Repository navigation
BUG: random: np.random.weibull(a=0) sometimes returns inf. #12371
Description
Over in #8056, there was a test failure that appears to be a numpy bug. The test that is failing is TestRandomDist.test_weibull_0, which is:
def test_weibull_0(self):
assert_equal(np.random.weibull(a=0), 0)
assert_raises(ValueError, np.random.weibull, a=-0.)
The assertion assert_equal(np.random.weibull(a=0), 0) is failing, because sometimes np.random.weibull(a=0) returns inf:
In [36]: np.__version__
Out[36]: '1.15.4'
In [37]: np.random.seed(123)
In [38]: np.random.weibull(a=0, size=10)
Out[38]: array([inf, 0., 0., 0., inf, 0., inf, inf, 0., 0.])
Those inf values should not be there. I get the same result with the current master branch, '1.16.0.dev0+e34f5bb'.
Activity
WarrenWeckesser commentedon Nov 12, 2018
Here's the code to generate Weibull variates:
If
ais 0, then1/aisinf. If0 < r < 1, thenr**infis 0, but ifr == 1thenr**infis 1, and ifr > 1thenr**infisinf. Those are not correct values for this degenerate case of the Weibull distribution. The only value that should be returned bynp.random.weibll(a=0)is 0.This could be fixed by handling
a = 0explicitly, e.g.(The even edgier edge case of
a = -0is handled before this C code is called.)BUG: random: Fix handling of a=0 for numpy.random.weibull.
WarrenWeckesser commentedon Nov 12, 2018
Pull request to fix this: #12372
BUG: random: Fix handling of a=0 for numpy.random.weibull.