MNT: when clearing an Axes via clear/cla fully detach children #24627
Conversation
Reset the Axes and Figure of the children to None to help break cycles. Closes matplotlib#6982
c68d4de to
ffcc8d3
Compare
| self._children = [] | ||
| for chld in old_children: |
There was a problem hiding this comment.
Is there a reason for not doing for chld in getattr(self, "_children", []): ... before resetting _children?
There was a problem hiding this comment.
I am more worried about the state where the Axes thinks the Artist is in the draw list rather than the case where the Artist still thinks it is in the Axes (which has been the status quo).
My first attempt was
old_children, self._children = self._children, []
but that had issues where either you check self._children before it exists or you get other attribute errors because other parts of the code are using the existence of self._children to determine if we are still in the init method or not.
I also thought about doing for chld in self._chidren: chld.remove() but was trying to avoid a lot of thrashing / maybe some premature optimization.
There was a problem hiding this comment.
not sure if compressing to one line is better or worse ....
There was a problem hiding this comment.
Perhaps a good idea to first get rid of this "not having _children" problem by pre-initing it early in __init__? e.g. something like
diff --git i/lib/matplotlib/axes/_base.py w/lib/matplotlib/axes/_base.py
index 3e665cda42..9eb65d3fc4 100644
--- i/lib/matplotlib/axes/_base.py
+++ w/lib/matplotlib/axes/_base.py
@@ -665,6 +665,8 @@ class _AxesBase(martist.Artist):
self.set_box_aspect(box_aspect)
self._axes_locator = None # Optionally set via update(kwargs).
+ self._children = []
+
# placeholder for any colorbars added that use this Axes.
# (see colorbar.py):
self._colorbars = []
@@ -1078,7 +1080,7 @@ class _AxesBase(martist.Artist):
self.transScale.set(
mtransforms.blended_transform_factory(
self.xaxis.get_transform(), self.yaxis.get_transform()))
- for line in getattr(self, "_children", []): # Not set during init.
+ for line in self._children:
if not isinstance(line, mlines.Line2D):
continue
try:
@@ -2936,22 +2938,15 @@ class _AxesBase(martist.Artist):
x_stickies = y_stickies = np.array([])
if self.use_sticky_edges:
- # Only iterate over Axes and artists if needed. The check for
- # ``hasattr(ax, "_children")`` is necessary because this can be
- # called very early in the Axes init process (e.g., for twin Axes)
- # when these attributes don't even exist yet, in which case
- # `get_children` would raise an AttributeError.
if self._xmargin and scalex and self.get_autoscalex_on():
x_stickies = np.sort(np.concatenate([
artist.sticky_edges.x
for ax in self._shared_axes["x"].get_siblings(self)
- if hasattr(ax, "_children")
for artist in ax.get_children()]))
if self._ymargin and scaley and self.get_autoscaley_on():
y_stickies = np.sort(np.concatenate([
artist.sticky_edges.y
for ax in self._shared_axes["y"].get_siblings(self)
- if hasattr(ax, "_children")
for artist in ax.get_children()]))
if self.get_xscale() == 'log':
x_stickies = x_stickies[x_stickies > 0]There was a problem hiding this comment.
That will lead to AttributeErrors related to the Axes title attributes (and that was the point where I stopped debugging and used getattr) 🤣
I think the choice is either accept getattr or start pulling this string and see what we have to unravel which might spin out to "re-do all of the init logic in the Axes family".
|
Does this need some refactor now after #24634? |
522cc01 to
146b478
Compare
|
dropped the |
Co-authored-by: Elliott Sales de Andrade <[email protected]>
Reset the Axes and Figure of the children to None to help break cycles.
Closes #6982
PR Summary
PR Checklist
Documentation and Tests
pytestpasses)I do not think this needs any additional documentation.