[MRG] text vectorizers should raise warnings when user params will be unused

[getgaurav2]

Reference Issues/PRs

#14580

What does this implement/fix? Explain your changes.

Add Warning and test cases for unused parameters for text vectorizers.

1. Add all warnings in a new method - _warn_for_unused_params() and call this method in the fit() method for all vectorizers .
2. Add Unit test case for all warning checks .

Any other comments?

[getgaurav2]

@jnothman , @rth - can you pls. give any feedback on the test case ?
I can add the other text vectorizers in the same function and iterate over the objects of each vectorizers and check different warnings msg for different parameter combinations
OR
I can create one unit test function per vectorizer per warning .

[jnothman]

Looking good

[getgaurav2]

[getgaurav2]

Adding only CountVectorizer for now . Will add others as I add warnings for them in the text.py file.

[getgaurav2]

@jnothman , @rth - can you please check why some of the checks are failing for this commit.
Here is the log :
scikit-learn.scikit-learn (Linux py35_conda_openblas)
PackagesNotFoundError: The following packages are not available from current channels:

• numpy=1.11.0
• scipy=0.17.0
• matplotlib=1.5.1
• pillow=4.0.0

Linux pylatest_conda_mkl_pandas
~/work/tmp_folder ~/work/1/s
build_tools/azure/test_pytest_soft_dependency.sh: line 21: coverage: command not found

[thomasjpfan]

It is a unrelated issue.

It should be fixed when #14619 get merged.

[rth]

LGTM, apart for the two comments below.

[jnothman]

Thanks

[jnothman]

Tests are failing . Let us know if you need help

[getgaurav2]

Tests are failing . Let us know if you need help

Thanks. A different unit test test_callable_analyzer_change_behavior seems to be failing ...I'll check the reason and get back to you .

[getgaurav2]

Tests are failing . Let us know if you need help

Thanks. A different unit test test_callable_analyzer_change_behavior seems to be failing ...I'll check the reason and get back to you .

@jnothman - Could you help me understand the reason for adding assert len(records) == 1 in test_callable_analyzer_change_behavior . This condition is failing for HashingVectorizer ( screen shot below). It passes for CountVectorizer and HashingVectorizer . I have commented this condition in the test case for now but will add it back once I understand it better.

The difference b/w the fit() method for HashingVectorizer () vs. the others is calling build_analyzer() . https://github.com/scikit-learn/scikit-learn/pull/14602#issuecomment-521613247

[jnothman]

It's saying that HashingVectorizer is now issuing the ChangedBehaviorWarning twice. It should only issue it once. By putting build_analyzer in its fit, it might now be repeating that warning.

[getgaurav2]

It's saying that HashingVectorizer is now issuing the ChangedBehaviorWarning twice. It should only issue it once. By putting build_analyzer in its fit, it might now be repeating that warning.

Yes , build_analyzer () is now being called inside fit() as well as transform() methods which is likely causing the double warnings . Looking into it now .
@jnothman - Do you think it may be easier to go back to the original approach and keep the analyzer != ‘word’ warnings outside of build_ analyzer ?
#14602 (comment)

[getgaurav2]

@jnothman - just wanted to make sure that you are getting a notification for the comment above.

[jnothman]

Otherwise LGTM. (Though I've not checked that the set of parameter pairs is complete.)

Please add an Enhancement entry to the change log at doc/whats_new/v0.22.rst. Like the other entries there, please reference this pull request with :pr: and credit yourself (and other contributors if applicable) with :user:

[jnothman]

Assuming you also think this is ready for merge, please change WIP -> MRG

[jnothman]

Please add a note in what's new as requested above. Thanks!

[getgaurav2]

Please add a note in what's new as requested above. Thanks!

Yup ..Was just working on it. Added now .

[jnothman]

Please describe changes only in terms of public API: "a warning is now raised if a parameter choice means that another parameter will be unused" or something more clear than that.

[getgaurav2]

Got it ..Can you please check if the latest commit makes more sense .

[jnothman]

Makes more sense. We might yet find clearer words before release, but as far as I'm concerned we can merge. Thanks, @getgaurav2!