Search code examples
pythonnumpyzerodivide

Numpy in what situation will return "Divide by zero error"?

I have an expression as below, where as data is a matrix and theta is a vector, python constantly raise divide by 0 error. But the only denominator is (1+e^x), every element in this matrix cannot be 0 anyway

np.sum(-data[:,-1]*np.log(1.0/(1.0+np.exp(-np.dot(data[:,:-1],theta.transpose())))))

Is there any occasion I've ignored that will generate this error?

Weird thing is, the program runs well and results is right even with this error showed every time.

Solution

  • From the error message given in the comments to the question:

    py:21: RuntimeWarning: divide by zero encountered in log return (np.sum(-data[:,-1]*np.log(1.0/(1.0+np.exp(-np.dot(data[:,:-1],theta.transpose()‌​))))-(1-data[:,-1])*np.log(1-1.0/(1.0+np.exp(-np.dot(data[:,:-1],theta.transpose(‌​))))))+np.sum(theta[1:]**2)*lamda/2.0)*1.0/data.shape[0]

    it appears that you are computing np.log(0). E.g.

    >>> import numpy as np
>>> np.log(0)
__main__:1: RuntimeWarning: divide by zero encountered in log
-inf
>>>

    And that is likely happening because np.exp(...) is overflowing, or 1.0 + np.exp(...) is returning 1.0 (because np.exp(...) is smaller than the "machine epsilon") and you are computing np.log(1.0 - 1.0/(1.0 + np.exp(...))).