Register scorer objects for d2_log_loss_score and d2_brier_score
#32356
Conversation
|
@ogrisel I am looking into this if it's okay |
|
Let me add a changelog entry to document the changes. |
|
@OmarManzoor I am converting to draft because after profiling the new code I realized that most computation is spent in multiple redundant calls to |
|
I think this is ready for review. I did a bit of local profiling and the redundant computation has gone aways as expected, and I see a ~1.2x to ~1.5x speed-up compared to main. Let me redo a quick benchmark for |
|
Performance improvement over 1.7 confirmed and new changelog for the performance improvement pushed. I think we could easily do a follow-up PR to add array API support to those metrics, but I think this PR is ready as it is. |
OmarManzoor
left a comment
There was a problem hiding this comment.
This looks good. Thank you @ogrisel
virchan
left a comment
There was a problem hiding this comment.
LGTM! Thanks, @ogrisel!
@OmarManzoor, could we merge this?
I also tried to register those metrics in the common tests, but there are 2 failures to investigate in
sklearn/metrics/tests/test_common.py.EDIT: I also renamed
THRESHOLDED_METRICStoCONTINOUS_CLASSIFICATION_METRICS(and similar) in the test code because I found the original names very confusing, since those metrics accept unthresholded predictions contrary to other classification metrics.