Search code examples
pythonpython-3.xnumpydivision

Faster division in python

I have this lambda-function:

transfer = lambda x: 1.0 / (1.0 + np.e ** (-2.0*x))

This function is called very often (more than 80 000 times). (I think division 1 by the expression is not effective, but cannot find good alternative for that.) Can I make this faster?

Solution

  • Rewrite your code such that you can collect all x values in a numpy array first and then use one operation on your numpy array directly

    import time
import numpy as np

data = np.random.random(size=100000)

result = list()

transfer = lambda x: 1.0 / (1.0 + np.e ** (-2.0 * x))

t1 = time.perf_counter()
for value in data:
    y = transfer(value)
    result.append(y)
result = np.array(result)
t2 = time.perf_counter()
print(f"time lambda = {t2 - t1} with sum {result.sum()}")

t3 = time.perf_counter()

# this line is the numpy expression of your transfer function
results2 = 1 / (1 + np.e ** (-2 * data))
t4 = time.perf_counter()
print(f"time numpy = {t4 - t3} with sum {results2.sum()}")

    Output of this code is:

    time lambda = 0.09890314500080422 with sum 71682.58495450807
time numpy = 0.0029907969874329865 with sum 71682.58495450807

    So the numpy approach is about 35 times faster