Remove internal use of `Artist.figure` #28550

rcomer · 2024-07-12T13:17:45Z

PR summary

The Artist.figure attribute is inconsistent: on most artists, it is the figure the artist is attached to, which may be a SubFigure. OTOH SubFigure.figure is always the root Figure.

A lot of the internal usage of this attribute was written before subfigures existed, so there would not have been any decision to make about which figure is required. This means that in places we could be getting hold of a SubFigure when we really need a Figure.

At #28177 (comment), @timhoffm said

Internally, we should not be using figure anymore.

so here is my attempt to remove that internal usage. If I comment out the figure property in both figure.py and artist.py, the tests still pass locally (except the one that explicitly tests the property).

`get_figure` parameter choices

Wherever I use get_figure in the main code, I made a choice about the root parameter. In many places, the end result will be the same whether root is True or False:

artist.get_figure(root=False) will be None if and only if artist.get_figure(root=True) is None

In many cases we are just accessing an attribute on the figure, which for a subfigure is the same as its parent, and therefore the same as the root.

matplotlib/lib/matplotlib/figure.py

Lines 2290 to 2297 in 830361d

    
           self._axstack = parent._axstack 
        
           self.subplotpars = parent.subplotpars 
        
           self.dpi_scale_trans = parent.dpi_scale_trans 
        
           self._axobservers = parent._axobservers 
        
           self.transFigure = parent.transFigure 
        
           self.bbox_relative = Bbox.null() 
        
           self._redo_transform_rel_fig() 
        
           self.figbbox = self._parent.figbbox

or accessing a property which returns the equivalent property of the parent

matplotlib/lib/matplotlib/figure.py

Lines 2310 to 2316 in 830361d

    
           @property 
        
           def canvas(self): 
        
               return self._parent.canvas 
        
           @property 
        
           def dpi(self): 
        
               return self._parent.dpi

or calling _get_renderer which calls _get_renderer on the parent

matplotlib/lib/matplotlib/figure.py

Lines 2339 to 2340 in 830361d

def _get_renderer(self):

return self._parent._get_renderer()

For the above cases, I set root=False as that avoids the call to _parent_figure.get_figure within Artist.get_figure, so seems more efficient for most cases.

matplotlib/lib/matplotlib/artist.py

Lines 733 to 736 in 830361d

    
           if root and self._parent_figure is not None: 
        
               return self._parent_figure.get_figure(root=True) 
        
           return self._parent_figure

For tests that do not involve subfigures, I left the root parameter out, because there is no ambiguity about what we are getting hold of and it would seem like unnecessary clutter.

PR checklist

[n/a] "closes #0000" is in the body of the PR description to link the related issue
new and changed code is tested
[n/a] Plotting related features are demonstrated in an example
[n/a] New Features and API Changes are noted with a directive and release note
[n/a] Documentation complies with general and docstring guidelines

rcomer

Added some comment about what I think are less obvious choices.

rcomer · 2024-07-12T13:21:53Z

lib/matplotlib/artist.py

-            if (renderer._rasterizing and artist.figure and
-                    artist.figure.suppressComposite):
+            if (renderer._rasterizing and (fig := artist.get_figure(root=True)) and
+                    fig.suppressComposite):


I have not really understood what it does, but suppressComposite appears to only exist on Figure.

In vector outputs if you have multiple images or things you are pre-rasterizing you may want to rasterize them all into one image in the final output (which is good because it saves space if they overlap, the pathological example is 100 4kx4k images with alpha that are all exactly on top of each other, by pre-merging them you get a 100x size win!). On the other hand, if you have a bunch of images that only slightly overlap or you want to be able to re-extract the individual layers later you do not want to merge them. As both are reasonable we have an (esoteric) flag to control it 🤣 .

Thank you for the explanation.

rcomer · 2024-07-12T13:37:01Z

lib/matplotlib/collections.py

+            if (extents.width < self.get_figure(root=True).bbox.width
+                    and extents.height < self.get_figure(root=True).bbox.height):


I think this should be root figure because it is about what is possible for the renderer, but I am definitely out of my depth here.

I'm also not sure why we are looking at the total size not what it might be clipped to, but it has been like this from 2014 so 👍 on leaving it unless we understand why it should be changed to the subfigure.

rcomer · 2024-07-12T13:39:41Z

lib/matplotlib/contour.py

            mpl._blocking_input.blocking_input_loop(
-                self.axes.figure, ["button_press_event", "key_press_event"],
+                self.axes.get_figure(root=True),


blocking_input_loop calls figure.show and subfigures do not have show - so this needs to be the root Figure I think.

rcomer · 2024-07-12T13:44:49Z

lib/matplotlib/image.py

        clip = ((self.get_clip_box() or self.axes.bbox) if self.get_clip_on()
-                else self.figure.bbox)
+                else self.get_figure(root=True).bbox)


I think this is saying clip the image to the figure if it's not clipped to anything else. I think that implies it should be the root Figure: if the user wants to clip to a subfigure they can presumably pass the subfigure's bbox in set_clip_box.

rcomer · 2024-07-12T13:47:18Z

lib/matplotlib/image.py

@@ -1385,14 +1385,15 @@ def get_extent(self):

    def make_image(self, renderer, magnification=1.0, unsampled=False):
        # docstring inherited
-        fac = renderer.dpi/self.figure.dpi
+        fig = self.get_figure(root=True)


I'm not too sure about this one. The use of get_size_inches below means that this needs to be Figure, but it's not obvious to me that this code could ever involve a subfigure: I think it's only called within figimage and there is no SubFigure.figimage.

as noted above, the subfigures .dpi is a property that gets the root figures dpi so this is fine.

rcomer · 2024-07-12T14:00:25Z

lib/matplotlib/text.py

            to match regions with a figure saved with a custom dpi value.
        """
        if not self.get_visible():
            return Bbox.unit()
+
+        fig = self.get_figure(root=True)


I think this needs to be the root Figure since below within the context manager we set its dpi and calculate display coordinate for the empty text string.

I think anything with dpi we should be hitting the root figure.

rcomer · 2024-07-12T14:07:37Z

lib/mpl_toolkits/axes_grid1/axes_grid.py

@@ -410,7 +410,7 @@ def _init_locators(self):
                self._colorbar_pad = self._vert_pad_size.fixed_size
        self.cbar_axes = [
            _cbaraxes_class_factory(self._defaultAxesClass)(
-                self.axes_all[0].figure, self._divider.get_position(),
+                self.axes_all[0].get_figure(root=False), self._divider.get_position(),


I am completely lost in here...

Agree, but working at the sub-figure level seems correct....

rcomer · 2024-07-12T14:13:20Z

lib/mpl_toolkits/mplot3d/tests/test_axes3d.py

@@ -667,8 +667,6 @@ def test_surface3d_label_offset_tick_position():
    ax.set_ylabel("Y label")
    ax.set_zlabel("Z label")

-    ax.figure.canvas.draw()


Think this was redundant since it's an image comparison test, so the draw will happen anyway.

lib/matplotlib/axes/_base.py

ksunden · 2024-07-14T18:05:48Z

lib/matplotlib/axes/_base.py

@@ -4754,6 +4759,9 @@ def __init__(self, figure, artists):
            self.figure = figure
            self.artists = artists

+        def get_figure(self, **kwargs):


Do we want to do **kwargs here?

The version elsewhere will also accept root as a positional arg, so it is not fully compatible with that signature (of course, it isn't actually changing the behavior here)

That said, this is an internal class inside of of a private function, which is only called in one place, and I'm not even convinced calls get_figure ever... so perhaps not worth being overly prescriptive here.

Hmmm. If we remove this method then this test
lib/matplotlib/tests/test_axes.py::test_zorder_and_explicit_rasterization
Fails here

matplotlib/lib/matplotlib/artist.py

Lines 78 to 79 in 7c70bea

if (renderer._rasterizing and (fig := artist.get_figure(root=True)) and

fig.suppressComposite):

Since that uses the root figure, I think we should also be getting the root figure here

matplotlib/lib/matplotlib/axes/_base.py

Lines 3149 to 3150 in 7c70bea

if artists_rasterized:

_draw_rasterized(self.get_figure(root=False), artists_rasterized, renderer)

As for the **kwargs, I don't really remember why I chose that, so happy to switch it to just the root keyword.

ksunden

Take or leave the minor comment about the private/internal method signature, I don't actually think it is likely to be problematic

rcomer · 2024-07-15T17:41:48Z

Rebased to fix Circle, which was trying and failing to build with python3.9. It must be looking at the branch for the environment spec but the merged result for the mpl install requirement.

tacaswell · 2024-07-18T02:35:09Z

It must be looking at the branch for the environment spec but the merged result for the mpl install requirement.

Yes, circle builds on the PR branch not the proposed merge (like most other CI services) so we do the merge in our setup steps. Mostly works, but cases like this are broken exactly the way you describe.

lib/matplotlib/axes/_base.py

tacaswell · 2024-07-19T18:58:33Z

lib/matplotlib/collections.py

@@ -1001,7 +1001,7 @@ def set_sizes(self, sizes, dpi=72.0):

    @artist.allow_rasterization
    def draw(self, renderer):
-        self.set_sizes(self._sizes, self.figure.dpi)
+        self.set_sizes(self._sizes, self.get_figure(root=False).dpi)


Subfigures ask their parent for the dpi anyway so just go straight there and save a property call.

Suggested change

self.set_sizes(self._sizes, self.get_figure(root=False).dpi)

self.set_sizes(self._sizes, self.get_figure(root=True).dpi)

tacaswell · 2024-07-19T18:58:42Z

lib/matplotlib/collections.py

@@ -1310,7 +1310,7 @@ def get_rotation(self):

    @artist.allow_rasterization
    def draw(self, renderer):
-        self.set_sizes(self._sizes, self.figure.dpi)
+        self.set_sizes(self._sizes, self.get_figure(root=False).dpi)


Suggested change

self.set_sizes(self._sizes, self.get_figure(root=False).dpi)

self.set_sizes(self._sizes, self.get_figure(root=True).dpi)

tacaswell · 2024-07-19T19:06:56Z

lib/matplotlib/lines.py

@@ -467,11 +467,12 @@ def contains(self, mouseevent):
        yt = xy[:, 1]

        # Convert pick radius from points to pixels
-        if self.figure is None:
+        fig = self.get_figure(root=False)


Suggested change

fig = self.get_figure(root=False)

fig = self.get_figure(root=True)

for the same reason as the rest of the DPI related comments.

tacaswell · 2024-07-19T19:07:09Z

lib/matplotlib/lines.py

@@ -640,7 +641,7 @@ def get_window_extent(self, renderer=None):
                                 ignore=True)
        # correct for marker size, if any
        if self._marker:
-            ms = (self._markersize / 72.0 * self.figure.dpi) * 0.5
+            ms = (self._markersize / 72.0 * self.get_figure(root=False).dpi) * 0.5


Suggested change

ms = (self._markersize / 72.0 * self.get_figure(root=False).dpi) * 0.5

ms = (self._markersize / 72.0 * self.get_figure(root=True).dpi) * 0.5

tacaswell · 2024-07-19T19:07:48Z

lib/matplotlib/lines.py

@@ -1648,7 +1649,7 @@ def __init__(self, line):
            'pick_event', self.onpick)
        self.ind = set()

-    canvas = property(lambda self: self.axes.figure.canvas)
+    canvas = property(lambda self: self.axes.get_figure(root=False).canvas)


Suggested change

canvas = property(lambda self: self.axes.get_figure(root=False).canvas)

canvas = property(lambda self: self.axes.get_figure(root=True).canvas)

Similarly to dpi, canvas always falls back to the root.

lib/matplotlib/axis.py

tacaswell · 2024-07-19T19:39:43Z

lib/matplotlib/axis.py

        else:
            # Union with extents of the top spine if present, of the axes otherwise.
            bbox = mtransforms.Bbox.union([
                *bboxes2, self.axes.spines.get("top", self.axes).get_window_extent()])
-            self.label.set_position((x, bbox.y1 + self.labelpad * self.figure.dpi / 72))
+            self.label.set_position(
+                (x, bbox.y1 + self.labelpad * self.get_figure(root=False).dpi / 72))


Suggested change

(x, bbox.y1 + self.labelpad * self.get_figure(root=False).dpi / 72))

(x, bbox.y1 + self.labelpad * self.get_figure(root=True).dpi / 72))

tacaswell · 2024-07-19T19:40:01Z

lib/matplotlib/axis.py

@@ -2429,14 +2432,14 @@ def _update_offset_text_position(self, bboxes, bboxes2):
            else:
                bbox = mtransforms.Bbox.union(bboxes)
                bottom = bbox.y0
-            y = bottom - self.OFFSETTEXTPAD * self.figure.dpi / 72
+            y = bottom - self.OFFSETTEXTPAD * self.get_figure(root=False).dpi / 72


Suggested change

y = bottom - self.OFFSETTEXTPAD * self.get_figure(root=False).dpi / 72

y = bottom - self.OFFSETTEXTPAD * self.get_figure(root=True).dpi / 72

tacaswell · 2024-07-19T19:40:16Z

lib/matplotlib/axis.py

        else:
            if not len(bboxes2):
                top = self.axes.bbox.ymax
            else:
                bbox = mtransforms.Bbox.union(bboxes2)
                top = bbox.y1
-            y = top + self.OFFSETTEXTPAD * self.figure.dpi / 72
+            y = top + self.OFFSETTEXTPAD * self.get_figure(root=False).dpi / 72


Suggested change

y = top + self.OFFSETTEXTPAD * self.get_figure(root=False).dpi / 72

y = top + self.OFFSETTEXTPAD * self.get_figure(root=True).dpi / 72

lib/matplotlib/axis.py

lib/matplotlib/backend_bases.py

lib/matplotlib/artist.py

tacaswell · 2024-07-19T19:50:51Z

Sigh, sorry about the far too many comments. I thought it would not be too many to do and then fell into the sunk-cost fallacy of being "almost done" (I was not) until I got to the point where I realized it might be easier for Ruth to do find-and-replace herself than to click "accept" a zillion times an finally gave up...

For dpi and canvas these things only make sense to be defined on the root figure, but to make the transition easier when we added SubFigure (to essentially avoid doing this work!) we added properties that forwarded to the parent. So in any case where you are accessing canvas, dpi, or _get_renderer (and a couple of the transforms that we hold a shared reference to in both the sub figure and the parent but I left those alone) it is not wrong to get it from the subfigure but going all the way to the root makes it a bit clearer to the future reader that this is a "figure global" type thing and is marginally faster to call (by skipping at least one property lookup).

I'm 👍 on this merging with or without my wall of suggestions going in.

tacaswell · 2024-07-19T19:53:02Z

@rcomer or anyone can merge at their discretion.

rcomer · 2024-07-19T20:11:41Z

Thanks @tacaswell!

My thinking for favouring root=False was that it should be more efficient in the case when there are no subfigures involved. E.g.

$ ipython
Python 3.9.19 | packaged by conda-forge | (main, Mar 20 2024, 12:50:21) 
Type 'copyright', 'credits' or 'license' for more information
IPython 8.18.1 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import matplotlib.pyplot as plt

In [2]: l, = plt.plot([0, 1])

In [3]: %timeit l.get_figure(root=False)
68.8 ns ± 0.657 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

In [4]: %timeit l.get_figure(root=True)
156 ns ± 1.9 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

But maybe these numbers are too small to worry about, or maybe there's a change that could make root=True more efficient when we're already at the root Figure...

I think I need to think about this again when I'm less tired.

rcomer · 2024-07-20T08:17:46Z

Pro tip: if you use GitHub's "Add suggestion to batch" feature, do not navigate away from the page before you commit the batch the batch will be lost and you have to add all the suggestions again 😢

rcomer · 2024-07-20T11:29:14Z

I decided to take @tacaswell's suggestions as it just smells better to get those Figure-specific properties directly from the Figure. Latest push has some more I found locally. I was hoping to remove SubFigure._get_renderer() but it's still used in at least one place. I do not know how to reliably find all the places it's used because we obviously are not testing every possible combination and I can only get so far with grep.

Co-authored-by: Thomas A Caswell <[email protected]>

rcomer · 2024-07-20T13:02:42Z

Rebased to pick up #28597.

tacaswell · 2024-07-22T13:54:51Z

@rcomer I'm sorry, my review on this appears to have been done in the optimally bad way for the tool!

rcomer · 2024-07-26T10:40:41Z

Although I had permission to self-merge this one, I'm not sure I should because I made a bunch more changes after accepting Tom's suggestions. The pyplot changes are marginally less trivial than the rest.

tacaswell added this to the v3.10.0 milestone Jul 12, 2024

rcomer commented Jul 12, 2024

View reviewed changes

rcomer force-pushed the remove-figure-use branch from 1610045 to 7c70bea Compare July 12, 2024 14:41

ksunden reviewed Jul 14, 2024

View reviewed changes

ksunden approved these changes Jul 14, 2024

View reviewed changes

rcomer force-pushed the remove-figure-use branch 2 times, most recently from 7b496fa to 8fc4464 Compare July 15, 2024 17:37