Add n_components="auto" to NMF when H and W are provided #26634

Alexlandeau · 2023-06-20T14:01:24Z

adding "auto" as default value for n_components. In this mode, n_components will be inferred from H and W if they are provided
adding additional validation on inferred n_components to prevent inconsistencies
updating parameter validation, deprecation notice, documentation and tests accordingly

* adding "auto" as default value for n_components * updating parameter validation, deprecation notice, documentation and tests accordingly --------- Co-authored-by: Alexandre Landeau <[email protected]> Co-authored-by: avigny <[email protected]>

github-actions · 2023-06-20T14:04:01Z

✔️ Linting Passed

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

_{Generated for commit: 024ed75. Link to the linter CI: here}

jeremiedbb

Thanks for the PR @Alexlandeau.

Here are some comments. In particular, this pr should target version 1.4 since the feature freeze for 1.3 already happened. Also, the addition of auto and future change of default should also be done for NMF and MiniBatchNMF.

Please also add an entry in the v1.4.rst changelog.

sklearn/decomposition/_nmf.py

sklearn/decomposition/tests/test_nmf.py

Co-authored-by: Jérémie du Boisberranger <[email protected]>

adding `init="custom"` so that we actually test the modified code

…nto bug/nmf_26392

This allows to check the shapes of w or h if given to the estimator. added the corresponding test as well

jeremiedbb

Looking good. Here's a new batch of comments, mostly nitpicks.

doc/whats_new/v1.4.rst

sklearn/decomposition/_nmf.py

sklearn/decomposition/tests/test_nmf.py

jeremiedbb · 2023-06-22T11:35:40Z

sklearn/decomposition/tests/test_nmf.py

+        X, H=H_true, n_components="auto", update_H=False
+    )  # should not fail


you can check that H.shape has not changed H.shape == H_true.shape.
you can also check that W has the appropriate shape.

done in e5cffdb

jeremiedbb · 2023-06-22T11:41:06Z

sklearn/decomposition/tests/test_nmf.py

+def test_nmf_n_components_auto(Estimator):
+    rng = np.random.RandomState(0)


Please add a small description of the goal of this test, like "Check that n_components is correctly inferred from the provided custom initialization."

And similar for other tests

done in 0e65acd

jeremiedbb · 2023-06-22T12:03:31Z

CI is failing because this PR introduces many FutureWarnings. You need to check every place where we call (MiniBatch)NMF or non_negative_factorization where n_components is left to its default value and change it to "auto" :)

This reverts commit 918cdbe.

avigny · 2023-06-23T09:10:59Z

Should we also add n_components="auto" this example? I haven't found other places where the warning could occur

scikit-learn/examples/compose/plot_compare_reduction.py

Line 57 in bcbc2fc

"reduce_dim": [PCA(iterated_power=7), NMF(max_iter=1_000)],

jeremiedbb · 2023-06-23T09:25:21Z

Should we also add n_components="auto" this example? I haven't found other places where the warning could occur

Yes, we don't want warnings in the examples either

avigny · 2023-06-23T09:41:17Z

Should we also add n_components="auto" this example? I haven't found other places where the warning could occur

Yes, we don't want warnings in the examples either

I was wrong, the n_components values are set by the GridSearchCV so there is no need to add a value for n_components

scikit-learn/examples/compose/plot_compare_reduction.py

Lines 53 to 70 in bcbc2fc

    
           N_FEATURES_OPTIONS = [2, 4, 8] 
        
           C_OPTIONS = [1, 10, 100, 1000] 
        
           param_grid = [ 
        
               { 
        
                   "reduce_dim": [PCA(iterated_power=7), NMF(max_iter=1_000)], 
        
                   "reduce_dim__n_components": N_FEATURES_OPTIONS, 
        
                   "classify__C": C_OPTIONS, 
        
               }, 
        
               { 
        
                   "reduce_dim": [SelectKBest(mutual_info_classif)], 
        
                   "reduce_dim__k": N_FEATURES_OPTIONS, 
        
                   "classify__C": C_OPTIONS, 
        
               }, 
        
           ] 
        
           reducer_labels = ["PCA", "NMF", "KBest(mutual_info_classif)"] 
        
           grid = GridSearchCV(pipe, n_jobs=1, param_grid=param_grid) 
        
           grid.fit(X, y)

jeremiedbb

LGTM. Thanks @Alexlandeau and @avigny

I fixed the remaining failing test and triggered the CI to run the examples to make sure we don't miss future warnings there. It'll be a longer run.

sklearn/decomposition/_nmf.py

ogrisel · 2023-06-23T15:11:09Z

sklearn/decomposition/_nmf.py

+            if H is not None:
+                self._n_components = _num_samples(H)
+            elif W is not None:
+                self._n_components = _num_features(W)


Could you please extend the tests to cover this branch?

Small note about this branch:

if H is None and W is not None we now set self._n_components = _num_features(W) instead of self._n_components = X.shape[1]

Situation before to the PR

If you have init = "custom" ==> Error because W is not set
If you have init != "custom" ==> A new W is created with the shape n_samples, X.shape[1]. This is misleading because X.shape[1] can be different from _num_features(W)
see

scikit-learn/sklearn/decomposition/_nmf.py

Lines 1208 to 1216 in 9cbcc1f

if self.solver == "mu":

avg = np.sqrt(X.mean() / self._n_components)

W = np.full((n_samples, self._n_components), avg, dtype=X.dtype)

else:

W = np.zeros((n_samples, self._n_components), dtype=X.dtype)

else:

W, H = _initialize_nmf(

X, self._n_components, init=self.init, random_state=self.random_state

)

Situation after the PR

If you have init = "custom" [Same behaviour] ==> Error because W is not set
If you have init != "custom" [Changed behaviour]==> a new W is created with but now the shape of W will be the same as the shape of the W passed as argument

the test was added in d8b6bf8

If you have init != "custom" [Changed behaviour]==> a new W is created with but now the shape of W will be the same as the shape of the W passed as argument

I don't think this is a desirable behavior. It's also a breaking change. In that case we should completely ignore W and fallback to X.shape[1].

It requires a little bit of refactoring because it means that we can't compute n_components before checking in which situation we are. I directly pushed some changes to fix the behavior. The benefit is that we no longer have to rely on _num_samples and _num_features :)

sklearn/decomposition/tests/test_nmf.py

ogrisel

Once the missing test is added, LGTM.

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

…nto bug/nmf_26392

Alexlandeau · 2023-06-27T08:32:36Z

Thanks @jeremiedbb for the last review and suggestions !

avigny · 2023-06-27T08:39:05Z

Thanks!!

yotamcons · 2023-06-27T09:44:26Z

Thank you @Alexlandeau for this!

…cikit-learn#26634) Co-authored-by: avigny <[email protected]> Co-authored-by: Jérémie du Boisberranger <[email protected]> Co-authored-by: Olivier Grisel <[email protected]>

Fix issue 26392 on NMF

01431f4

* adding "auto" as default value for n_components * updating parameter validation, deprecation notice, documentation and tests accordingly --------- Co-authored-by: Alexandre Landeau <[email protected]> Co-authored-by: avigny <[email protected]>

github-actions bot added the module:decomposition label Jun 20, 2023

ogrisel changed the title ~~Fix issue 26392 on NMF~~ Add n_components="auto" to NMF when H and W are provided Jun 20, 2023

Alexandre Landeau added 2 commits June 21, 2023 10:23

fixed linting

118da1d

Merge branch 'main' into bug/nmf_26392

4a5010b

jeremiedbb reviewed Jun 21, 2023

View reviewed changes

Alexandre Landeau and others added 16 commits June 21, 2023 14:13

linting

ada48ab

Apply suggestions from code review

15f066e

Co-authored-by: Jérémie du Boisberranger <[email protected]>

import _param_validation.hidden and close parenthesis

3b2e45e

Doc for fit_transform of MiniBatchNMF and NMF classes updated

28a595f

Adding versionchanged directive about ading 'auto` value

1c6994b

Moving the n_components validation in the _check_params function

9f031da

fixed _n_components initialization when init method is custom

04ac4ee

added additional validation on W and H shapes when provided

261e540

Update tests

af56554

adding `init="custom"` so that we actually test the modified code

Merge branch 'main' into bug/nmf_26392

046f813

merge conflicts

7e2b92d

merge conflicts pt.2

e86435b

Merge branch 'main' into bug/nmf_26392

2aa8d0b

added changelog

6f43e74

Merge branch 'bug/nmf_26392' of github.com:Alexlandeau/scikit-learn i…

b8fc2dc

…nto bug/nmf_26392

Allow n_components="auto" independently of the init method.

5042418

This allows to check the shapes of w or h if given to the estimator. added the corresponding test as well

avigny force-pushed the bug/nmf_26392 branch from 535893e to 5042418 Compare June 22, 2023 10:01

jeremiedbb reviewed Jun 22, 2023

View reviewed changes

Alexandre Landeau and others added 4 commits June 22, 2023 14:41

PR feedback: doc improvement and warning suppression

b3529bb

merge conflicts

b524958

Warn the user if W or H is set and will be overwritten at init

82fbc4c

do an actual test in test_nmf_n_components_auto_no_h_update

e5cffdb

avigny added 2 commits June 23, 2023 10:57

Revert "repair tests failing due to FutureWarning not caught"

dcbe421

This reverts commit 918cdbe.

filter warnings related to n_components

bcbc2fc

avigny added 3 commits June 23, 2023 11:21

test repair once again with Future Warning about n_components

219a9c2

More precise changelog

b398b6e

remove linebreak

445a943

[doc build]

c965924

jeremiedbb approved these changes Jun 23, 2023

View reviewed changes

Alexlandeau added 2 commits June 23, 2023 14:43

Merge branch 'main' into bug/nmf_26392

9eca680

Merge branch 'main' into bug/nmf_26392

0062004

ogrisel reviewed Jun 23, 2023

View reviewed changes

ogrisel approved these changes Jun 23, 2023

View reviewed changes

Alexandre Landeau and others added 10 commits June 23, 2023 18:18

added test when W is used and not H

d8b6bf8

Update sklearn/decomposition/tests/test_nmf.py

cc0886e

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

Update sklearn/decomposition/tests/test_nmf.py

88820f3

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

Update sklearn/decomposition/tests/test_nmf.py

cd40814

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

Update sklearn/decomposition/_nmf.py

58e8c26

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

Merge branch 'main' into bug/nmf_26392

1eabeb6

linting

9f2140c

Merge branch 'bug/nmf_26392' of github.com:Alexlandeau/scikit-learn i…

7d9a75f

…nto bug/nmf_26392

never use W to infer n_components when init!=custom

15bd41f

improve coverage

024ed75

jeremiedbb merged commit ee5d94e into scikit-learn:main Jun 26, 2023

		X, H=H_true, n_components="auto", update_H=False
		) # should not fail

		def test_nmf_n_components_auto(Estimator):
		rng = np.random.RandomState(0)

	if self.solver == "mu":
	avg = np.sqrt(X.mean() / self._n_components)
	W = np.full((n_samples, self._n_components), avg, dtype=X.dtype)
	else:
	W = np.zeros((n_samples, self._n_components), dtype=X.dtype)
	else:
	W, H = _initialize_nmf(
	X, self._n_components, init=self.init, random_state=self.random_state
	)

Uh oh!

Add n_components="auto" to NMF when H and W are provided #26634

Add n_components="auto" to NMF when H and W are provided #26634

Uh oh!

Conversation

Alexlandeau commented Jun 20, 2023

Uh oh!

github-actions bot commented Jun 20, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✔️ Linting Passed

Uh oh!

jeremiedbb left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jeremiedbb left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

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!

jeremiedbb commented Jun 22, 2023

Uh oh!

avigny commented Jun 23, 2023

Uh oh!

jeremiedbb commented Jun 23, 2023

Uh oh!

avigny commented Jun 23, 2023

Uh oh!

jeremiedbb left a comment

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

Situation before to the PR

Situation after the PR

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ogrisel left a comment

Choose a reason for hiding this comment

Uh oh!

Alexlandeau commented Jun 27, 2023

Uh oh!

avigny commented Jun 27, 2023

Uh oh!

yotamcons commented Jun 27, 2023

Uh oh!

Reviewers

Assignees

Labels

github-actions bot commented Jun 20, 2023 •

edited

Loading