Codestin Search App

jeremiedbb · 2023-07-05T09:35:07Z

knn.predict doesn't validate X right away because it call kneighbors that does validate X.
The issue is that ArgKMin.is_usable_for assumes that it's a numpy array. This PR simply adds a check there. Note that it is not usable for a pandas dataframe anyway because it's not c-contiguous.

github-actions · 2023-07-05T09:36:52Z

✔️ Linting Passed

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

_{Generated for commit: 61467a0. Link to the linter CI: here}

jeremiedbb · 2023-07-05T10:11:28Z

cc/ @jjerphan

jjerphan

LGTM. Thank you, @jeremiedbb.

jjerphan · 2023-07-05T11:55:09Z

@@ -96,7 +96,9 @@ def is_usable_for(cls, X, Y, metric) -> bool:
        """

        def is_numpy_c_ordered(X):


Side-note: if there are any other libraries that implement the buffer protocol in Python and that we could support, we might want to adapt this.

betatim · 2023-07-05T13:43:42Z


        def is_numpy_c_ordered(X):
-            return hasattr(X, "flags") and X.flags.c_contiguous
+            if not hasattr(X, "flags"):


I don't quite understand why this fixes it :( I always thought that something like bool(expr1 and expr2) will not evaluate expr2 if expr1 is False. From the bug report it looks like hasattr(X, "flags") will be False, but then why did it continue into evaluating X.flags.c_contiguous. And even more confusingly the exception is AttributeError: 'Flags' object has no attribute 'c_contiguous', which makes me think that hasattr(X, "flags") evaluates as True when you use a DF as input. But then there is a new test and that passes. So overall I am left confused :-/

Or is the thing that actually fixes it the getattr(X.flags, "c_contiguous", False)?

which makes me think that hasattr(X, "flags") evaluates as True when you use a DF as input.

Yep, a dataframe has a flags attribute, but this flags has no c_contiguous attribute :)

Or is the thing that actually fixes it the getattr(X.flags, "c_contiguous", False)?

indeed. Your understanding of the conditional evaluation of expr1 and expr2 is correct

So we could have kept a one-liner then:

return hasattr(X, "flags") and getattr(X.flags, "c_contiguous", False)

hi every one so i have same prblm here 'AttributeError: 'Flags' object has no attribute 'c_contiguous''
and i dont understand the solution u given before . any one can help ?

Co-authored-by: Julien Jerphanion <[email protected]>

jeremiedbb · 2023-07-05T15:39:51Z

the doc failure is probably due to #26773

thomasjpfan

Thanks for the PR! Overall looks good.

thomasjpfan · 2023-07-05T17:44:47Z

+                return False
+            return getattr(X.flags, "c_contiguous", False)


Nit: (Ask for forgiveness and not permission)

def is_numpy_c_ordered(X): try: return X.flags.c_contiguous except AttributeError: return False

I'm usually not a huge fan of try excepts because it hides stuff. With that it would not have crashed in the first place but it means that we would have never noticed that since we're not validating X at the beginning of predict, passing a dataframe will always take the legacy route.

jeremiedbb i am very new to this. I have the exact same issue with c_contiguous flag.

Can you please explain what steps I must follow to fix this? I am afraid I can't quite follow what I should do to fix this issue. Please help.

This bug will be fixed in scikit-learn 1.3.1 that we plan to release soon. In the mean time you can use version 1.2 which does not have this bug.

@jeremiedbb I upgraded to scikit-learn 1.4.0 but im still getting this issue c_contiguous flag when I run knn.predict. please help!

ogrisel · 2023-07-06T09:24:52Z

The issue is that ArgKMin.is_usable_for assumes that it's a numpy array. This PR simply adds a check there. Note that it is not usable for a pandas dataframe anyway because it's not c-contiguous.

But shouldn't we just validate the data in predict to make it a c-contiguous array instead?

ogrisel · 2023-07-06T09:26:38Z

I think it's not the role of the lower level ArgKMin API to have to deal with non-numpy inputs. I think it's the estimators role to make sure that it calls its Cython backend with the proper datastructure.

jeremiedbb · 2023-07-06T09:27:27Z

Yes, but X is validated in kneighbors right after. So in order to do that we need a bit of refactoring, probably through a private _kneighbors function

jeremiedbb · 2023-07-06T09:29:18Z

I think it's not the role of the lower level ArgKMin API to have to deal with non-numpy inputs. I think it's the estimators role to make sure that it calls its Cython backend with the proper datastructure.

I agree but I wanted a quick fix. I'm okay to make the refactoring I explained above as part of this PR if you think it's better

jjerphan · 2023-07-06T12:08:05Z

But shouldn't we just validate the data in predict to make it a c-contiguous array instead?

This might be costly to from Fortran-contiguous data to C-contiguous datasets, but we might want to consider a configuration option to specify the behavior.

FYI, this is one item of #25888.

ogrisel · 2023-07-06T16:18:26Z

This might be costly to from Fortran-contiguous data to C-contiguous datasets, but we might want to consider a configuration option to specify the behavior.

Indeed, this might increase memory usage by a lot for such users. But it should also make the code run significantly faster. Maybe we could extend the pairwise distance + reduction Cython classes to accept large F-aligned input, in which case the outer loop would load rowise slices of the input in parallel into C-contiguous chunks to perform the actual computation in parallel on rows of X_test.

Ok let's be pragmatic and merge this quick fix and see later what to do in a PR for 1.4.

jjerphan · 2023-07-06T17:48:54Z

Indeed, this might increase memory usage by a lot for such users. But it should also make the code run significantly faster.

Generally in pipeline we might have different steps which necessitates either C- or F-contiguous inputs to work (efficiently). Thus I think converting is a tradeoff whose choice which should be left to the user via the configuration for instance.

Maybe we could extend the pairwise distance + reduction Cython classes to accept large F-aligned input, in which case the outer loop would load rowise slices of the input in parallel into C-contiguous chunks to perform the actual computation in parallel on rows of X_test.

I thought of it in the past, and I do not think coming up with equivalent and efficient implementations for F-contiguous data is doable for PairwiseDistancesReductions since their memory access patterns operating on rows would not operate on continuous adresses in this case, making implementations of algorithms complex or even impossible to translate. Though, if only X_test is F-contiguous, then copying or converting it to C-contiguous vectors to use those implementations might be competitive.

00kira00 · 2023-07-08T18:36:51Z

hi every one so i have same prblm here 'AttributeError: 'Flags' object has no attribute 'c_contiguous''
and i dont understand the solution u given before . any one can help ?

jjerphan · 2023-07-09T14:10:29Z

@00kira00: the problem you describe is fixed by this PR.

betatim · 2023-07-10T07:26:41Z

Should we make an issue to discuss #26772 (comment) and what to do?

jjerphan · 2023-07-10T07:27:44Z

Should we make an issue to discuss #26772 (comment) and what to do?

I think it makes sense.

Co-authored-by: Julien Jerphanion <[email protected]> Co-authored-by: Olivier Grisel <[email protected]>

fix dataframe support knn predict

173e3d8

github-actions Bot added module:metrics module:neighbors labels Jul 5, 2023

add what's new entry

949ca1e

jeremiedbb mentioned this pull request Jul 5, 2023

AttributeError "Flags object has no attribute 'c_contiguous'" when using KNeighborsClassifier predict #26768

Closed

ref

fa77f00

jjerphan approved these changes Jul 5, 2023

View reviewed changes

jjerphan added Quick Review For PRs that are quick to review Waiting for Second Reviewer First reviewer is done, need a second one! labels Jul 5, 2023

betatim reviewed Jul 5, 2023

View reviewed changes

jeremiedbb and others added 2 commits July 5, 2023 16:42

Update doc/whats_new/v1.3.rst

98c1389

Co-authored-by: Julien Jerphanion <[email protected]>

Update v1.3.rst

0514516

thomasjpfan reviewed Jul 5, 2023

View reviewed changes

jeremiedbb added 2 commits July 6, 2023 10:52

one liner

c1d853d

Merge branch 'main' into fix-usable-for-dataframe

822607c

ogrisel approved these changes Jul 6, 2023

View reviewed changes

Merge branch 'main' into fix-usable-for-dataframe

61467a0

jjerphan removed the Quick Review For PRs that are quick to review label Jul 7, 2023

betatim merged commit f473d7e into scikit-learn:main Jul 10, 2023

jjerphan removed the Waiting for Second Reviewer First reviewer is done, need a second one! label Jul 10, 2023

punndcoder28 pushed a commit to punndcoder28/scikit-learn that referenced this pull request Jul 29, 2023

FIX Dataframe support in KNN.predict (scikit-learn#26772)

b1bbc12

Co-authored-by: Julien Jerphanion <[email protected]> Co-authored-by: Olivier Grisel <[email protected]>

martinfleis mentioned this pull request Aug 3, 2023

CI: properly test min and dev pysal/esda#258

Merged

trevorcampbell mentioned this pull request Aug 18, 2023

Various compilation fixes UBC-DSCI/introduction-to-datascience-python#206

Merged

glemaitre pushed a commit to glemaitre/scikit-learn that referenced this pull request Sep 18, 2023

FIX Dataframe support in KNN.predict (scikit-learn#26772)

cf6e6c3

Co-authored-by: Julien Jerphanion <[email protected]> Co-authored-by: Olivier Grisel <[email protected]>

jeremiedbb added a commit that referenced this pull request Sep 20, 2023

FIX Dataframe support in KNN.predict (#26772)

7f4c690

Co-authored-by: Julien Jerphanion <[email protected]> Co-authored-by: Olivier Grisel <[email protected]>

REDVM pushed a commit to REDVM/scikit-learn that referenced this pull request Nov 16, 2023

FIX Dataframe support in KNN.predict (scikit-learn#26772)

8b31d0f

Co-authored-by: Julien Jerphanion <[email protected]> Co-authored-by: Olivier Grisel <[email protected]>

jeremiedbb deleted the fix-usable-for-dataframe branch September 3, 2026 16:08

		@@ -96,7 +96,9 @@ def is_usable_for(cls, X, Y, metric) -> bool:
		"""

		def is_numpy_c_ordered(X):

Uh oh!

Uh oh!

Conversation

jeremiedbb commented Jul 5, 2023

Uh oh!

github-actions Bot commented Jul 5, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✔️ Linting Passed

Uh oh!

jeremiedbb commented Jul 5, 2023

Uh oh!

jjerphan left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jeremiedbb commented Jul 5, 2023

Uh oh!

thomasjpfan left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ogrisel commented Jul 6, 2023

Uh oh!

ogrisel commented Jul 6, 2023

Uh oh!

jeremiedbb commented Jul 6, 2023

Uh oh!

jeremiedbb commented Jul 6, 2023

Uh oh!

jjerphan commented Jul 6, 2023

Uh oh!

ogrisel commented Jul 6, 2023

Uh oh!

jjerphan commented Jul 6, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

00kira00 commented Jul 8, 2023

Uh oh!

jjerphan commented Jul 9, 2023

Uh oh!

betatim commented Jul 10, 2023

Uh oh!

jjerphan commented Jul 10, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

github-actions Bot commented Jul 5, 2023 •

edited

Loading

jjerphan commented Jul 6, 2023 •

edited

Loading