ipympl changes MPL backend on import

[sappelhoff]

• xref: Preserve Matplotlib backend across sys_info() mne-tools/mne-python#13963 (review)

In mne-python (see above), we have a functionality called sys_info, which gives us (among other things) the versions of all sorts of installed packages.

For this reason, we also import ipympl.

At import time, ipympl actually changes the MPL backend.

Is this something that could be avoided? This change in backend was very surprising to us and it took an in-depth profiling to identify it as the cause of a massive memory leak in a pipeline that produced several plots on each iteration, never closing them due to the backend change.

[ianhi]

Hmm I agree that that is a bad behavior. I suspect this can be changed depending on how the %matplotlib ipympl magic and matplotlib backend import works.

Some relevant code paths from a quick check

https://github.com/matplotlib/matplotlib/blob/fe78309725e02eb679c29ce24029c5576ab4bb14/lib/matplotlib/backends/registry.py#L139-L151

https://github.com/ipython/ipython/blob/fdbdf1c699e80f76a71e3dc5b46a3d164939b3f2/IPython/core/pylabtools.py#L38-L40

The actual lines that cause the backend to be changed on import.

ipympl/ipympl/__init__.py

Lines 23 to 28 in 36ae86d

	# __init__.py is used by the nbextension installation.
	# Conda cannot have dependencies for post-link scripts.
	if 'matplotlib' in sys.modules:
	import matplotlib
	matplotlib.use('module://ipympl.backend_nbagg')

which are 8 years old and nbextension is no longer relevant so I'm pretty sure we can just remove that entirely. But would love to hear what @martinRenou thinks before doing that

As a workaround for now you can try directly importing from _version

In [1]: from ipympl._version import version_info

In [2]: version_info
Out[2]: (0, 10, 0)