[MNT]: FigureCanvasSVG has no attribute .get_renderer

[raphaelquast]

Summary

Hey,

while investigating possibilities for svg exports with EOmaps I realized that the FigureCanvasSVG class does not expose the renderer via a .get_renderer() method but initializes the renderer in .pring_svg().

matplotlib/lib/matplotlib/backends/backend_svg.py

Lines 1330 to 1338 in 2f778fd

	dpi = self.figure.dpi
	self.figure.dpi = 72
	width, height = self.figure.get_size_inches()
	w, h = width * 72, height * 72
	renderer = MixedModeRenderer(
	self.figure, width, height, dpi,
	RendererSVG(w, h, fh, image_dpi=dpi, metadata=metadata),
	bbox_inches_restore=bbox_inches_restore)
	self.figure.draw(renderer)

Is this intentional or is this a API inconsistency that should be fixed?

• The same is also true for the pdf backend

Proposed fix

Implement .get_renderer() similar to the agg backend:

matplotlib/lib/matplotlib/backends/backend_agg.py

Lines 406 to 413 in 2f778fd

	def get_renderer(self):
	w, h = self.figure.bbox.size
	key = w, h, self.figure.dpi
	reuse_renderer = (self._lastKey == key)
	if not reuse_renderer:
	self.renderer = RendererAgg(w, h, self.figure.dpi)
	self._lastKey = key
	return self.renderer

[anntzer]

The implementation is currently private at Figure._get_renderer/backend_bases._get_renderer. It would be nice to decide once and for all where the public API should be -- I think putting that implementation as FigureCanvasBase.get_renderer, letting canvas subclasses override that with their normal implementation if needed (which is already the case e.g. for agg), would be normal.

(See also #13099, though.)

[jklymak]

It's might help decide the right way forward to explain why one would need/want the renderer? Usually folks just switch to svg to print.

[tacaswell]

get_renderer is a feature of the Agg backend family not of the base.

My recollection is that the vector backends stream their output so outside of the save code it is not clear to me that we can even give the user a Renderer is a usable state.

I would have to check the history, but I suspect get_renderer was added to support blitting which really only makes sense with the raster (Agg derived) backends.

[raphaelquast]

Hey, thanks a lot for all the responses!!

In general I agree with @tacaswell that this primarily makes sense for blitting (which is exactly where I ran into this) and in 99% of the use-cases this will concern Agg derived backends...

Now, to answer the question of @jklymak on why one would need the renderer in the first place...
In EOmaps I somewhat abuse the blitting capabilities to create "multi-layered figures" to delay/avoid draws of very large datasets etc.

To do this, I handle drawing artists myself which is currently implemented (very roughly) like this:

import matplotlib.pyplot as plt
f, ax = plt.subplots()
ax.set_animated(True)
l, = ax.plot([0, 0.5, 1], [0, 0.5, 1], lw=20, animated=True)

def drawit(*args, **kwargs):
    renderer = f.canvas.get_renderer()
    
    l.set_color(plt.cm.prism(min(ax.get_xlim())))
    l.draw(renderer)
    for _, s in ax.spines.items():
        s.draw(renderer)

f.canvas.mpl_connect("draw_event", drawit)

This worked nicely for all my needs so far (including png, jpeg... export) but it fails on svg and pdf exports since there is no .get_renderer() for the svg/pdf canvas.

It might be that there is an entirely better way of implementing all of this (any insights are highly welcome!!) ... a general use-case concerning this issue would be:

• how to trigger drawing an individual artist in the svg/pdf backend without relying on methods from the private API? (ax.draw_artist uses get_renderer renderer as well)

[tacaswell]

When saving the animated attribute is ignored and all of the artists are re-drawn from scratch:

matplotlib/lib/matplotlib/axes/_base.py

Lines 3031 to 3034 in d464cbf

	if not self.figure.canvas.is_saving():
	artists = [
	a for a in artists
	if not a.get_animated() or isinstance(a, mimage.AxesImage)]

There is still the timing issue of when that will report true / false, when the draw_event callbacks fire, and when the backends are swapped back, but I suspect there is a working path through.

[raphaelquast]

@tacaswell
thanks! this is why i explicitly use ax.set_animated(True) right now to completely avoid triggering the axes-draw cycle.
(I do this so that I can selectively cache backgrounds, avoid re-drawing complex artists, allow interactive comparison of cached layers and to have a fast-export of figures at the current-dpi)

Since for "svg/pdf" exports cached (agg-based) backgrounds are anyways irrelevant, a painless way to fix "svg/pdf" export would be to temporarily use ax.set_animated(False) to re-enable the ordinary axes draw cycle (and make sure only relevant artists are visible).... thanks for triggering this idea!

If there is in fact no way to ensure a meaningful object is returned by get_renderer for svg/pdf backends, feel free to close this issue if you think there's no point in discussing this further!

[tacaswell]

But if you are in fig.savefig we ignore the state of a.get_animated() (unless you are doing really wild things....)

[raphaelquast]

yes, but only for children of the axes in the ax.draw() call. However, if the whole axes is set to animated, ax.draw() is never triggered in the first place and so no artists (animated or not) of the axes are drawn at all. (not sure this is intentional, but that's what I use so far to avoid the default draw-cycle)

[tacaswell]

Oh, that is a bug:

matplotlib/lib/matplotlib/figure.py

Lines 214 to 216 in d464cbf

	artists = sorted(
	(artist for artist in artists if not artist.get_animated()),
	key=lambda artist: artist.get_zorder())

That should be following the same rules as the filtering inside of Axes.

[raphaelquast]

ok... that's unfortunate... 😅
please ping me in case you tackle this since it would have quite an affect on some of EOmaps functionalities
(I'll see to get this out of my code asap).

At the moment this was the most easy way I found to completely avoid the normal draw-cycle of an axes... is there a better way (preferably without subclassing Axes or Figure)?