[MRG+2] FIX DummyEstimator and a single output 2d input list/array

[adrinjalali]

Fixes #10786, closes #10926

Assumes the y is 1d if there's only one column to it.

[NicolasHug]

LGTM

[jnothman]

Is this the behaviour of other multioutput estimatiors?

Needs a what's new.

[adrinjalali]

@jnothman borrowing your gist:

import warnings
import sklearn
warnings.simplefilter('ignore')
from sklearn import *

X = [[0], [0], [0], [0], [0]]
y_1d = [1, 2, 1, 1, 1]
y = [[1], [2], [1], [1], [1]]

for clf in [tree.DecisionTreeClassifier(),
            neighbors.KNeighborsClassifier(),
            neural_network.MLPClassifier(),
            multioutput.MultiOutputClassifier(linear_model.LogisticRegression()),
            multiclass.OneVsRestClassifier(linear_model.LogisticRegression()),
            dummy.DummyClassifier(),
           ]:
    try:
        clf.fit(X, y)
        for method in ['predict', 'decision_function', 'predict_proba']:
            if not hasattr(clf, method):
                continue
            s = getattr(clf, method)(X[-3:])
            if hasattr(s, 'shape'):
                print(type(clf).__name__, method, s.shape)
            else:
                print(type(clf).__name__, method, [x.shape for x in s])
    except Exception as e:
        print(type(clf).__name__, e)

The output:

DecisionTreeClassifier predict (3,)
DecisionTreeClassifier predict_proba (3, 2)
KNeighborsClassifier predict (3,)
KNeighborsClassifier predict_proba (3, 2)
MLPClassifier predict (3,)
MLPClassifier predict_proba (3, 2)
MultiOutputClassifier predict (3, 1)
MultiOutputClassifier predict_proba [(3, 2)]
OneVsRestClassifier unhashable type: 'list'
DummyClassifier predict (3,)
DummyClassifier predict_proba (3, 2)

[jnothman]

Otherwise lgtm

[jnothman]

"1d dimensional y with ndim=2" makes no sense to me. Do you mean "column vector"?

[adrinjalali]

Yep, it was a copy paste from the old issue, sorry. Fixed.

[NicolasHug]

Thanks @adrinjalali !