Crash when trying to get roc_auc score

[alexitkes]

Describe the bug

AutoSklearnClassifier crashes when trying to get its score if metric is set to roc_auc

To Reproduce

Steps to reproduce the behavior:
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from autosklearn.classification import AutoSklearnClassifier
from autosklearn.metrics import roc_auc

# Load some classification dataset
bc = load_breast_cancer()
(X, y) = (bc['data'], bc['target'])
(X_train, X_test, y_train, y_test) = train_test_split(X, y, random_state=0, stratify=y)
# Make sure it is a binary classification dataset
print(len(set(y)))
# 2 will be printed

# Accurcy metrics
clf = AutoSklearnClassifier(time_left_for_this_task=45, per_run_time_limit=15, n_jobs=-1)
clf.fit(X_train, y_train)
print(clf.score(X_train, y_train))
print(clf.score(X_test, y_test))
#  Displayed:
# 0.9929577464788732
# 0.9440559440559441

# And now...
clf = AutoSklearnClassifier(time_left_for_this_task=45, per_run_time_limit=15, n_jobs=-1)
clf.fit(X_train, y_train, metric=roc_auc)
print(clf.score(X_train, y_train))
print(clf.score(X_test, y_test))
# Crash!
IndexError                                Traceback (most recent call last)
<ipython-input-9-54810915749d> in <module>
      1 clf = AutoSklearnClassifier(time_left_for_this_task=45, per_run_time_limit=15, n_jobs=-1)
      2 clf.fit(X_train, y_train, metric=roc_auc)
----> 3 print(clf.score(X_train, y_train))
      4 print(clf.score(X_test, y_test))

~/venv-auto/lib/python3.6/site-packages/autosklearn/estimators.py in score(self, X, y)
    535 
    536     def score(self, X, y):
--> 537         return self._automl[0].score(X, y)
    538 
    539     def show_models(self):

~/venv-auto/lib/python3.6/site-packages/autosklearn/automl.py in score(self, X, y)
    731                                task_type=self._task,
    732                                metric=self._metric,
--> 733                                all_scoring_functions=False)
    734 
    735     @property

~/venv-auto/lib/python3.6/site-packages/autosklearn/metrics/__init__.py in calculate_score(solution, prediction, task_type, metric, all_scoring_functions)
    300             score = metric(solution, cprediction)
    301         else:
--> 302             score = metric(solution, prediction)
    303 
    304     return score

~/venv-auto/lib/python3.6/site-packages/autosklearn/metrics/__init__.py in __call__(self, y_true, y_pred, sample_weight)
    123 
    124         if y_type == "binary":
--> 125             y_pred = y_pred[:, 1]
    126         elif isinstance(y_pred, list):
    127             y_pred = np.vstack([p[:, -1] for p in y_pred]).T

Expected behavior

It should display two numbers, just like for accuracy scoring

Environment and installation:

Ubuntu 18.04.4
Python 3.6.9
Auto-Sklearn 0.7.0

Note that I can try to fix it myself, I think I know how it happens.

[mfeurer]

Note that I can try to fix it myself, I think I know how it happens.

This would be great. Otherwise, please let us know what you think is going wrong.

[alexitkes]

It's looking like score function is passing the predicted class vector to roc_auc scorer while it needs the probability matrix. I thought it could be fixed by setting needs_proba flag for roc_auc metric
https://github.com/alexitkes/auto-sklearn/blob/880eff289b40473bca4d23caa8c143c87505583d/autosklearn/metrics/__init__.py#L214

But it dit not work - now there are more errors while fitting.

[mfeurer]

Fixed via #879. Please reopen if this issue still persists with the development branch.