fig.colorbar messes up subplots centering

[anntzer]

from pylab import *
fig, axs = plt.subplots(2)
axs[0].imshow(np.random.rand(32, 32), extent=(10, 0, 0, 10))
img = axs[1].imshow(np.random.rand(32, 32), extent=(10, 0, 0, 10))
fig.colorbar(img, ax=list(axs))
plt.show()

(click on the image to confirm that it is indeed mis-centered)

[WeatherGod]

Confirmed. And it happens regardless if use_gridspec is True or False.

On Wed, Nov 11, 2015 at 1:45 PM, Antony Lee notifications@github.com
wrote:

from pylab import *
fig, axs = plt.subplots(2)
axs[0].imshow(np.random.rand(32, 32), extent=(10, 0, 0, 10))
img = axs[1].imshow(np.random.rand(32, 32), extent=(10, 0, 0, 10))
fig.colorbar(img, ax=list(axs))
plt.show()

[image: colorbar]
https://cloud.githubusercontent.com/assets/1322974/11099602/27f4af18-8861-11e5-8fdc-9f0971a22d9a.png
(click on the image to confirm that it is indeed mis-centered)

—
Reply to this email directly or view it on GitHub
#5463.

[mdboom]

I'm milestoning this as 2.0, but because I think we can guarantee we can fix this for then, but because fixing this may require changing behavior in a backward-incompatible way.

[anntzer]

Note to self and whoever hits this bug: this is due to the fact that axes with a set aspect ratio (directly or via an image) only positioned at drawing time (why?...) but the colorbar is positioned when it's drawn, so the colorbar thinks it's there for the normal, "rectangular" axes and everything collapes.

A stopgap measure is to call fig.canvas.draw() before creating the colorbar (because that'll resolve the positions of the axes). In fact I wonder whether plt.colorbar should just do that... (at least in interactive mode, where that can't(?) be too bad except for a performance hit)

[efiring]

The automatic generation of the colorbar axes requires that it steal space from the main axes; setting the aspect ratio has to be done after this, hence at drawing time for the main axes. This is the sort of chicken and egg problem that could be solved better with a constraints engine. I don't think that triggering a full draw before creating the colorbar solves the problem; it just adds useless overhead.

[anntzer]

It's not perfect, but because triggering an initial draw brings the positions of the axes much closer to their final position, the final layout still looks much less unbalanced.

[efiring]

Sorry, I don't think the benefit is worth the cost. It's just the wrong solution. At the stage of doing interactive plots, there is no functional advantage to improving the centering in this way.

[tacaswell]

Another possible solution is to use a function like:

def squeeze_fig_aspect(fig, preserve='h'):
    preserve = preserve.lower()
    bb = bbase.union([ax.bbox for ax in fig.axes])

    w, h = fig.get_size_inches()
    if preserve == 'h':
        new_size = (h * bb.width / bb.height, h)
    elif preserve == 'w':
        new_size = (w, w * bb.height / bb.width)
    else:
        raise ValueError(
            'preserve must be "h" or "w", not {}'.format(preserve))
    fig.set_size_inches(new_size, forward=True)

after the first draw which re-sizes the figure to have the same aspect ratio as the bounding box around all of it's axes.

This probably needs some epicycles added like ignoring non-visible axes etc.

[anntzer]

Seems like a useful functionailty to have in general, but I'm not sure how this helps in this case?

[tacaswell]

And I am pretty sure that the fixed-aspect are positioned on the figure at draw time because the axes does not necessarily know the data-limits before then and the default behavior of a fixed aspect ratio axes is to resize the axes, not change the data limits, to get same ratio between data space and pixel space in x and y.

Subjectively, I think that the current behavior gives a better looking figure

import matplotlib.pyplot as plt
import numpy as np
from matplotlib.transforms import BboxBase as bbase

def squeeze_fig_aspect(fig, preserve='h'):
    preserve = preserve.lower()
    bb = bbase.union([ax.bbox for ax in fig.axes])

    w, h = fig.get_size_inches()
    if preserve == 'h':
        new_size = (h * bb.width / bb.height, h)
    elif preserve == 'w':
        new_size = (w, w * bb.height / bb.width)
    else:
        raise ValueError(
            'preserve must be "h" or "w", not {}'.format(preserve))
    fig.set_size_inches(new_size, forward=True)


fig, axs = plt.subplots(2)
axs[0].imshow(np.random.rand(32, 32), extent=(10, 0, 0, 10),
              cmap='viridis', interpolation='nearest')
img = axs[1].imshow(np.random.rand(32, 32), extent=(10, 0, 0, 10),
                    cmap='viridis', interpolation='nearest')

#fig.canvas.draw()
fig.colorbar(img, ax=list(axs))
squeeze_fig_aspect(fig)

[anntzer]

I agree that the lower layout is much better (well, perhaps that raises other issues about automatic ticking...) but I don't see the figure being resized, either when calling the script directly or via %run test.py in IPython with %matplotlib mode on (so everything is still off-center). Am I missing something?

[tacaswell]

Are you using inline?

[anntzer]

I wasn't. In %inline mode the issue is not there to start with...