should we say "got predict_proba of shape"?

That seems like a good idea, because that is what is actually happening.

But, will the user understand that? Is there some place in the documentation where we can write:

When obtaining predictions from a classifier, if it is only one class, the returned ValueError will be "got predict_proba of shape". That means it returns one column all with same probabilities (equal to 1).
Q: How to fix this?
A: Must use a classifier with 2 classes.

Or will they be able to see it from the test below:
with pytest.raises(ValueError, match="use classifier with two classes")
I would feel better if the user had explicit feedback on how to solve the error.

cc: @AMDonati

Are you saying the current error message is not informative enough? The test is just to make sure that the error that is raised is actually informative.
I suggested adding the predict_proba shape to the error message, not replace the message that's there now.

That works.
I tried making the update myself but ran into multiple problems: errors list

ok, making progress. Please let me know next steps.

cc: @AMDonati

you're not actually providing the shape.

ok, what is happening is this:

1. previously, when there was only one classifier, but expecting two, the y_pred was not resolving. This was the error: IndexError: index 1 is out of bounds for axis 1 with size 1
2. so, we made a check and an adjustment so that if the shape was 2 (meaning it only returned one column vector probabilities, which are all 1's), we want to raise a ValueError.
3. What should we call it? I don't think returning the shape (even if we did that) is informative. How about:

• "vector of probabilities returnedpredict_proba is all 1's because only one classifier is in the model."

Does that work?

aren't you meant to output the shape here?

Why is it a good idea to output the shape here? Isn't this more informative:
"vector of probabilities returnedpredict_proba is all 1's because only one classifier is in the model."

do you mean "only one class" rather than "only one classifier"?

Either way the proposed error message here says "got predict of shape; ..." that simply doesn't make sense.

How about I change the error message to:
"vector of probabilities returned (predict_proba) is all 1's because there is only one class in the input to the model."

It can also fail with GMMs where the shape would be wrong but the entries wouldn't be all ones.

yes, but you should use actual y_pred.shape[1] not hard-code 1 as there might be code-paths that have >2. (unless we can make sure that's not happening)

you mean like this?

            if y_pred.shape[1] == 2:
                y_pred = y_pred[:, 1]
            elif y_pred.shape[1] == 1:
                raise ValueError('got predict_proba of shape;'

Or, I think you mean this:

predict_proba has shape (n_samples, y_pred.shape[1]) which returns only 1 vector of probabilities (because it is a single class) but 2 classes in the data are required.

No I mean like

ValueError("got predict_proba of shape {}, but need classifier with two classes for {} scoring".format(y_pred.shape, self._score_func.__name__))

ok, I have made the updates.