Lambdified Piecewise produces "RuntimeWarning: invalid value encountered in log"

[GeorgySk]

Simple example:

import sympy as sym

x = sym.Symbol('x')
f = sym.Piecewise((sym.log(-x), x < 0),
                  (sym.log(x), True))
lf = sym.lambdify(x, f)
print(lf(10))

gives:

2.302585092994046
/home/georgy/miniconda3/lib/python3.6/site-packages/numpy/__init__.py:1: RuntimeWarning: invalid value encountered in log
  """

Value is calculated correctly, then why is there a warning?

[moorepants]

Seems like it is evaluating both log(x) and log(-x). But the underlying printer doesn't seem to do this:

In [6]: from sympy.printing import lambdarepr

In [9]: lambdarepr.lambdarepr(f)
Out[9]: '((log(-x)) if (x < 0) else (((log(x)) if (True) else None)))'

In [12]: x = 10

In [13]: ((np.log(-x)) if (x < 0) else (((np.log(x)) if (True) else None)))
Out[13]: 2.3025850929940459

[asmeurer]

You want the NumPyPrinter.

>>> from sympy.printing.pycode import NumPyPrinter
>>> NumPyPrinter().doprint(f)
'numpy.select([numpy.less(x, 0),True], [numpy.log(-x),numpy.log(x)], default=numpy.nan)'

It has to use np.select so that the if-then-else gets vectorized, meaning the "incorrect" branch is always evaluated. This was also the problem with #11306. Is there a better NumPy function we can use with lambdify that doesn't evaluate unused branches, but still works on arrays? Such a function would be more efficient too.

Regarding the issue, the warning can be ignored. The invalid value isn't actually used, assuming your piecewise is setup correctly.