Scikit-learn: [0.23.1] doctest GradientBoostingClassifier failes on arm(rhel) processors

Created on 1 Jul 2020 · 3Comments · Source: scikit-learn/scikit-learn

Describe the bug

On arm processors (AWS:gravition2, rhel), I get the following failure in version 0.23.1

1038     >>> from sklearn.model_selection import train_test_split
1039     >>> X, y = make_classification(random_state=0)
1040     >>> X_train, X_test, y_train, y_test = train_test_split(
1041     ...     X, y, random_state=0)
1042     >>> clf = GradientBoostingClassifier(random_state=0)
1043     >>> clf.fit(X_train, y_train)
1044     GradientBoostingClassifier(random_state=0)
1045     >>> clf.predict(X_test[:2])
1046     array([1, 0])
1047     >>> clf.score(X_test, y_test)
Expected:
    0.88
Got:
    0.84

Steps/Code to Reproduce

pytest -v sklearn/ensemble/_gb.py::sklearn.ensemble._gb.GradientBoostingClassifier

Expected Results

PASSED is thrown.

Actual Results

FAILED is thrown.

1047     >>> clf.score(X_test, y_test)
Expected:
    0.88
Got:
    0.84

Versions

System:
    python: 3.6.8 (default, Dec  5 2019, 16:02:25)  [GCC 8.3.1 20191121 (Red Hat 8.3.1-5)]
executable: /usr/bin/python3
   machine: Linux-4.18.0-193.1.2.el8_2.aarch64-aarch64-with-redhat-8.2-Ootpa

Python dependencies:
          pip: 20.1.1
   setuptools: 39.2.0
      sklearn: 0.23.1
        numpy: 1.14.3
        scipy: 1.0.0
       Cython: 0.29
       pandas: 1.0.5
   matplotlib: 3.2.1
       joblib: 0.14.0
threadpoolctl: 2.1.0

Built with OpenMP: True
Linux-4.18.0-193.1.2.el8_2.aarch64-aarch64-with-redhat-8.2-Ootpa
Traceback (most recent call last):
  File "<string>", line 3, in <module>
NameError: name 'Python' is not defined

Bug arm help wanted ensemble

Source

murata-yu

All 3 comments

Thanks for the report @murata-yu , I can reproduce in https://github.com/scikit-learn/scikit-learn/pull/17996

Not yet sure if it's an indication of an actual issue of it we should just increase the tolerance.

rth on 26 Jul 2020

👍1

From a user point of view, 4% of change in accuracy in the 0.8 range looks like more than a small numerical rounding discrepancy. It's worth investigating.

ogrisel on 27 Jul 2020

I can reproduce the problem locally by building scikit-learn in an arm64 miniforge environment in docker / qemu container as described in https://github.com/scikit-learn/scikit-learn/pull/17644#issuecomment-663857435.

ogrisel on 28 Jul 2020

👍1

Was this page helpful?

0 / 5 - 0 ratings