TST improve messages raised in check_classifier_multioutput #30235
Conversation
| "The prediction for multioutput data is expected to be of the same type " | ||
| f"as the input y. Got dtype={y_pred.dtype}, dtype.kind={y_pred.dtype.kind} " | ||
| f"instead, while input was dtype={y.dtype}, dtype.kind={y.dtype.kind}." |
There was a problem hiding this comment.
I suppose this is what we're testing. But not entirely sure.
|
|
||
| if hasattr(estimator, "decision_function"): | ||
| decision = estimator.decision_function(X) | ||
| assert isinstance(decision, np.ndarray) | ||
| assert isinstance(decision, np.ndarray), ( |
There was a problem hiding this comment.
So here we're checking the actual type of the output of decision_function. But this seems wrong(?) when I think of the array API work. WDYT @betatim @ogrisel @OmarManzoor
There was a problem hiding this comment.
Maybe we could check for xp.array?
| dec_pred = (decision > 0).astype(int) | ||
| dec_exp = estimator.classes_[dec_pred] | ||
| assert_array_equal(dec_exp, y_pred) |
There was a problem hiding this comment.
This parts is just odd to me. estimator.classes_=[0, 1, 2], and we do some fancy indexing using the results from the decision function. I can't imagine a case where this could fail, or to be meaningful, especially since the value 2 from classes_ would never be chosen here.
| np.argmax(y_prob[i], axis=1).astype(int), y_pred[:, i] | ||
| ) | ||
| elif not tags.classifier_tags.poor_score: | ||
| if not tags.classifier_tags.poor_score: |
There was a problem hiding this comment.
decoupling the shape check from the value check, and skipping only the value check if poor_score
|
@glemaitre I think you've worked on this quite a bit in the past. |
This PR improves error messages raised in
check_classifier_multioutput