FIX Work around likely SIMD issue in tree export on 32bit OS #29327

lesteve · 2024-06-21T04:50:12Z

I have put @mr-c,@gspr and @sergiopasra as co-authors, thanks for your inputs on this tricky issue 🙏!

I tested the fix in the Docker image provided by @mr-c in #27506 (comment) and it works

Note for Distro package managers the patch may not work directly because _IS_32BIT was moved from sklearn.utils to sklearn.utils.fixes . Honestly I am pretty sure that this code is not performance critical and that just using -1.0 * tree.impurity even on 64bit OS is a completely fine patch. After all this code was not SIMDed before using numpy 1.26 and nobody ever complained that it was slow ...

The main reason, I used _IS_32BIT is that it makes it more explicit and I hope that one day it can be cleaned up 🤞

Co-authored-by: Michael R. Crusoe <1330696+mr-c@users.noreply.github.com> Co-authored-by: SGard Spreemann <gspr@nonempty.org> Co-authored-by: Sergio Pascual <sergiopr@fis.ucm.es>

github-actions · 2024-06-21T04:51:31Z

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: d1c65fd. Link to the linter CI: here}

lesteve · 2024-06-21T08:25:08Z

The CI is 🚨 but I will merge main when #29328 is auto-merged and that should be all 💚

…nto export-tree-32bit

lesteve · 2024-06-21T10:04:53Z

@mr-c @gspr, just curious, would it make the Linux distribution packaging easier (my understanding is that you are involved in it somehow) if this fix was included in the soonish to be released bug-fix release scikit-learn 1.5.1? Maybe it would not make a big difference anyway, because you are going to package an older scikit-learn release like 1.3 or 1.4?

mr-c · 2024-06-21T10:16:10Z

@mr-c @gspr, just curious, would it make the Linux distribution packaging easier (my understanding is that you are involved in it somehow) if this fix was included in the soonish to be released bug-fix release scikit-learn 1.5.1? Maybe it would not make a big difference anyway, because you are going to package an older scikit-learn release like 1.3 or 1.4?

Debian is already carrying a version of this patch on top of version 1.4.2 in our development distribution and that will soon migrate to our testing distribution

https://tracker.debian.org/pkg/scikit-learn

We are carrying other patches as well, perhaps some of them could also be upstreamed

https://salsa.debian.org/science-team/scikit-learn/-/tree/master/debian/patches?ref_type=heads

(I'm not one of the primary maintainers of scikit-learn in Debian, but I'm on the team and recently merged some fixes)

We have not yet packaged scikit-learn 1.5.x (possibly due to a bug in how we check for new versions, I will look into that now)

So, no rush needed for Debian. Thank you for asking!

gspr · 2024-06-21T10:17:26Z

@mr-c @gspr, just curious, would it make the Linux distribution packaging easier (my understanding is that you are involved in it somehow) if this fix was included in the soonish to be released bug-fix release scikit-learn 1.5.1? Maybe it would not make a big difference anyway, because you are going to package an older scikit-learn release like 1.3 or 1.4?

To add to what @mr-c said: Thank you for asking! This kind of thought and consideration given to distributions really gives me confidence in the scikit-learn developers ❤️

mr-c · 2024-06-21T10:28:55Z

possibly due to a bug in how we check for new versions, I will look into that now)

Ah, the source distribution now has an underscore instead of a hyphen: https://pypi.org/project/scikit-learn/1.5.0/#files vs https://pypi.org/project/scikit-learn/1.4.2/#files

I have fixed our regex to account for that

lesteve · 2024-06-21T11:50:56Z

Ah, the source distribution now has an underscore instead of a hyphen: pypi.org/project/scikit-learn/1.5.0/#files vs pypi.org/project/scikit-learn/1.4.2/#files

Oh well, probably a side-effect of moving to Meson (and away from setuptools). This was noticed in #28757 (comment) actually, but would I have believed that someone out there was relying on this? Certainly not 😅 ...

jeremiedbb

Needs a changelog entry. I would put it for 1.5.1

jeremiedbb · 2024-06-21T13:16:45Z

sklearn/tree/_export.py

+                minus_impurity = -1.0 * tree.impurity if _IS_32BIT else -tree.impurity
+                self.colors["bounds"] = (np.min(minus_impurity), np.max(minus_impurity))


would the following work ? It avoids 1 unnecessary array multiplication and it avoids to specialize the behavior for 32bit OS.

Suggested change

minus_impurity = -1.0 * tree.impurity if _IS_32BIT else -tree.impurity

self.colors["bounds"] = (np.min(minus_impurity), np.max(minus_impurity))

# equivalent to (np.min(-tree.impurity), np.max(-tree.impurity))

self.colors["bounds"] = (-np.max(tree.impurity), -np.min(tree.impurity))

That was my original thought in #27506 (comment) 😉.

I find it slightly too clever and I kind of doubt this matters that much in practice since this is not in a performance-critical part of the code (visualization/debugging)

Having said that, the min(-arr) == -max(arr) trick is now the official Debian patch for scikit-learn 1.4 apparently: https://salsa.debian.org/science-team/scikit-learn/-/blob/master/debian/patches/i386-impurity-nan.patch?ref_type=heads 😅

That was my original thought [...] I find it too clever

😄

It's a matter of taste indeed. I find it more intuitive.
One argument: the workaround should be removed when the bug is fixed upstream, while the suggestion can be let as is for ever. I updated the suggestion to add a comment to make it less surprising

@jeremiedbb OK you won me over by talking to my ex-physicist side 😉, I have added a comment to kind of indicate there is a reason why we write it like this.

I have also added an entry in the 1.5.1 changelog.

thomasjpfan

LGTM

jeremiedbb

LGTM

…learn#29327) Co-authored-by: Michael R. Crusoe <1330696+mr-c@users.noreply.github.com> Co-authored-by: SGard Spreemann <gspr@nonempty.org> Co-authored-by: Sergio Pascual <sergiopr@fis.ucm.es>

Co-authored-by: Michael R. Crusoe <1330696+mr-c@users.noreply.github.com> Co-authored-by: SGard Spreemann <gspr@nonempty.org> Co-authored-by: Sergio Pascual <sergiopr@fis.ucm.es>

lesteve and others added 2 commits June 21, 2024 06:31

FIX Work around likely SIMD issue on 32bit OS

22b5003

Co-authored-by: Michael R. Crusoe <1330696+mr-c@users.noreply.github.com> Co-authored-by: SGard Spreemann <gspr@nonempty.org> Co-authored-by: Sergio Pascual <sergiopr@fis.ucm.es>

[azure parallel]

0d27fba

github-actions bot added the module:tree label Jun 21, 2024

lesteve changed the title ~~Export tree 32bit~~ FIX Work-around likely SIMD issue on 32bit OS Jun 21, 2024

lesteve changed the title ~~FIX Work-around likely SIMD issue on 32bit OS~~ FIX Work around likely SIMD issue on 32bit OS Jun 21, 2024

lesteve changed the title ~~FIX Work around likely SIMD issue on 32bit OS~~ FIX Work around likely SIMD issue in tree export on 32bit OS Jun 21, 2024

lesteve added 2 commits June 21, 2024 11:06

Merge branch 'main' of https://github.com/scikit-learn/scikit-learn i…

872eec3

…nto export-tree-32bit

[azure parallel]

828d895

lesteve added the Quick Review For PRs that are quick to review label Jun 21, 2024

jeremiedbb reviewed Jun 21, 2024

View reviewed changes

lesteve added 2 commits June 21, 2024 17:05

go back to the min(-arr) = -max(arr)

14f69ee

add changelog

d1c65fd

thomasjpfan approved these changes Jun 21, 2024

View reviewed changes

jeremiedbb approved these changes Jun 21, 2024

View reviewed changes

jeremiedbb merged commit a67ebbe into scikit-learn:main Jun 21, 2024
30 checks passed

jeremiedbb mentioned this pull request Jul 2, 2024

Release 1.5.1 #29382

Merged

11 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FIX Work around likely SIMD issue in tree export on 32bit OS #29327

FIX Work around likely SIMD issue in tree export on 32bit OS #29327

lesteve commented Jun 21, 2024 •

edited

Loading

github-actions bot commented Jun 21, 2024 •

edited

Loading

lesteve commented Jun 21, 2024 •

edited

Loading

lesteve commented Jun 21, 2024

mr-c commented Jun 21, 2024

gspr commented Jun 21, 2024

mr-c commented Jun 21, 2024

lesteve commented Jun 21, 2024 •

edited

Loading

jeremiedbb left a comment

jeremiedbb Jun 21, 2024 •

edited

Loading

lesteve Jun 21, 2024 •

edited

Loading

jeremiedbb Jun 21, 2024

jeremiedbb Jun 21, 2024 •

edited

Loading

lesteve Jun 21, 2024

thomasjpfan left a comment

jeremiedbb left a comment

		minus_impurity = -1.0 * tree.impurity if _IS_32BIT else -tree.impurity
		self.colors["bounds"] = (np.min(minus_impurity), np.max(minus_impurity))

FIX Work around likely SIMD issue in tree export on 32bit OS #29327

FIX Work around likely SIMD issue in tree export on 32bit OS #29327

Conversation

lesteve commented Jun 21, 2024 • edited Loading

github-actions bot commented Jun 21, 2024 • edited Loading

✔️ Linting Passed

lesteve commented Jun 21, 2024 • edited Loading

lesteve commented Jun 21, 2024

mr-c commented Jun 21, 2024

gspr commented Jun 21, 2024

mr-c commented Jun 21, 2024

lesteve commented Jun 21, 2024 • edited Loading

jeremiedbb left a comment

Choose a reason for hiding this comment

jeremiedbb Jun 21, 2024 • edited Loading

Choose a reason for hiding this comment

lesteve Jun 21, 2024 • edited Loading

Choose a reason for hiding this comment

jeremiedbb Jun 21, 2024

Choose a reason for hiding this comment

jeremiedbb Jun 21, 2024 • edited Loading

Choose a reason for hiding this comment

lesteve Jun 21, 2024

Choose a reason for hiding this comment

thomasjpfan left a comment

Choose a reason for hiding this comment

jeremiedbb left a comment

Choose a reason for hiding this comment

lesteve commented Jun 21, 2024 •

edited

Loading

github-actions bot commented Jun 21, 2024 •

edited

Loading

lesteve commented Jun 21, 2024 •

edited

Loading

lesteve commented Jun 21, 2024 •

edited

Loading

jeremiedbb Jun 21, 2024 •

edited

Loading

lesteve Jun 21, 2024 •

edited

Loading

jeremiedbb Jun 21, 2024 •

edited

Loading