Combining multiple existing figures into one figure

Many higher level applications produce figures, typically when the figure itself is a composite of multiple axes, hence the option to pass an Axes instance is not available. Often the actual creation of the Figure is buried deeply in a call stack, making it difficult to intercept. I’ve seen the new subfigures feature of matplotlib, which is so new that limiting a search to the past one year finds only a handful of pages on the internet. I know there is no public API and design intention at the moment to convert a Figure to a SubFigure and include it in a main Figure. However, because this feature is so much desired and would be so useful for many, I thought it is worth to look a bit deeper whether it’s a low hanging fruit, or there are major obstacles. I show here just a little example of creating two figures and trying to make them subfigures. The to_subfig is more or less a copy of SubFigure.__init__.

import matplotlib as mpl
from matplotlib.transforms import BboxTransformTo
import numpy as np

def to_subfig(fig, sfig):

    fig.__class__ = mpl.figure.SubFigure
    fig._subplotspec = sfig._subplotspec
    fig._parent = sfig._parent
    fig.figure = sfig.figure
    # subfigures use the parent axstack
    fig._axstack = sfig._axstack
    fig.subplotpars = sfig.subplotpars
    fig.dpi_scale_trans = sfig.dpi_scale_trans
    fig._axobservers = sfig._axobservers
    fig.canvas = sfig.canvas
    fig.transFigure = sfig.transFigure
    fig.bbox_relative = None
    fig.figbbox = fig._parent.figbbox
    fig.bbox = sfig.bbox
    fig.transSubfigure = BboxTransformTo(fig.bbox)

    fig.patch = sfig.patch
    fig.patch.figure = fig

    return fig

fig1 = mpl.figure.Figure(constrained_layout = True)
ax1 = fig1.add_subplot()

fig2 = mpl.figure.Figure(constrained_layout = True)
ax2 = fig2.add_subplot()
x = np.arange(0, 20, .1)

mainfig = mpl.figure.Figure(constrained_layout = True, figsize = (10, 4))
subf = mainfig.subfigures(1, 2)
fig1 = to_subfig(fig1, subf[0])
fig2 = to_subfig(fig2, subf[1])
mainfig.subfigs = [fig1, fig2]

Looking at the output, both figures are drawn but at wrong position and scaling:

Despite the bboxes seem to be alright:

# Bbox([[0.0, 0.0], [0.5, 1.0]])
# Bbox([[0.5, 0.0], [1.0, 1.0]])

I am wondering if someone with more insight about matplotlib design might be able to tell what is missing here? Is it just a few steps away, or is it impossible?

We do not have an API to convert a Figure into a SubFigure, but anyplace you can pass in a Figure you should be able to pass in a SubFigure instance and do your composition that way.

The core of the problem with post-facto combining of Figures / Axes is that in the Figure → Axes → DataArtist trees there are shared Transform objects (which form their own tree) that manage the translation from the (many) notional coordinate systems Matplotlib maintains (“data”, “axes”, “figure”, the blended transfrom, …) to “screen” (which can be pixels or physical units depending on the back end). Additionally the Artists also hold references to their (grand) parents in the tree.

While it is theoretically possible to swap opt the transform at the top and update all of the references, I am not confident it will “just work”. In a couple of places we explicitly prevent moving objects between unrelated Figures because we thought it would be better to raise than to give the user “working” (in that the code runs without error) but very very broken output (because something was inconsistent and violated the internal assumptions/in-variants).

It in my estimate that it is probably 10-20 hours of work (and as with all software estimates multiply by π) to chase through and sort out if it is possible and a similar amount of time to test and document it well enough.

1 Like

Thanks a lot @tacaswell, your answer gives valuable insight. In my opinion, 10-20h of work would well worth it, even a few days or weeks of work, as this could become a very powerful feature. I am not too familiar with the internals of matplotlib, but maybe at some point I will try to look into it.