FIX check_estimator fails when validating SGDClassifier with log_loss #24071
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
glemaitre
reviewed
Oct 17, 2022
sklearn/utils/estimator_checks.py
Outdated
Comment on lines
3463
to
3470
| a = estimator.predict_proba(X_test)[:, 1].round(decimals=10) | ||
| b = estimator.decision_function(X_test).round(decimals=10) | ||
| assert_array_equal(rankdata(a), rankdata(b)) | ||
|
|
||
| prob_rank = rankdata(a) | ||
| # calculate the average decision score groupby the rank of predicted probability | ||
| avg_decision_score = [np.mean(b[prob_rank == i]) for i in np.unique(prob_rank)] | ||
| # check if the average decision score is strictly increasing | ||
| assert all(x < y for x, y in zip(avg_decision_score, avg_decision_score[1:])) |
There was a problem hiding this comment.
I think that we can relax in this manner.
I think that just for readability, I would propose the following:
y_proba = estimator.predict_proba(X_test)[:, 1].round(decimals=10)
y_score = estimator.decision_function(X_test).round(decimals=10)
rank_proba = rankdata(y_proba)
rank_score = rankdata(y_score)
try:
assert_array_almost_equal(rank_proba, rank_score)
except AssertionError:
# Sometimes, the rounding applied on the probabilities will results
# on ties that are not present in the scores because it is
# numerically more precise. In this case, we relax the test by
# grouping the decision function scores based on the probability
# rank and check that the score is monotonically increasing.
grouped_y_score = np.array(
[y_score[rank_proba == group].mean() for group in np.unique(rank_proba)]
)
sorted_idx = np.argsort(grouped_y_score)
assert_array_equal(sorted_idx, np.arange(len(sorted_idx)))There was a problem hiding this comment.
That's much more readable, updated :)
glemaitre
reviewed
Oct 17, 2022
There was a problem hiding this comment.
We can also add a non-regression test in sklearn/utils/tests/test_estimator_checks.py:
def test_decision_proba_tie_ranking():
"""Check that in case with some probabilities ties, we relax the
ranking comparison with the decision function.
Non-regression test for:
https://github.com/scikit-learn/scikit-learn/issues/24025
"""
# Move these imports to the top of the file
from sklearn.linear_model import SGDClassifier
from sklearn.utils.estimator_checks import check_decision_proba_consistency
estimator = SGDClassifier(loss="log_loss")
check_decision_proba_consistency("SGDClassifier", estimator)
glemaitre
approved these changes
Oct 18, 2022
glemaitre
pushed a commit
to glemaitre/scikit-learn
that referenced
this pull request
Oct 31, 2022
…scikit-learn#24071) Co-authored-by: Thomas J. Fan <thomasjpfan@gmail.com>
andportnoy
pushed a commit
to andportnoy/scikit-learn
that referenced
this pull request
Nov 5, 2022
…scikit-learn#24071) Co-authored-by: Thomas J. Fan <thomasjpfan@gmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Reference Issues/PRs
Fixes #24025.
What does this implement/fix? Explain your changes.
The test that failed was
check_decision_proba_consistencywhich checks that the output ofpredict_probanddecision functionhas perfect rank correlation. Meanwhile the probability output may have ties, which causes the rank to be different.Instead of checking the rank to be exactly the same, the proposed fix checks that the average decision score is strictly increasing after grouped by the rank of predicted probability.