SimpleImputer doesn't support infinite values

[gjimzhou]

Description

SimpleImputer will throw ValueError if the array passed in contains np.inf which is not the case in the old Imputer.

This may be caused by calling the _validate_input before actual fitting or transforming.

Steps/Code to Reproduce

s = np.array([-np.inf, 2, 3, 4, 5, 6, np.inf]).reshape(-1, 1)
from sklearn.impute import SimpleImputer
transformer = SimpleImputer(strategy='median')
transformer.fit(s)

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-28-ee05450ab4c6> in <module>()
      3 transformer = SimpleImputer(strategy='median')
      4 
----> 5 transformer.fit(s)

~/anaconda3/lib/python3.6/site-packages/sklearn/impute.py in fit(self, X, y)
    221         self : SimpleImputer
    222         """
--> 223         X = self._validate_input(X)
    224 
    225         # default fill_value is 0 for numerical input and "missing_value"

~/anaconda3/lib/python3.6/site-packages/sklearn/impute.py in _validate_input(self, X)
    195                                  "".format(self.strategy, X.dtype.kind))
    196             else:
--> 197                 raise ve
    198 
    199         _check_inputs_dtype(X, self.missing_values)

~/anaconda3/lib/python3.6/site-packages/sklearn/impute.py in _validate_input(self, X)
    188         try:
    189             X = check_array(X, accept_sparse='csc', dtype=dtype,
--> 190                             force_all_finite=force_all_finite, copy=self.copy)
    191         except ValueError as ve:
    192             if "could not convert" in str(ve):

~/anaconda3/lib/python3.6/site-packages/sklearn/utils/validation.py in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, warn_on_dtype, estimator)
    571         if force_all_finite:
    572             _assert_all_finite(array,
--> 573                                allow_nan=force_all_finite == 'allow-nan')
    574 
    575     shape_repr = _shape_repr(array.shape)

~/anaconda3/lib/python3.6/site-packages/sklearn/utils/validation.py in _assert_all_finite(X, allow_nan)
     54                 not allow_nan and not np.isfinite(X).all()):
     55             type_err = 'infinity' if allow_nan else 'NaN, infinity'
---> 56             raise ValueError(msg_err.format(type_err, X.dtype))
     57 
     58 

ValueError: Input contains infinity or a value too large for dtype('float64').

Expected Results

s = np.array([-np.inf, 2, 3, 4, 5, 6, np.inf]).reshape(-1, 1)
from sklearn.preprocessing import Imputer
transformer = Imputer(strategy='median')
transformer.fit(s)

/home/jzhou/anaconda3/lib/python3.6/site-packages/sklearn/utils/deprecation.py:58: DeprecationWarning:

Class Imputer is deprecated; Imputer was deprecated in version 0.20 and will be removed in 0.22. Import impute.SimpleImputer from sklearn instead.

Imputer(axis=0, copy=True, missing_values='NaN', strategy='median', verbose=0)

Versions

System:
    python: 3.6.8 |Anaconda, Inc.| (default, Dec 30 2018, 01:22:34)  [GCC 7.3.0]
executable: /home/jzhou/anaconda3/bin/python
   machine: Linux-4.10.0-42-generic-x86_64-with-debian-stretch-sid

BLAS:
    macros: SCIPY_MKL_H=None, HAVE_CBLAS=None
  lib_dirs: /home/jzhou/anaconda3/lib
cblas_libs: mkl_rt, pthread

Python deps:
       pip: 19.0.3
setuptools: 40.8.0
   sklearn: 0.20.2
     numpy: 1.16.2
     scipy: 1.2.1
    Cython: 0.28.2
    pandas: 0.24.1

[jnothman]

What use case do you have for infinite values?

[glemaitre]

If there is a real use case to keep the np.inf you can use the sklearn.config_context without checking for infinite values. However, I think that the current implementation is better in general.