check_array(X, dtype='numeric') should fail if X has strings

Currently, dtype=‘numeric’ is defined as “dtype is preserved unless array.dtype is object”. This seems overly lenient and strange behaviour, as in #9342 where @qinhanmin2014 shows that check_array(['a', 'b', 'c'], dtype='numeric') works without error and produces an array of strings! This behaviour is not tested and it’s hard to believe that it is useful and intended. Perhaps we need a deprecation cycle, but I think dtype=‘numeric’ should raise an error, or attempt to coerce, if the data does not actually have a numeric, real-valued dtype.

[zeromh]

I’m confused about the fix here. I’m using sklearn 0.23.2, and the behavior that @jnothman called out as a problem is still the same as he described. To reproduce:

arr = np.array([[1, 's'],
                [1, 1]])
check_array(arr, dtype='numeric')

FutureWarning: Beginning in version 0.22, arrays of bytes/strings will be converted to decimal numbers
 if dtype='numeric'. It is recommended that you convert the array to a float dtype before using it in 
scikit-learn, for example by using your_array = your_array.astype(np.float64).
  return f(**kwargs)

array([['1', 's'],
       ['1', '1']], dtype='<U21')

Now everything’s a string. It looks like the warning message was added in 2018 but the behavior was never changed. Am I missing something?

[jnothman]

We wouldn’t deprecate check_array entirely, but we would warn for two releases that “In the future, this data with dtype(‘Uxx’) would be rejected because it is not of a numeric dtype.”