[MRG] FIX numpy 1.3 issues with the new multilabel metrics

[arjoly]

With the support of multilabel input for the zero_one_loss, accuracy_score and hamming_loss, there were 2 bugs with numpy 1.3:

• np.setxor1d does not make unique by default (and so, there were some failures with redundant labels)
• there were some failures with 0-dimensionality array

This patch intend to solve both of these issues.
For more information, see #1606 and jenkins console log.

[ogrisel]

typo: required

[ogrisel]

BTW, this typo is replicated several times.

[ogrisel]

I am not sure how np.setxor1d works, but if the tests pass, I am ok with merging this :)

[arjoly]

I fix the typo!

[ogrisel]

@larsmans @amueller this looks good to me. Shall we merge?

[amueller]

maybe add a comment both are in list of labels format

[arjoly]

Why both?

[amueller]

y_pred and y_true.

[amueller]

to me it looks a bit as if the updated part should be refactored into a separate function. otherwise it looks good.

[arjoly]

If you want, I can backport numpy.setxor1d.

There are other part that could be simplify once #1643 will be merged.
I will do that in a future pr.

[amueller]

I rather thought about refactoring a larger part. Can we not just implement one of the functions in terms of the other?

[larsmans]

If all the tests pass, +1 for merge.

[amueller]

ok,lets merge and refactor later.

Lars Buitinck notifications@github.com schrieb:

If all the tests pass, +1 for merge.

---

Reply to this email directly or view it on GitHub:
#1741 (comment)

Diese Nachricht wurde von meinem Android-Mobiltelefon mit K-9 Mail gesendet.

[ogrisel]

I pushed the green button. Hopefully jenkins will be happy.

[arjoly]

jenkins is happy now !!!

[arjoly]

I rather thought about refactoring a larger part. Can we not just implement one of the functions in terms of the other?

@amueller I though about it and decided to avoid memory copy each time the data aren't properly formatted.

[amueller]

Sorry, I don't think I understand the argument. Can you elaborate?

[arjoly]

For the moment, there is one implementation of the measure per format type (multiclass, indicator-matrix, list-of-labels). If you write one implementation for all format, you will need to perform data transformation each time you aren't in the proper format.

[amueller]

What I meant was to do #1748, not refactor the format types.