This should have been fixed in #12625, and I can confirm that the corresponding commit is part of the 0.20.1 tag.

Could you please double-check that you are using 0.20.1. Also please post the full traceback, which will help us investigate.

I think @greghoop is using 0.20.1, seems that it's not appropriate to use len(array.dtypes)

@greghoop Please provide self-contained example code, including imports and data (if possible), so that other contributors can just run it and reproduce your issue. Ideally your example code should be minimal.

And please provide the output of sklearn.show_versions(). Thanks.

I cannot reproduce locally. And I think we have a test in Travis CI.

@qinhanmin2014

Please provide the output of sklearn.show_versions()

I've already downgraded to 0.20.0 so that won't help.

@greghoop Please provide self-contained example code

If I get a second I'll create an example, but I can't upload the current code.

The problem occurs for me on both Mac and Ubuntu in Python 3.7 with Pandas 0.23.4

Given the initial error message, I think this is indeed 0.20.1. The Series.dtypes attrribute returns a single dtype, and numpy dtypes have a length of 0:

In [21]: len(np.dtype('int64'))                                                                                                                                                 
Out[21]: 0

that was the characteristics we were using in

to deal with Series objects.

But I am now thinking of a case where the error can come up: with the custom pandas dtypes ...

In [25]: s = pd.Series(['a', 'b', 'c']).astype('category')                                                                                                                      

In [26]: s.dtypes                                                                                                                                                               
Out[26]: CategoricalDtype(categories=['a', 'b', 'c'], ordered=False)

In [27]: len(s.dtypes)                                                                                                                                                          
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-27-3d39f9ba3f74> in <module>
----> 1 len(s.dtypes)

TypeError: object of type 'CategoricalDtype' has no len()

@greghoop can it be the case that you were also using categorical data?

@greghoop

I've already downgraded to 0.20.0 so that won't help.

Downgrading to 0.20.0 won't help (downgrading to 0.19.X will)

If I get a second I'll create an example, but I can't upload the current code.

Thanks, I think you only need to provide the data type of y (you passed into check_array).

Apart from fixing it here in sklearn, we should consider on the pandas side if we want to fix this (pandas-dev/pandas#24033).
But that said, relying on the length of the dtype was maybe not a good idea. Because eg structured dtypes actually do have a length > 0. Although, at the moment you cannot store a structured array in a Series, so we should not be able to encounter that case.

I agree that we should avoid using len
@jorisvandenbossche Do you know when we'll get object of type 'DTYPE NAME OF THE SERIES' has no len()? I can't find relevant information in pandas.

@qinhanmin2014 Yes, see my example above: #12699 (comment) (at least, I assume that is what happened, @greghoop needs to confirm that, as you will never actually get DTYPE NAME OF THE SERIES in practice)

Yes, it was specifically for a categorical dtype series. Apologies for the confusion.

FYI, this isn’t a problem in 0.20.0 (I don’t need to downgrade to 0.19.x to get this working).

I don’t not know if it affects other dtypes

@jorisvandenbossche Thanks, I've seems your example above. I'm wondering what's type 'DTYPE NAME OF THE SERIES' since we need to reproduce the reporter's error. Any insights?

I'm wondering what's type 'DTYPE NAME OF THE SERIES' since we need to reproduce the reporter's error. Any insights?

That's just something that @greghoop put there, he apologized for the confusion :-)

Ahh, I see the comment from the contributor, let's replace len with other options.

@greghoop And thanks for the report!

@qinhanmin2014 I would maybe replace the and len(array.dtypes) with for example and hasattr(array.dtypes; 'iloc') (the duck type check we use in several places for pandas objects).
Ideally, I think we should start doing actual instance checks instead of those duck types, but that's for another issue.

Shall I do a PR, or are you already on it?

@jorisvandenbossche I'm not available now so please go ahead :)
FYI I proposed a solution in #12622 (comment)