AutoPipeline predict_proba and roc_auc score.

[hirzel]

• Implement the predict_proba method for the AutoPipeline operator.
• Add test for AutoPipeline with scoring="roc_auc", which has been reported to raise "AttributeError: The underlying operator impl does not define predict_proba", and get that to work as well.

[AmirhosseinHonardoust]

Hi @hirzel, I reproduced this issue locally.

with AutoPipeline(scoring="roc_auc"), fitting succeeds, but calling predict_proba on the trained AutoPipeline currently raises:

AttributeError: The underlying operator implementation class does not define predict_proba

I implemented a small fix that delegates AutoPipeline.predict_proba() to the selected best trained pipeline, and added input_predict_proba / output_predict_proba schemas.

I also added a regression test in test_core_pipeline.py for AutoPipeline(scoring="roc_auc") that fits the model and checks the shape of predict_proba.

the targeted test passes locally:

python -m pytest test/test_core_pipeline.py -k predict_proba_with_roc_auc

Result: 1 passed, 79 deselected.

I will open a PR for this shortly.