Codestin Search App

randomgeek78 · 2022-03-25T19:59:51Z

Currently, checking if the feature union is fitted fails. This is because it does not have any attributes of its own that end with '_' and it does not have sklearn_is_fitted implemented.

This PR delegates checking if a feature union was fitted to one of its sub-transformers.

Reference Issues/PRs

What does this implement/fix? Explain your changes.

Any other comments?

thomasjpfan

Thank you for the PR!

thomasjpfan · 2022-03-25T20:15:41Z


+    def __sklearn_is_fitted__(self):
+        # Delegate whether feature union was fitted
+        check_is_fitted(self.transformer_list[0][1])


I think we need to check all the transformers to make sure the whole FeatureUnion is fitted.

Also transformer_list can be configured with "passthrough" and "drop" as the estimator, so those needs to be avoided in this check. (In other words, transformer_list[0][1] can be string.)

Thanks @thomasjpfan. Let me know if my fix works. Tx

thomasjpfan

Thanks for the update!

thomasjpfan · 2022-03-26T18:21:11Z



+def test_feature_union_check_if_fitted():
+    class Estimator(BaseEstimator):


We can use from sklearn.utils._testing import MinimalTransformer for a minimal transformer.

thomasjpfan · 2022-03-26T18:26:34Z

+    union = clone(union)
+    with pytest.raises(NotFittedError):
+        check_is_fitted(union)
+
+    union.set_params(clf="drop")
+    check_is_fitted(union)
+
+    union.set_params(clf=Estimator())
+    with pytest.raises(NotFittedError):
+        check_is_fitted(union)
+
+    union.fit(iris.data, iris.target)
+    check_is_fitted(union)


I think we can remove these tests. It's fundamentally testing that clone works. In other words, if clone works and the tests above pass, then these should pass as well.

@thomasjpfan I just pushed out an update. Tx

thomasjpfan

Minor nits, otherwise LGTM

thomasjpfan · 2022-03-28T18:51:09Z

+- |Fix| Checking if :class:`pipeline.FeatureUnion` was fitted was broken.
+  :pr:`22953` by :user:`randomgeek78 <randomgeek78>` provides a fix.


We can be more explicit about the fix:

Suggested change

- |Fix| Checking if :class:`pipeline.FeatureUnion` was fitted was broken.

:pr:`22953` by :user:`randomgeek78 <randomgeek78>` provides a fix.

- |Fix| Defines `__sklearn_is_fitted__` in

:class:`pipeline.FeatureUnion` to return correct result

with :func:`utils.validation.check_is_fitted`.

:pr:`22953` by :user:`randomgeek78 <randomgeek78>`

thomasjpfan · 2022-03-28T18:51:27Z

    check_is_fitted(pipeline)


+def test_feature_union_check_if_fitted():


Suggested change

def test_feature_union_check_if_fitted():

def test_feature_union_check_if_fitted():

"""Check __sklearn_is_fitted__ is defined correctly."""

Great feedback. Just pushed an update. Tx

@thomasjpfan I think linting is currently broken due to a black dependency.

@thomasjpfan This PR does not build because of this message: Could not find a version that satisfies the requirement black==22.1.0. Do you know anything about this? Thanks.

randomgeek78 · 2022-04-03T07:21:47Z

@adrinjalali @ogrisel Could you help review this PR? It requires a second approval. Thanks!

adrinjalali

we should also test how drop works here.

adrinjalali · 2022-04-04T09:42:40Z

+        for _, transformer, _ in self._iter():
+            check_is_fitted(transformer)
+        return True


the issue with this implementation is that we allow all transformers to be None in FeatureUnion. fit's code is:

transformers = self._parallel_func(X, y, fit_params, _fit_one) if not transformers: # All transformers are None return self self._update_transformer_list(transformers) return self

wouldn't it be better here to add a private _fitted attribute in fit, and in this method to check for that?

If we want to also support pre-fitting transformers and then putting them in a FeatureUnion, we could do something like:

if getattr(self, '_fitted', None) or all_transformers_fitted: return True

I think the current solution works because _iter does not yield dropped estimators:

scikit-learn/sklearn/pipeline.py

Lines 1052 to 1057 in 231dd7b

for name, trans in self.transformer_list:

if trans == "drop":

continue

if trans == "passthrough":

trans = FunctionTransformer()

yield (name, trans, get_weight(name))

(Also FeatureUnion no longer supports None and uses `"drop" instead. )

@adrinjalali If all the transformers are None, it is stateless and is by definition fitted (imho). This is the approach that is also taken by FunctionTransformer:

def __sklearn_is_fitted__(self): """Return True since FunctionTransfomer is stateless.""" return True

Instead of all transformers being None, I am fine returning true even if all transformers are stateless (like FunctionTransformer). What do you think? If you feel strongly about introducing state to FeatureUnion through _fitted, I can push a PR with the required changes. Thanks for your feedback.

Regarding drop, what would you like to test for it? As you know, transformers that are dropped are not yielded by _iter.

@thomasjpfan Thanks. I didn't see your message while composing my responses.

…learn#22953) * Fix to checking if feature union was fitted * Updated test and linting * Added entry in whats new 1.1 * Expanded tests to all transformers and strings * Expanded test * Linting fix * Updated tests * Addressed additional feedback

Fix to checking if feature union was fitted

deb0260

github-actions Bot added the module:pipeline label Mar 25, 2022

randomgeek78 added 2 commits March 25, 2022 13:05

Updated test and linting

5d99959

Added entry in whats new 1.1

30db1b8

thomasjpfan reviewed Mar 25, 2022

View reviewed changes

randomgeek78 added 3 commits March 25, 2022 13:36

Expanded tests to all transformers and strings

f61f6e2

Expanded test

1bd53e5

Linting fix

d9d2456

thomasjpfan reviewed Mar 26, 2022

View reviewed changes

Updated tests

71c1543

randomgeek78 force-pushed the patch-2 branch from 4322f9c to 71c1543 Compare March 28, 2022 18:15

thomasjpfan approved these changes Mar 28, 2022

View reviewed changes

randomgeek78 and others added 2 commits March 28, 2022 12:10

Addressed additional feedback

274f675

Merge branch 'scikit-learn:main' into patch-2

d0f1834

adrinjalali reviewed Apr 4, 2022

View reviewed changes

adrinjalali approved these changes Apr 5, 2022

View reviewed changes

adrinjalali merged commit 106ea32 into scikit-learn:main Apr 5, 2022

randomgeek78 deleted the patch-2 branch April 6, 2022 05:16



		def test_feature_union_check_if_fitted():
		class Estimator(BaseEstimator):

		- \|Fix\| Checking if :class:`pipeline.FeatureUnion` was fitted was broken.
		:pr:`22953` by :user:`randomgeek78 <randomgeek78>` provides a fix.

		check_is_fitted(pipeline)


		def test_feature_union_check_if_fitted():

	def test_feature_union_check_if_fitted():
	def test_feature_union_check_if_fitted():
	"""Check __sklearn_is_fitted__ is defined correctly."""

	for name, trans in self.transformer_list:
	if trans == "drop":
	continue
	if trans == "passthrough":
	trans = FunctionTransformer()
	yield (name, trans, get_weight(name))

Uh oh!

Uh oh!

Conversation

randomgeek78 commented Mar 25, 2022

Reference Issues/PRs

What does this implement/fix? Explain your changes.

Any other comments?

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!

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!

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!

randomgeek78 commented Apr 3, 2022

Uh oh!

adrinjalali 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!

randomgeek78 Apr 4, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

randomgeek78 Apr 4, 2022 •

edited

Loading