[Bug] inset_axes fails under inline backend with bbox_inches=tight (renderer None)

[rowan-stein]

User Request

[Bug]: Error while creating inset axes using mpl_toolkits.axes_grid1.inset_locator.inset_axes

Bug summary:
Unable to create the inset axes in a plot using the code (following the first example on the website as posted here).

Code for reproduction:

import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1.inset_locator import inset_axes

fig, (ax, ax2) = plt.subplots(1, 2, figsize=[5.5, 2.8])
axins = inset_axes(ax, width=1.3, height=0.9)
plt.show()

Actual outcome (stack trace excerpt):

AttributeError: 'NoneType' object has no attribute '_get_renderer'
... at matplotlib/offsetbox.py:399 in OffsetBox.get_window_extent
... called from mpl_toolkits/axes_grid1/inset_locator.py:73 in AnchoredLocatorBase.__call__
... triggered via matplotlib/_tight_bbox.py during inline backend bbox_inches='tight'

Expected outcome:
An empty box towards the top right of the first subplot (with axes ax), as shown in the demo.

Environment:

• OS: Arch linux: 6.4.2-arch1-1
• Matplotlib Version: 3.7.2
• Backend: module://matplotlib_inline.backend_inline
• Python: 3.8.17
• Jupyter lab: 3.6.5
• Installation: conda

Notes:

• Inline backend applies bbox="tight" to figure display.
• The axes created by axes_grid1.insetlocator.inset_axes are not compatible with tight layout; however, get_window_extent should still work.

Task reference: ID 284

---

Researcher Specification (Emerson Gray)

Overview:
Creating inset axes via mpl_toolkits.axes_grid1.inset_locator.inset_axes fails under IPython inline backend when bbox="tight" is applied implicitly. The failure occurs because AnchoredLocatorBase.__call__ forwards renderer=None to OffsetBox.get_window_extent, which then tries to use self.figure._get_renderer(). Since the locator object is not attached to a figure (self.figure is None), this raises AttributeError.

Minimal fix: Ensure AnchoredLocatorBase.__call__ obtains a valid renderer from the parent Axes’ figure when the passed renderer is None.

Files and exact change:

• Target file: lib/mpl_toolkits/axes_grid1/inset_locator.py

Patch diff chunk:

*** Begin Patch
*** Update File: lib/mpl_toolkits/axes_grid1/inset_locator.py
@@
 class AnchoredLocatorBase(AnchoredOffsetbox):
@@
     def draw(self, renderer):
         raise RuntimeError("No draw method should be called")
 
     def __call__(self, ax, renderer):
+        # Some callers (e.g., tight bbox adjustment or inset axes setup)
+        # invoke locators with renderer=None. OffsetBox.get_window_extent
+        # will fallback to self.figure._get_renderer() if renderer is None,
+        # but this locator is not attached to a Figure (self.figure is None).
+        # Ensure a valid renderer is used.
+        if renderer is None:
+            renderer = ax.figure._get_renderer()
         self.axes = ax
         bbox = self.get_window_extent(renderer)
         px, py = self.get_offset(bbox.width, bbox.height, 0, 0, renderer)
         bbox_canvas = Bbox.from_bounds(px, py, bbox.width, bbox.height)
         tr = ax.figure.transSubfigure.inverted()
         return TransformedBbox(bbox_canvas, tr)
*** End Patch

Compatibility and behavior:

• Inline backend with implicit bbox_inches='tight': fixed.
• Non-inline backends and non-tight usage: unaffected.
• Other axes_grid1 locators benefit similarly.
• We do not attach the locator to the figure; passing a valid renderer is sufficient.

Testing plan:

• Manual repro (pre-fix failure, post-fix success):

import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1.inset_locator import inset_axes
fig, (ax, ax2) = plt.subplots(1, 2, figsize=(5.5, 2.8))
axins = inset_axes(ax, width=1.3, height=0.9)
fig.savefig('out.png', bbox_inches='tight')

• Unit test proposal: Add a smoketest to lib/mpl_toolkits/axes_grid1/tests/test_axes_grid1.py that creates an inset axes and saves with bbox_inches='tight' to ensure no exception.

Edge cases:

• tight_layout: existing warnings unchanged; fix prevents error.
• constrained_layout: unaffected; the renderer derivation is safe.

Reason:
Prevent OffsetBox.get_window_extent from using self.figure when the locator is not attached to a figure, avoiding AttributeError during inline backend tight bbox processing.