cross_val_score passes a list rather than a DataFrame to Estimator

[cancan101]

Docs say:

sklearn-pandas provides a wrapper on sklearn's cross_val_score function which passes a pandas DataFrame to the estimator rather than a numpy array

but I see:

test = pd.DataFrame({'a':[1,2,3]})
class Model(BaseEstimator):
    def fit(self, x):
        print type(x)
    def score(self, x):
        return 1
​
_ = cross_val_score(Model(), test)

<type 'list'>
<type 'list'>
<type 'list'>

[cancan101]

The issue is in _safe_split:

type(_safe_split(Model(), DataWrapper(test), None, [1, 2])[0])
list

[cancan101]

It actually looks like the newest version of scikit handles Dataframes natively.

[dukebody]

@cancan101 you are probably right regarding your last comment.

Regarding the cross_val_score wrapper, it was intended to be able to be able to use a dataframe with the cross-validation functions in pipelines. The list you see is in fact a somewhat dirty trick to avoid sklearn to turn the dataframe into an array. That list is later mapped to a DataFrame inside the DataFrameMapper. But you cannot use the cv-wrapper directly without the DataFrameMapper.

If you want to use a dataframe in cv with old versions of sklearn, simply pass df.values to the sklearn cv functions.

I'm closing this because I don't consider it a bug.