scikit-learn: Test failure in i686 with version 1.3.1

Describe the bug

During the build of scikit-learn for Fedora Linux, I’m obtaining an error runing the tests in i686. The test that fails is:

sklearn/tree/tests/test_export.py::test_graphviz_toy

Steps/Code to Reproduce

In a i686 machine

pytest sklearn/tree/tests/test_export.py

Expected Results

Test passes

Actual Results

sklearn/tree/tests/test_export.py::test_graphviz_toy FAILED              [ 93%]
=================================== FAILURES ===================================
______________________________ test_graphviz_toy _______________________________
    def test_graphviz_toy():
        # Check correctness of export_graphviz
        clf = DecisionTreeClassifier(
            max_depth=3, min_samples_split=2, criterion="gini", random_state=2
        )
        clf.fit(X, y)
    
        # Test export code
        contents1 = export_graphviz(clf, out_file=None)
        contents2 = (
            "digraph Tree {\n"
            'node [shape=box, fontname="helvetica"] ;\n'
            'edge [fontname="helvetica"] ;\n'
            '0 [label="x[0] <= 0.0\\ngini = 0.5\\nsamples = 6\\n'
            'value = [3, 3]"] ;\n'
            '1 [label="gini = 0.0\\nsamples = 3\\nvalue = [3, 0]"] ;\n'
            "0 -> 1 [labeldistance=2.5, labelangle=45, "
            'headlabel="True"] ;\n'
            '2 [label="gini = 0.0\\nsamples = 3\\nvalue = [0, 3]"] ;\n'
            "0 -> 2 [labeldistance=2.5, labelangle=-45, "
            'headlabel="False"] ;\n'
            "}"
        )
    
        assert contents1 == contents2
    
        # Test plot_options
        contents1 = export_graphviz(
            clf,
            filled=True,
            impurity=False,
            proportion=True,
            special_characters=True,
            rounded=True,
            out_file=None,
            fontname="sans",
        )
        contents2 = (
            "digraph Tree {\n"
            'node [shape=box, style="filled, rounded", color="black", '
            'fontname="sans"] ;\n'
            'edge [fontname="sans"] ;\n'
            "0 [label=<x<SUB>0</SUB> &le; 0.0<br/>samples = 100.0%<br/>"
            'value = [0.5, 0.5]>, fillcolor="#ffffff"] ;\n'
            "1 [label=<samples = 50.0%<br/>value = [1.0, 0.0]>, "
            'fillcolor="#e58139"] ;\n'
            "0 -> 1 [labeldistance=2.5, labelangle=45, "
            'headlabel="True"] ;\n'
            "2 [label=<samples = 50.0%<br/>value = [0.0, 1.0]>, "
            'fillcolor="#399de5"] ;\n'
            "0 -> 2 [labeldistance=2.5, labelangle=-45, "
            'headlabel="False"] ;\n'
            "}"
        )
    
        assert contents1 == contents2
    
        # Test max_depth
        contents1 = export_graphviz(clf, max_depth=0, class_names=True, out_file=None)
        contents2 = (
            "digraph Tree {\n"
            'node [shape=box, fontname="helvetica"] ;\n'
            'edge [fontname="helvetica"] ;\n'
            '0 [label="x[0] <= 0.0\\ngini = 0.5\\nsamples = 6\\n'
            'value = [3, 3]\\nclass = y[0]"] ;\n'
            '1 [label="(...)"] ;\n'
            "0 -> 1 ;\n"
            '2 [label="(...)"] ;\n'
            "0 -> 2 ;\n"
            "}"
        )
    
        assert contents1 == contents2
    
        # Test max_depth with plot_options
        contents1 = export_graphviz(
            clf, max_depth=0, filled=True, out_file=None, node_ids=True
        )
        contents2 = (
            "digraph Tree {\n"
            'node [shape=box, style="filled", color="black", '
            'fontname="helvetica"] ;\n'
            'edge [fontname="helvetica"] ;\n'
            '0 [label="node #0\\nx[0] <= 0.0\\ngini = 0.5\\n'
            'samples = 6\\nvalue = [3, 3]", fillcolor="#ffffff"] ;\n'
            '1 [label="(...)", fillcolor="#C0C0C0"] ;\n'
            "0 -> 1 ;\n"
            '2 [label="(...)", fillcolor="#C0C0C0"] ;\n'
            "0 -> 2 ;\n"
            "}"
        )
    
        assert contents1 == contents2
    
        # Test multi-output with weighted samples
        clf = DecisionTreeClassifier(
            max_depth=2, min_samples_split=2, criterion="gini", random_state=2
        )
        clf = clf.fit(X, y2, sample_weight=w)
    
>       contents1 = export_graphviz(clf, filled=True, impurity=False, out_file=None)
sklearn/tree/tests/test_export.py:131: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
sklearn/utils/_param_validation.py:211: in wrapper
    return func(*args, **kwargs)
sklearn/tree/_export.py:905: in export_graphviz
    exporter.export(decision_tree)
sklearn/tree/_export.py:465: in export
    self.recurse(decision_tree.tree_, 0, criterion=decision_tree.criterion)
sklearn/tree/_export.py:528: in recurse
    ', fillcolor="%s"' % self.get_fill_color(tree, node_id)
sklearn/tree/_export.py:285: in get_fill_color
    return self.get_color(node_val)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
self = <sklearn.tree._export._DOTTreeExporter object at 0xc5080408>
value = -0.4691358024691358
    def get_color(self, value):
        # Find the appropriate color & intensity for a node
        if self.colors["bounds"] is None:
            # Classification tree
            color = list(self.colors["rgb"][np.argmax(value)])
            sorted_values = sorted(value, reverse=True)
            if len(sorted_values) == 1:
                alpha = 0.0
            else:
                alpha = (sorted_values[0] - sorted_values[1]) / (1 - sorted_values[1])
        else:
            # Regression tree or multi-output
            color = list(self.colors["rgb"][0])
            alpha = (value - self.colors["bounds"][0]) / (
                self.colors["bounds"][1] - self.colors["bounds"][0]
            )
        # compute the color as alpha against white
>       color = [int(round(alpha * c + (1 - alpha) * 255, 0)) for c in color]
E       ValueError: cannot convert float NaN to integer

Versions

Version is 1.3.1

About this issue

  • Original URL
  • State: closed
  • Created 3 years ago
  • Comments: 16 (16 by maintainers)

Commits related to this issue

Most upvoted comments

I opened an issue in Numpy: https://github.com/numpy/numpy/issues/26775. Let’s see what they say!

Great thanks a lot @mr-c for the reproducer inside the Docker image, I confirm I can reproduce indeed.

About the patch, rereading https://github.com/scikit-learn/scikit-learn/issues/27506#issuecomment-2154762231 more carefully I think we could even do something simpler by multiplying by -1. instead of using array negation (or the min(-arr) == -max(arr) trick I originally thought of). Go ahead if you want to open a a PR with something like the following otherwise I’ll try to pick it up in the not too far future:

from .utils.fixes import _IS_32BIT

# Avoid what looks like an issue with SIMD on non memory aligned arrays on 32bit in numpy >= 1.26,
# see https://github.com/scikit-learn/scikit-learn/issues/27506 for more details
minus_impurity = -1. * tree_.impurity if _IS32BIT else -tree_.impurity 
self.colors["bounds"] = (np.min(minus_impurity), np.max(minus_impurity))

I’ll also try to open an issue in Numpy, because they will probably more insights than me/us on this kind of tricky topics.

Thanks a lot for the details!

So I would say this bug is a bit of a edge case, since this happens on a 32bit machine, which represent a small portion of our users (for example for Pillow when they had 32bit wheels on Linux, that was less than 0.1% of the Linux downloads, see https://discuss.python.org/t/dropping-32-bit-packages/5476/20) and also happens in some scikit-learn visualization/debugging code.

The simplest fix would probably be to avoid using array negation, using the fact that np.min(-array) == -np.max(array) i.e. something like this:

diff --git a/sklearn/tree/_export.py b/sklearn/tree/_export.py
index dd3c655173..43909a17ee 100644
--- a/sklearn/tree/_export.py
+++ b/sklearn/tree/_export.py
@@ -266,7 +266,7 @@ class _BaseTreeExporter:
             self.colors["rgb"] = _color_brew(tree.n_classes[0])
             if tree.n_outputs != 1:
                 # Find max and min impurities for multi-output
-                self.colors["bounds"] = (np.min(-tree.impurity), np.max(-tree.impurity))
+                self.colors["bounds"] = (-np.max(tree.impurity), -np.min(tree.impurity))
             elif tree.n_classes[0] == 1 and len(np.unique(tree.value)) != 1:
                 # Find max and min values in leaf nodes for regression
                 self.colors["bounds"] = (np.min(tree.value), np.max(tree.value))

It could well be the case that a similar issue happens in other places of scikit-learn.

For further reference, could you post the output of

python -c 'import numpy; print(numpy.show_config())'

and

python -c 'import sklearn; sklearn.show_versions()'

I don’t know too much about the SIMD intricacies, but in an ideal world, it would be nice to understand if scikit-learn is to blame for not creating an array on a 64bit boundary (I am not sure at all what the boundary should be by the way) or numpy should be able to deal with unaligned arrays better …

I don’t know either, whether there is some hope to reproduce this kind of issues inside a 32bit Docker image on a 64bit host …