Different failures for scikit-learn (Was: Failures for sklearn)
Hi,
Am Tue, Feb 15, 2022 at 01:39:30PM +0100 schrieb Andreas Tille:
>
> I'll try relaxing tolerance limits a bit to solve this issue.
This worked for i386.
Unfortunately there are several failing architectures remaining[1]
(not finished building yet for all at the time of writing).
I checked some logs and noticed that armel and armhf are failing with
"Bus error". While I subscribe to the reasoning that this could be
caused by some coding bug which hides somewhere and is not detected on
other architectures I think we should draw some line here. Obviously we
do not have the power to maintain this code for all architectures
neither do I expect that there is any honest user of this code on the
said architectures. So I would simply exclude these from the list
of architectures. Please raise your hand if you disagree and volunteer
to work on these problems.
For arm64[2] its a single failing test:
=================================== FAILURES ===================================
________ [doctest] sklearn.ensemble._weight_boosting.AdaBoostRegressor _________
1004 Examples
1005 --------
1006 >>> from sklearn.ensemble import AdaBoostRegressor
1007 >>> from sklearn.datasets import make_regression
1008 >>> X, y = make_regression(n_features=4, n_informative=2,
1009 ... random_state=0, shuffle=False)
1010 >>> regr = AdaBoostRegressor(random_state=0, n_estimators=100)
1011 >>> regr.fit(X, y)
1012 AdaBoostRegressor(n_estimators=100, random_state=0)
1013 >>> regr.predict([[0, 0, 0, 0]])
Expected:
array([4.7972...])
Got:
array([5.74049295])
/<<PKGBUILDDIR>>/.pybuild/cpython3_3.9/build/sklearn/ensemble/_weight_boosting.py:1013: DocTestFailure
= 1 failed, 22012 passed, 404 skipped, 7 deselected, 59 xfailed, 40 xpassed, 2350 warnings in 2090.20s (0:34:50) =
Its the very same failure for ppc64el - I'll report this upstream.
This error occures also on s390x where we see also an additional one[3]
___________ [doctest] sklearn.feature_extraction._hash.FeatureHasher ___________
081 DictVectorizer : Vectorizes string-valued features using a hash table.
082 sklearn.preprocessing.OneHotEncoder : Handles nominal/categorical features.
083
084 Examples
085 --------
086 >>> from sklearn.feature_extraction import FeatureHasher
087 >>> h = FeatureHasher(n_features=10)
088 >>> D = [{'dog': 1, 'cat':2, 'elephant':4},{'dog': 2, 'run': 5}]
089 >>> f = h.transform(D)
090 >>> f.toarray()
Expected:
array([[ 0., 0., -4., -1., 0., 0., 0., 0., 0., 2.],
[ 0., 0., 0., -2., -5., 0., 0., 0., 0., 0.]])
Got:
array([[ 0., 0., 0., -1., 0., 0., 0., 4., 0., 2.],
[ 0., 0., 0., -2., -5., 0., 0., 0., 0., 0.]])
/<<PKGBUILDDIR>>/.pybuild/cpython3_3.9/build/sklearn/feature_extraction/_hash.py:90: DocTestFailure
ppc64 behaves like s390x and risc64 has also only the first error.
So I'll contact upstream with two issues and will see what might
happen next.
Kind regards
Andreas.
[1] https://buildd.debian.org/status/package.php?p=scikit-learn
[2] https://buildd.debian.org/status/fetch.php?pkg=scikit-learn&arch=arm64&ver=1.0.2-1&stamp=1644956264&raw=0
[3] https://buildd.debian.org/status/fetch.php?pkg=scikit-learn&arch=s390x&ver=1.0.2-1&stamp=1644952396&raw=0
--
http://fam-tille.de
Reply to: