OneHotEncoder generates a `2` and inconsistent data.

Description

When you create a OneHotEncoder with certain categories and then fit it with a numpy.array with dtype=object, when calling transform with different type of dtype it will generate an incoherent output.

Steps/Code to Reproduce

Example:

import numpy as np
from sklearn.preprocessing import OneHotEncoder

categories = np.array([['mse'], ['mae']], dtype='object')

encoder = OneHotEncoder(categories=[['mse','mae']], sparse=False)
encoder.fit(categories)

values = np.array([['mae'], ['mae'], ['mse'], ['mae'], ['mse'], ['mse']])
encoder.transform(values)

which outputs:

array([[1., 0.],
       [1., 0.],
       [0., 0.],
       [2., 0.],
       [0., 0.],
       [1., 0.]])

However if we run values.astype(object) we get the expected output:

encoder.transform(values.astype(object))

array([[0., 1.],
       [0., 1.],
       [1., 0.],
       [0., 1.],
       [1., 0.],
       [1., 0.]])

Versions

scikit-learn==0.21.3 scipy==1.3.3 numpy==1.17.4

[absognety]

switching the input for categories to auto is giving expected output. the handling of encoders is changed in version 0.22.

import numpy as np
from sklearn.preprocessing import OneHotEncoder

categories = np.array([['mse'], ['mae']], dtype='object')

encoder = OneHotEncoder(categories='auto',sparse=False)
encoder.fit(categories)

values = np.array([['mae'], ['mae'], ['mse'], ['mae'], ['mse'], ['mse']])
transformed = encoder.transform(values)
print (transformed)

array([[1., 0.],[1., 0.],[0., 1.],[1., 0.],[0., 1.],[0., 1.]])

versions:

scikit-learn == 0.21.3
scipy == 1.1.0
numpy == 1.15.1

[jnothman]

Does #15763 fix this?