Repository navigation
backend partially ignores "figure.figsize" in plt.rc_context in jupyter #23513
Description
Bug summary
This possibly related to #20738
When I run the code below on a new jupyter kernel, only the tick styling is working. The figure.size and figure.dpi is ignored first.
One has to run something interacting with the backend first (like matplotlib.pyplot.show() ) in order to get plt.rc_context() working correctly.
Is this normal behavior, since the correct backend already seems to be activated or did I oversee something?
Code for reproduction
import matplotlib.pyplot as plt
rc = {"figure.figsize": [6, 1], "figure.dpi": 150, "xtick.direction": "in", "ytick.direction": "in"}
# Uncommenting this helps, but is this the intended behavoiur in jupyter?
# plt.show()
with plt.rc_context(rc):
# Sanity check.
print(f"{plt.rcParams['figure.figsize']=}")
print(f"{plt.rcParams['figure.dpi']=}")
print(f"{plt.rcParams['backend']=}")
fig, ax = plt.subplots()
# ax.plot(x, x)
plt.show()Actual outcome
plt.rcParams['figure.figsize']=[6.0, 1.0]
plt.rcParams['figure.dpi']=150.0
plt.rcParams['backend']='module://matplotlib_inline.backend_inline'
Expected outcome
Runnig the code twice, which has then plt.show() already called once delivers the correct output:
Additional information
No response
Operating system
No response
Matplotlib Version
3.5.2
Matplotlib Backend
module://matplotlib_inline.backend_inline
Python version
Python 3.9.7
Jupyter version
6.4.12
Installation
No response
Activity
tacaswell commentedon Jul 29, 2022
I suspect that this is fixed by ipython/matplotlib-inline#14
The interaction here is:
Due to the delayed import matplotlib-inline is now imported inside of the context manager which overrides your values (and which the context sets back to what it was before on exit which hides the state change!). The call to
plt.show()fixes the problem because it forces the import to be earlier. Simlarly, running the cell a second time works because the module is already imported and hence does not change the rcparams.Adding
plt.get_backend()will also force the import and is the work-around I would go with for now (it should work on all versions of mpl / matplotlib-inline and is mostly harmless).I am going to close this issue because we understand it and a fix has been merged downstream. Sorry we broke this on you.
juhuebner commentedon Jul 29, 2022
@tacaswell
Thank you very much for the really quick and insightful answer. However, when I put
plt.get_backend()in-between the import and the context section, it has no effect likeplt.show(). I noticedplt.clf()works as well. However, in both cases I don't know if I generate some potential pitfalls.tacaswell commentedon Jul 29, 2022
One day I will learn to always test my suggestions before suggesting....
I (incorrectly) thought that you would be going through the backend fallback logic so checking the would do import the backend module, but there is actually a string already set so
get_backendis fast and has no side effects 🤦🏻 .Instead doing:
will definitely import the module and trigger the rcparam changing logic.
juhuebner commentedon Aug 3, 2022
There is still something fishy with this import.
You are probably referring to https://github.com/ipython/matplotlib-inline/blob/f764b4354b2358aa001471215494d7dfdc505593/matplotlib_inline/backend_inline.py#L227
However, somewhere in the machinery the module already got imported, which then hinders the re-execution of
_enable_matplotlib_integration()You can see it via, e.g.,
However, explicitly importing it into a new namespace works:
tacaswell commentedon Aug 3, 2022
I did not learn my lesson about testing in the post where I noted needing to learn my lesson about testing 🤦🏻