BUG: recent numpy fails on non-int shape

[perimosocordiae]

With the current numpy master branch, the WebAgg backend fails during the draw() method

 File "/usr/local/lib/python2.7/dist-packages/matplotlib-1.5.0rc2+149.g50f373f-py2.7-linux-x86_64.egg/matplotlib/backends/backend_webagg_core.py", line 287, in handle_event
    self.draw()
  File "/usr/local/lib/python2.7/dist-packages/matplotlib-1.5.0rc2+149.g50f373f-py2.7-linux-x86_64.egg/matplotlib/backends/backend_webagg_core.py", line 181, in draw
    self.manager.refresh_all()
  File "/usr/local/lib/python2.7/dist-packages/matplotlib-1.5.0rc2+149.g50f373f-py2.7-linux-x86_64.egg/matplotlib/backends/backend_webagg_core.py", line 473, in refresh_all
    diff = self.canvas.get_diff_image()
  File "/usr/local/lib/python2.7/dist-packages/matplotlib-1.5.0rc2+149.g50f373f-py2.7-linux-x86_64.egg/matplotlib/backends/backend_webagg_core.py", line 211, in get_diff_image
    buff.shape = (renderer.height, renderer.width)
TypeError: 'numpy.float64' object cannot be interpreted as an index

This patch remedies the problem, though not at its source. It might be useful to track down what causes the width and height to be float64 in the first place.

[perimosocordiae]

The Travis failures seem to be unrelated to this change.

[mdboom]

Thanks for noticing and providing a fix. I wonder if we could track down why renderer.width and renderer.height are floats in the first place. It's kind of nonsensical for them to be that.

[mdboom]

I'm remilestoning this -- as it doesn't affect released versions of numpy, I don't think we need to worry about it for 1.5.0

[efiring]

@mdboom, isn't this a time-bomb? If the root problem isn't apparent, I would be inclined to accept the patch with the addition of a comment as to its rationale. The requirement that indices be integers is already in effect; I see it in python 3.4 with numpy 1.10.1 at least.

[mdboom]

No, I think we need to get to the root of the problem, not pave it over like this. I can't reproduce on 1.5.x with Python 2.7/3.4 or 3.5 with Numpy 1.10.1. That's what lead me to postpone this, given that 1.5.0 is already way overdue. Can you reproduce it with a released Numpy?

[efiring]

I don't even know how to to try to reproduce it. I'm OK with your decision.

[mdboom]

I can't even get matplotlib to build with numpy master right now. @perimosocordiae: What was your last working Numpy git hash?

[perimosocordiae]

Here's a script that reproduces the error for me:

import matplotlib
matplotlib.use('webagg')
from matplotlib import pyplot as plt
import numpy as np

print 'numpy:', np.__version__
print 'MPL:', matplotlib.__version__

plt.figure()
plt.plot([0,1,0.5,2,1,3])
plt.show()

The versions reported are:

numpy: 1.11.0.dev0+e24895c
MPL: 1.5.0rc2+149.g50f373f

I built matplotlib using that numpy version without issues on Ubuntu 14.10 with Python 2.7.8.

[perimosocordiae]

I've tracked down where the floats are coming from: https://github.com/matplotlib/matplotlib/blob/master/lib/matplotlib/backends/backend_webagg_core.py#L251

The bbox object stores its data as numpy floats. I'll update the PR shortly.

[mdboom]

Great. Thanks for tracking this down.

[mdboom]

Backported to v1.5.x as 921aa41