如何捕获像异常一样的numpy警告（不仅用于测试）？

I have to make a Lagrange polynomial in Python for a project I’m doing. I’m doing a barycentric style one to avoid using an explicit for-loop as opposed to a Newton’s divided difference style one. The problem I have is that I need to catch a division by zero, but Python (or maybe numpy) just makes it a warning instead of a normal exception.

So, what I need to know how to do is to catch this warning as if it were an exception. The related questions to this I found on this site were answered not in the way I needed. Here’s my code:

import numpy as np
import matplotlib.pyplot as plt
import warnings

class Lagrange:
    def __init__(self, xPts, yPts):
        self.xPts = np.array(xPts)
        self.yPts = np.array(yPts)
        self.degree = len(xPts)-1 
        self.weights = np.array([np.product([x_j - x_i for x_j in xPts if x_j != x_i]) for x_i in xPts])

    def __call__(self, x):
        warnings.filterwarnings("error")
        try:
            bigNumerator = np.product(x - self.xPts)
            numerators = np.array([bigNumerator/(x - x_j) for x_j in self.xPts])
            return sum(numerators/self.weights*self.yPts) 
        except Exception, e: # Catch division by 0. Only possible in 'numerators' array
            return yPts[np.where(xPts == x)[0][0]]

L = Lagrange([-1,0,1],[1,0,1]) # Creates quadratic poly L(x) = x^2

L(1) # This should catch an error, then return 1. 

When this code is executed, the output I get is:

Warning: divide by zero encountered in int_scalars

That’s the warning I want to catch. It should occur inside the list comprehension.

It seems that your configuration is using the print option for numpy.seterr:

>>> import numpy as np
>>> np.array([1])/0   #'warn' mode
__main__:1: RuntimeWarning: divide by zero encountered in divide
array([0])
>>> np.seterr(all='print')
{'over': 'warn', 'divide': 'warn', 'invalid': 'warn', 'under': 'ignore'}
>>> np.array([1])/0   #'print' mode
Warning: divide by zero encountered in divide
array([0])

This means that the warning you see is not a real warning, but it’s just some characters printed to stdout(see the documentation for seterr). If you want to catch it you can:

1. Use numpy.seterr(all='raise') which will directly raise the exception. This however changes the behaviour of all the operations, so it’s a pretty big change in behaviour.
2. Use numpy.seterr(all='warn'), which will transform the printed warning in a real warning and you’ll be able to use the above solution to localize this change in behaviour.

Once you actually have a warning, you can use the warnings module to control how the warnings should be treated:

>>> import warnings
>>> 
>>> warnings.filterwarnings('error')
>>> 
>>> try:
...     warnings.warn(Warning())
... except Warning:
...     print 'Warning was raised as an exception!'
... 
Warning was raised as an exception!

Read carefully the documentation for since it allows you to filter only the warning you want and has other options. I’d also consider looking at which is a context manager which automatically resets the original filterwarnings function:

>>> import warnings
>>> with warnings.catch_warnings():
...     warnings.filterwarnings('error')
...     try:
...         warnings.warn(Warning())
...     except Warning: print 'Raised!'
... 
Raised!
>>> try:
...     warnings.warn(Warning())
... except Warning: print 'Not raised!'
... 
__main__:2: Warning: 

To add a little to @Bakuriu’s answer:

If you already know where the warning is likely to occur then it’s often cleaner to use the numpy.errstate context manager, rather than numpy.seterr which treats all subsequent warnings of the same type the same regardless of where they occur within your code:

import numpy as np

a = np.r_[1.]
with np.errstate(divide='raise'):
    try:
        a / 0   # this gets caught and handled as an exception
    except FloatingPointError:
        print('oh no!')
a / 0           # this prints a RuntimeWarning as usual

Edit:

In my original example I had a = np.r_[0], but apparently there was a change in numpy’s behaviour such that division-by-zero is handled differently in cases where the numerator is all-zeros. For example, in numpy 1.16.4:

all_zeros = np.array([0., 0.])
not_all_zeros = np.array([1., 0.])

with np.errstate(divide='raise'):
    not_all_zeros / 0.  # Raises FloatingPointError

with np.errstate(divide='raise'):
    all_zeros / 0.  # No exception raised

with np.errstate(invalid='raise'):
    all_zeros / 0.  # Raises FloatingPointError

The corresponding warning messages are also different: 1. / 0. is logged as RuntimeWarning: divide by zero encountered in true_divide, whereas 0. / 0. is logged as RuntimeWarning: invalid value encountered in true_divide. I’m not sure why exactly this change was made, but I suspect it has to do with the fact that the result of 0. / 0. is not representable as a number (numpy returns a NaN in this case) whereas 1. / 0. and -1. / 0. return +Inf and -Inf respectively, per the IEE 754 standard.

If you want to catch both types of error you can always pass np.errstate(divide='raise', invalid='raise'), or all='raise' if you want to raise an exception on any kind of floating point error.

To elaborate on @Bakuriu’s answer above, I’ve found that this enables me to catch a runtime warning in a similar fashion to how I would catch an error warning, printing out the warning nicely:

import warnings

with warnings.catch_warnings():
    warnings.filterwarnings('error')
    try:
        answer = 1 / 0
    except Warning as e:
        print('error found:', e)

You will probably be able to play around with placing of the warnings.catch_warnings() placement depending on how big of an umbrella you want to cast with catching errors this way.

Remove warnings.filterwarnings and add:

numpy.seterr(all='raise')