CODEWITHSUNDEEP
pythonnumpy
GrayLiterature
GrayLiterature

Reputation: 391

Why is numpy.exp() giving an Overflow warning in this case?

I have the following reproducible script:

import numpy as np

c0, c1 = 0.000032, 3

S_test = np.ones((3,3), dtype=float)*300000
S_test[1,1] = 2000


np.where(S_test > 2000, 1, 1 - 1 / np.exp((c0 * S_test) ** c1) )

RuntimeWarning: overflow encountered in exp

array([[1.00000000e+00, 1.00000000e+00, 1.00000000e+00],
       [1.00000000e+00, 2.62109643e-04, 1.00000000e+00],
       [1.00000000e+00, 1.00000000e+00, 1.00000000e+00]])

I am trying to figure out 1) Why I am getting this overflow error and 2) how can it be dealt with in a way that doesn't just ignore it because i'm not entirely certain what the consequences of ignoring it are in the first place.

The np.where() should first recognize where S_test > 2000 and then return 1 for those values, alternatively, where this condition does not hold np.where() should return 1 - 1 / np.exp((c0 * S_test) ** c1)

I have tried to separately compute 1 - 1 / np.exp((c0 * 2000) ** c1) and this evaluates just fine and as I expected.

Any insight would be great.

Upvotes: 1

Views: 242

Answers (2)

Han-Kwang Nienhuys
Han-Kwang Nienhuys

Reputation: 3254

You get the error because all arguments of np.where are fully evaluated. You can do this:

S_test_capped = np.where(S_test > 2000, 2000, S_test)
np.where(S_test > 2000, 1, 1 - 1 / np.exp((c0 * S_test_capped) ** c1) )

or just silence the warnings:

old_settings = np.seterr(over='ignore')
np.where(S_test > 2000, 1, 1 - 1 / np.exp((c0 * S_test) ** c1) )
np.seterr(**old_settings) # restore warning settings

Upvotes: 0

ev-br
ev-br

Reputation: 26050

np.where evaluates both operands.

If you really want to avoid the warning, you can do e.g. something like this: https://github.com/scipy/scipy/blob/master/scipy/_lib/_util.py#L31

Upvotes: 4

Related Questions