Numpy FloatingPointError: Invalid value obtained in subtraction - not reproducible?

Question

I meet FloatingPointError: invalid value encountered in subtract in a piece of test code. The exception was thrown without any changes to the code itself, so I had trouble understanding it.

My question is: what causes the invalid value encountered in subtract exception? Why should this behave differently on different python + numpy installations?

DETAILS:

This MWE does not raise a FloatingPointError :

 >>> import numpy as np >>> np.__version__ '1.6.1' >>> x = np.arange(5,dtype='float64') >>> y = np.ones(5,dtype='float64') >>> x[2]=np.nan >>> xy # array([ -1., 0., nan, 2., 3.]) 

However, deep inside the piece of code, I subtract two np.float64 ndarray and get a floating point exception. The arrays that throw the exception contain some pretty huge and tiny numbers (like 1e307 and 1e-307) and some nan s, but I didn't make any combinations of these numbers as a result of testing the exceptions myself.

Much more disturbing, I have a large grid of Jenkins tests that work with the same code with many versions of numpy, matplotlib, python, and scipy, and NONE of them throws this exception. I got lost at this moment - I do not know if there is an error, or if there is, how to track it.

If you are curious curious, this code is pyspeckit and the test does not work on line 20 of test_hr2421.py .

EDIT: Follow-up - I think this little snippet: np.seterr(invalid='raise') is called in the module that I imported, in particular pymc, and the transfer request has since prevented this change.

Answer 1

Numpy has custom behavior regarding how errors are handled. By default, some errors are ignored, others cause a warning. For each category, you can change this behavior. Someone had to install it to increase errors without changing it.

You can suppress this exception by calling numpy.seterr(invalid='warn') , or alternatively invalid='ignore' . See the numpy.seterr documentation for a complete list of possible errors.

You can also use the context manager:

 In [12]: x = np.arange(-5, 5,dtype='float64') In [13]: with np.errstate(divide="raise"): print(1/x) ....: --------------------------------------------------------------------------- FloatingPointError Traceback (most recent call last) <ipython-input-13-881589fdcb7a> in <module>() 1 with np.errstate(divide="raise"): ----> 2 print(1/x) 3 FloatingPointError: divide by zero encountered in true_divide In [14]: with np.errstate(divide="warn"): print(1/x) ....: /home/users/gholl/venv/stable-3.5/bin/ipython3:2: RuntimeWarning: divide by zero encountered in true_divide [-0.2 -0.25 -0.33333333 -0.5 -1. inf 1. 0.5 0.33333333 0.25 ] In [15]: with np.errstate(divide="ignore"): print(1/x) ....: [-0.2 -0.25 -0.33333333 -0.5 -1. inf 1. 0.5 0.33333333 0.25 ] 

I tend to wrap all my code inside a block with np.errstate(all="raise") , and then use the context manager, ignoring the specific condition, if I'm sure that the problem does not hide the error - usually it is.