How to make nan in an expression?

[louisabraham]

I would like to produce a 'nan' expression without any warning. 'log(-1)' works for example but produces a warning.

see https://github.com/sympy/sympy/pull/12984#discussion_r127817271

[robbmcleod]

As per the warnings docs,

https://docs.python.org/3/library/warnings.html#temporarily-suppressing-warnings

import numpy as np
import numexpr as ne
import warnings

with warnings.catch_warnings():
    warnings.simplefilter("ignore")
    foo = ne.evaluate('log(-1)')
print(foo)

[louisabraham]

Of course, but sometimes you want to see the relevant warnings. If you have a log error in the same expression, you will miss it and believe you got NaN for a legitimate reason.

[robbmcleod]

Warnings are not exceptions. This context manager will not suppress exceptions, only a try/except block would do that.

[louisabraham]

I think I wasn't very clear and don't understand your point.

Let me show an example:

ne.evaluate('where(log(a)<2, 1, log(-1))')

I would like a = 1000 not to show a warning, but a = -1 should show a warning. Do you think it is possible? Currently the best solution would be to find a free name and to give it the value float('nan').

[robbmcleod]

Edit: The warning will only ever pop once. So if you suppress it, it doesn't show up again.

For your code fragment, probably not in 2.6. Can you inject some variable 'NaN' variable into the locals dict of the calling frame? You would have the same problem with pi for example.

The 3.0 branch (https://github.com/pydata/numexpr/tree/numexpr-3.0) can do that already:

import numexpr3 as ne3
import numpy as np
a = 1000
ne3.evaluate('where(log(a)<2, 1, np.nan')

I'm hoping to find some time this long weekend to work on the benchmarking once more and then it should be ready for a beta. Integration into down-stream dependencies would likely take quite some time, because some of them are so enormous (like pandas).

[github-actions[bot]]

Message to comment on stale issues. If none provided, will not mark issues stale