On Sat, Mar 14, 2009 at 4:22 PM, Sturla Molden <stu...@molden.no> wrote: >> On Sat, Mar 14, 2009 at 1:37 PM, <josef.p...@gmail.com> wrote: > >> OK. One more question: how often do the tests fail? I want to include a >> note >> to repeat testing if the test fails. > > I don't like this. I think the prngs should use fixed seeds known to pass > the test. Depending on confidence intervals in the units tests is really, > really bad style. Tests should be deterministic. > > S.M. >
The hypergeometric tests are on the support of the distribution and should never fail. And the outcome is not random. The test of logser with N = 100000 also should be pretty exact and fail only with very low probability in the patched version. But again this is testet in scipy.stats. I think Sturlas idea to find a random seed that differentiates before and after will be better for numpy, and using only a small sample size e.g. N=1000, since it's pretty fast. But since I don't have an unpatched numpy version available right now, I cannot do this. Josef _______________________________________________ Numpy-discussion mailing list Numpy-discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
