ATLAS version and failures inscipy.test()
Pearu, I switched back to ATLAS3.4.1 http://www.cs.utk.edu/~rwhaley/ATLAS/errata3.4.1.html#rLLt Now, scipy.test() gives ====================================================================== FAIL: check_normal (scipy.stats.morestats.test_morestats.test_anderson) ---------------------------------------------------------------------- Traceback (most recent call last): File "/usr/local/lib/python2.3/site-packages/scipy/stats/tests/ test_morestats.py", line 49, in check_normal assert(scipy.all(A < crit[-2:])) AssertionError ---------------------------------------------------------------------- Ran 1028 tests in 2.348s FAILED (failures=1) Nils
On Wed, 26 Nov 2003, Nils Wagner wrote:
Pearu,
I switched back to ATLAS3.4.1 http://www.cs.utk.edu/~rwhaley/ATLAS/errata3.4.1.html#rLLt
Yes, I found that too. I'll fix it in scipy CVS soon. Thanks.
Now, scipy.test() gives
====================================================================== FAIL: check_normal (scipy.stats.morestats.test_morestats.test_anderson) ---------------------------------------------------------------------- Traceback (most recent call last): File "/usr/local/lib/python2.3/site-packages/scipy/stats/tests/ test_morestats.py", line 49, in check_normal assert(scipy.all(A < crit[-2:])) AssertionError
When you run scipy.test multiple times, this test may succeed. There is always a possibility that some stats tests fail occasionally. Pearu
On Wed, 26 Nov 2003, Pearu Peterson wrote:
I switched back to ATLAS3.4.1 http://www.cs.utk.edu/~rwhaley/ATLAS/errata3.4.1.html#rLLt
Yes, I found that too. I'll fix it in scipy CVS soon.
It turns out that the failures in scipy tests that occur with Atlas versions higher than 3.4.1 are related to Atlas bugs. I have contacted Clint regarding this and he has opened the corresponding Atlas bug issue: https://sourceforge.net/tracker/index.php?func=detail&aid=850346&group_id=23725&atid=379482 Until potrf routines get gets fixed in Atlas, scipy.linalg uses Fortran version of zpotrf function (which actually should call Atlas routine clapack_zpotrf, so there should be no remarkable performance hit because of that). And it is OK to use any version of ATLAS with scipy again. Regards, Pearu
participants (2)
-
Nils Wagner -
Pearu Peterson