8000 FIX np.divide undefined behaviour with where in gaussian processes by lesteve · Pull Request #24245 · scikit-learn/scikit-learn · GitHub
[go: up one dir, main page]
Skip to content

FIX np.divide undefined behaviour with where in gaussian processes#24245

Merged
jeremiedbb merged 2 commits intoscikit-learn:mainfrom
lesteve:np-divide-undefined
Sep 5, 2022
Merged

FIX np.divide undefined behaviour with where in gaussian processes#24245
jeremiedbb merged 2 commits intoscikit-learn:mainfrom
lesteve:np-divide-undefined

Conversation

@lesteve
Copy link
Member
@lesteve lesteve commented Aug 24, 2022 •
edited
Loading

Triggered by investigating #24221.

We are using code like:

result = np.divide(numerator, denominator, where=denominator!=0)

result[denominator == 0] values are undefined when using this. For CPython it seems like the np.empty allocated for the return value reuses a temporary array created in a previous line that computes (X[:, np.newaxis, :] - X[np.newaxis, :, :]) ** 2 so that result[denominator == 0] only contains 0. For PyPy this is not the case and at one point we get a 4. rather than a 0. (don't ask me where the 4. comes from ...).

A snippet to reproduce a similar behaviour:

import numpy as np

numerator = np.array([1., 1., 1.])
denominator = np.array([0., 1., 1.])
where = denominator != 0

tmp = np.array([-999., -999., -999.])
print(f"tmp address: {tmp.__array_interface__['data'][0]}")
del tmp
divide_result = np.divide(numerator, denominator, where=where)
print(f"div address: {divide_result.__array_interface__['data'][0]}")
print(f"{divide_result}")

Output with CPython (address is the same so the -999. of the tmp array is reused):

tmp address: 94515367946736
div address: 94515367946736
[-999.    1.    1.]

Output with pypy (address is not the same first value is random):

tmp address: 94369592900336
div address: 94369602251488
[4.66247785e-310 1.00000000e+000 1.00000000e+000]

@lesteve lesteve mentioned this pull request Aug 24, 2022
Closed
thomasjpfan
thomasjpfan approved these changes Aug 27, 2022
View reviewed changes
Copy link
Member
@thomasjpfan thomasjpfan left a comment •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the PR and the explanation!

Besides a minor comment on the whats new placement, LGTM

doc/whats_new/v1.2.rst
:mod:`sklearn.feature_selection`
................................

:mod:`sklearn.gaussian_process`
Copy link
Member
@thomasjpfan thomasjpfan Aug 27, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This section seems to be out of place in the whats new.

@thomasjpfan thomasjpfan added the Quick Review For PRs that are quick to review label Aug 27, 2022
jeremiedbb
jeremiedbb approved these changes Sep 5, 2022
View reviewed changes
Copy link
Member
@jeremiedbb jeremiedbb left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM, thanks @lesteve. We should almost always be using the out arg of divide.

@jeremiedbb jeremiedbb merged commit 056d5cf into scikit-learn:main Sep 5, 2022
@lesteve lesteve deleted the np-divide-undefined branch September 6, 2022 07:39
@lesteve lesteve mentioned this pull request Sep 7, 2022
Merged
glemaitre pushed a commit to glemaitre/scikit-learn that referenced this pull request Sep 12, 2022
@lesteve @glemaitre
FIX np.divide undefined behaviour with where in gaussian processes (s…
b3cc549 
…cikit-learn#24245)
@jeremiedbb jeremiedbb mentioned this pull request Feb 28, 2023
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@thomasjpfan thomasjpfan thomasjpfan approved these changes

@jeremiedbb jeremiedbb jeremiedbb approved these changes

Assignees

No one assigned

Labels

module:gaussian_process Quick Review For PRs that are quick to review

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@lesteve @thomasjpfan @jeremiedbb
0