[MRG + 1] Fix element-wise comparison for numpy - #8011

@lesteve: I followed your instructions which gave me the below error in this particular line:
multioutput = check_array(multioutput, ensure_2d=False)

Error:

TypeError: Singleton array array('variance_weighted', 
          dtype='|S17') cannot be considered a valid collection.

On a side note, what do you mean when you say install numpy from master ?

Error:

TypeError: Singleton array array('variance_weighted',
dtype='|S17') cannot be considered a valid collection.

Hmmm that means that somewhere the string 'variance_weighted' is turned into an array ... so the string_types clause is skipped. You will need to either:

1. understand where in our code we do this conversion 'variance_weighted' -> np.array('variance_weighted') and see how easy it is to get rid of it
2. treat the case of singleton string arrays in the string_types clause

My preference is for 1.

On a side note, what do you mean when you say install numpy from master ?

That means installing the numpy development version because that is what the original issue was about. You can still work on this issue without it but the only way to make sure the original issue is fixed is through Travis. This can be a bit cumbersome and frustrating because the feedback loop is a lot longer that testing things locally.

However, it does seem natural that check_array should have a default behavior if the array passed is a string, which it currently doesn't seem to have. Your views @lesteve?

@lesteve I think the string_types clause is not skipped but within the check_array condition it converts the string variance_weighted to an array using np.asarray("variance_weighted"). Later we check if that array is singleton or not and raise the above error if it is singleton. @dalmia Would you like to help me out here ?

Sure @aashil. From what I see, you need to add another check if multioutput is among the valid multioutput_options and do the corresponding functionality there. So just change:

elif multioutput is not None:

To:

elif multioutput is not None and multioutput not in multioutput_options:

That should be all.

@dalmia I believe you mean elif multioutput is not None and multioutput in multioutput_options: But the real problem is in check_array() method inside the elif clause which is not happy with the array being singleton. Take a look at my latest commit.

@aashil Sorry I made a small mistake. Also, I don't think I conveyed properly what I intended to say. Let me elaborate what I intend to say. _check_reg_targets intends to return a proper value of multioutput. So, if a valid multioutput string is passed as a parameter, it won't be modified. However, if it's not a string, only then it needs to go for check_array.
The whole patch then becomes:

multioutput_options = (None, 'raw_values', 'uniform_average',
                           'variance_weighted')
   # If it is a string, but not a valid option, raise an error
    if isinstance(multioutput, string_types) and multioutput not in multioutput_options:
        raise ValueError("Invalid multioutput option")
   # If it is not a string then check for the validity of the array
    elif multioutput is not None and not isinstance(multioutput, string_types):
        multioutput = check_array(multioutput, ensure_2d=False)
        if n_outputs == 1:

@dalmia Ahh, that makes it so clear. Thank you.

Sure, happy to help :)

@aashil I pushed a cosmetic change (I feel the if clause logic is more readable this way) and a test for the error message, have a look at a6efd19.

LGTM

Tested locally. LGTM

OK, merging then, thanks a lot @aashil!