/numpy/user/quickstart.html#basic-operations – hpaulj Jun 06 '19 at 20:46

0 Answers0