RuntimeWarning: Divide by Zero error: How to avoid? PYTHON, NUMPY

8
votes

I am running in to RuntimeWarning: Invalid value encountered in divide

 import numpy
 a = numpy.random.rand((1000000, 100))
 b = numpy.random.rand((1,100))
 dots = numpy.dot(b,a.T)/numpy.dot(b,b)
 norms = numpy.linalg.norm(a, axis =1)
 angles = dots/norms ### Basically I am calculating angle between 2 vectors

There are some vectors in my a which have norm as 0. so while calculating angles it is giving runtime warning.

Is there a one line pythonic way to compute angles while taking into account norms which are 0?

angles =[i/j if j!=0 else -2 for i,j in zip(dots, norms)] # takes 10.6 seconds

But it takes a lot of time. Since all angles will have values between 1 and -1 and I need only 10 max values this will help me. This takes around 10.6 seconds which is insane.

5
pythonnumpy
can you filter norms before passing it? - Padraic Cunningham
No I want to preserve the indices because I want indices of 10 largest angles in the end which are further matched in other list. So preserving indices is must for me. - pg2455
With the code you posted, the angles variable will have NaNs where norms is zero. Isn't that what you want? Note that if you want None instead of NaN, then you'll need to make angles of dtype object, which is bad for performance... (Also, it is possible to just suppress RuntimeWarnings. Is that what you want?) - unutbu
I want specific value at place where NaN is encountered. So that I can manipulate it while sorting. - pg2455
nans sort higher than anything else so when you are sorting you can just strip the top np.count_nonzero(np.isnan(d.ravel()) values - jtaylor

5 Answers

13
votes

you can ignore warings with the np.errstate context manager and later replace nans with what you want:

import numpy as np
angle = np.arange(-5., 5.) 
norm = np.arange(10.)
with np.errstate(divide='ignore'):
    print np.where(norm != 0., angle / norm, -2)
# or:
with np.errstate(divide='ignore'):
    res = angle/norm
res[np.isnan(res)] = -2
3
votes

You could use angles[~np.isfinite(angles)] = ... to replace nan values with some other value.

For example:

In [103]: angles = dots/norms

In [104]: angles
Out[104]: array([[ nan,  nan,  nan, ...,  nan,  nan,  nan]])

In [105]: angles[~np.isfinite(angles)] = -2

In [106]: angles
Out[106]: array([[-2., -2., -2., ..., -2., -2., -2.]])

Note that division by zero may result in infs, rather than nans, 

In [140]: np.array([1, 2, 3, 4, 0])/np.array([1, 2, 0, -0., 0])
Out[140]: array([  1.,   1.,  inf, -inf,  nan])

so it is better to call np.isfinite rather than np.isnan to identify the places where there was division by zero.

In [141]: np.isfinite(np.array([1, 2, 3, 4, 0])/np.array([1, 2, 0, -0., 0]))
Out[141]: array([ True,  True, False, False, False], dtype=bool)

Note that if you only want the top ten values from an NumPy array, using the np.argpartition function may be quicker than fully sorting the entire array, especially for large arrays:

In [110]: N = 3

In [111]: x = np.array([50, 40, 30, 20, 10, 0, 100, 90, 80, 70, 60])

In [112]: idx = np.argpartition(-x, N)

In [113]: idx
Out[113]: array([ 6,  7,  8,  9, 10,  0,  1,  4,  3,  2,  5])

In [114]: x[idx[:N]]
Out[114]: array([100,  90,  80])

This shows np.argpartition is quicker for even only moderately large arrays:

In [123]: x = np.array([50, 40, 30, 20, 10, 0, 100, 90, 80, 70, 60]*1000)

In [124]: %timeit np.sort(x)[-N:]
1000 loops, best of 3: 233 µs per loop

In [125]: %timeit idx = np.argpartition(-x, N); x[idx[:N]]
10000 loops, best of 3: 53.3 µs per loop
3
votes

In newer versions of numpy there is a third alternative option that avoids needing to use the errstate context manager.

All Numpy ufuncs accept an optional "where" argument. This acts slightly differently than the np.where function, in that it only evaluates the function "where" the mask is true. When the mask is False, it doesn't change the value, so using the "out" argument allows us to preallocate any default we want.

import numpy as np

angle = np.arange(-5., 5.)
norm = np.arange(10.)

# version 1
with np.errstate(divide='ignore'):
    res1 = np.where(norm != 0., angle / norm, -2)

# version 2
with np.errstate(divide='ignore'):
    res2 = angle/norm
res2[np.isinf(res2)] = -2

# version 3
res3 = -2. * np.ones(angle.shape)
np.divide(angle, norm, out=res3, where=norm != 0)

print(res1)
print(res2)
print(res3)

np.testing.assert_array_almost_equal(res1, res2)
np.testing.assert_array_almost_equal(res1, res3)
2
votes

You want to be using np.where. See the documentation.

angles = np.where(norms != 0, dots/norms, -2)

Angles will consist of downs/norms whenever norms != 0, and will be -2 otherwise. You will still get the RuntimeWarning, as np.where will still calculate the entire vector dots/norms internally, but you can safely ignore it.

0
votes

You can use np.where( condition ) to perform a conditional slice of where norms does not equal 0 before dividing:

norms = np.where(norms != 0 ) 
angles = dots/norms