It would be great to get @leejoon's thoughts on this, as the legend guy.

I like this because it brings Figure.legend in line with Axes.legend in terms of what it can do.

Should we reconsider the behavior of pyplot.legend()? IIRC, it operates upon gca(), not gcf(). I am inclined to keep it as gca(), but I just wanted to make sure a decision is made on it.

Doesn't pyplot have figlegend which is the equivalent of this in pyplot?

Yeah -- I think gca().legend() is the more common case (or at least that's how I see it after all these years of expecting that).

And as @jenshnielsen points out, we have pyplot.figlegend to get to Figure.legend(). And it has the same shortcoming -- that lines and labels are required as arguments.

Thank you for the confirmation. I've applied the same modification to pyplot.figlegend(). Also, the Figure.legend() can now work for 0, 2, and 3 arguments, because pyplot.figlegend() is intended to be called with handler, labels and loc without keywords.

I think it is good to merge.

A few minor points.

1. I am personally not quite convinced for the current behavior for identical labels.
2. The ordering of handles and labels rely on the order of dict.keys() which is arbitrary.

Unless there is no objection of merging this, I will go ahead and merge it.

Why not use OrderedDict from the standard library's collection module?
Note that you can't just simply import collections because that would pick
up matplotlib's collection module, so you would have to import from future
the absolute imports feature.

Unless there is no objection of merging this, I will go ahead and merge it.

I'm currently working on a re-factor of the legend code. It deprecates some of the plt.legend interface and I'd like to have the discussion about that interface before we merge this. I'm hoping to have a PR in the next couple of days and would really value your feedback on it.

Cheers,

@WeatherGod OrderedDict is new in python2.7.

@mrterry yeah, so? We support python2.6. That particular line should really be doing a single call to items() and splitting things out that way, but you do have to watch out for the empty dictionary case.

Just pointing out that it is not present in one of our supported python versions. I wanted to use it in setup.py, but got shot down. 😦 I really like OrderedDict.

This having issues with pyparsing on 3.3.

@sinhrks Are you still interested in working on this PR?

pushing off to 1.5 due to issues with py3k and in-activity.

Thank you for your comment and I apologize that I hadn't check it long.

I've changed the internal logic not to use dict, but use list to retain the drawing order.

About Legend Labels: My original intention is to create a figure legend easily in subplots which uses similar data, same type of chart and color schema. In such a case, "looks-identical" artists should have the same label, artist, colors, etc. Thus I'd like to ask on what condition 2 artist can be regarded as identical (or looks-similar enough to be omitted in figure legend) in most cases. In other detailed cases, handles and labels can be passed to legend explicitly.

I've implemented a proto which also check artist type, color, edgecolor and facecolor are identical before skipping. I'll try to implement it once what attributes are checked.

Current Outlook:

It looks like you have introduced a bunch of pep8 violations into the code.

@sinhrks Are you still interested in working on this?

@sinhrks This needs a rebase.

Does it make sense to group artist by label? It might be worth getting the simpler change (making the API match that of ax.legend) in and doing the smarter grouping in a second pass?

Since it's been over 2 years since last contact, I'll pick this up. Merged and rebased onto my own branch, just getting tests to run etc. and I'll open a new PR when ready.

I am closing this in favor of #7811