Fix problem with legend if data has NaN's [backport to 1.4.x]

[demotu]

If data has NaN's and the plot a legend with parameter 'best', matplotlib throws the following RuntimeWarning:

...\matplotlib\transforms.py:651: RuntimeWarning: invalid value encountered in sign
dx0 = np.sign(vertices[:, 0] - x0)
...\matplotlib\transforms.py:652: RuntimeWarning: invalid value encountered in sign
dy0 = np.sign(vertices[:, 1] - y0)
...\matplotlib\transforms.py:653: RuntimeWarning: invalid value encountered in sign
dx1 = np.sign(vertices[:, 0] - x1)
...\matplotlib\transforms.py:654: RuntimeWarning: invalid value encountered in sign
dy1 = np.sign(vertices[:, 1] - y1)

The following code reproduces this RuntimeWarning:
plt.plot([1,2,np.nan,4,5], label='test')
plt.legend(loc='best')

The solution is to check for NaN's in data and delete those rows, as suggested.

[tacaswell]

This looks reasonable to me. Someone who understand the transform stack better than I do should take a look at this.

[mdboom]

Seems like a reasonable fix. It's a shame it has to copy all of the vertices (which can be quite large) to do it, though.

Maybe just use the Numpy error context manager to suppress the warning instead?

[demotu]

Yes, using the context manager to suppress the warning will be faster and will use less resources.

We can use:

np.seterr(invalid='ignore')
...

Or:

with np.errstate(invalid='ignore'):
    dx0 = np.sign(vertices[:, 0] - x0)
    dy0 = np.sign(vertices[:, 1] - y0)
    dx1 = np.sign(vertices[:, 0] - x1)
    dy1 = np.sign(vertices[:, 1] - y1)
...

Any preference?

[mdboom]

Let's use the context manager. Setting it globally will hide legitimate warnings in user code.

Thanks!

[demotu]

Ok, should I change the code and do the pull request?

[tacaswell]

Just make a new commit and add it to this branch. The PR tracks the branch from your repo, not the commit.

You can even do a force-push to your repo and expunge all record of the current commit.

[efiring]

The test failure on 2.7 is odd; it is not just the usual unrelated timeout or other Travis glitch.

[tacaswell]

I have noticed that one go by every so often and it is normally a travis glitch. Restarted the tests on 2.7.

[demotu]

it worked, thanks.

[efiring]

@tacaswell, would you do the backport, please?

[tacaswell]

back-ported as 86f9f9c