ICA plot_components opens figure one by one (and might hang depending on the console)

[mscheltienne]

Description of the problem

A video will be clearer. Same behavior in a regular python console and in an IPython console.

Screencast.from.05-12-2023.04.45.47.PM.webm

The figure are shown one by one after you close the old one. At least in a python or IPython console, it works. In spyder, it hangs the console after you close the first figure 😞

Tested both in a working environment and in a fresh environment with only pip install ipython mne installed (MNE 1.4).

Steps to reproduce

from mne.datasets import sample
from mne.io import read_raw_fif
from mne.preprocessing import ICA


directory = sample.data_path() / "MEG" / "sample"
raw = read_raw_fif(directory / "sample_audvis_raw.fif", preload=False)
raw.pick_types(eeg=True)
raw.crop(0, 10)
raw.load_data()
raw.filter(1., 40.)
ica = ICA(n_components=None)
ica.fit(raw)
ica.plot_components(inst=raw)

No idea what is causing it at the moment, will try to look into it. Might be related to #11654 and to #11510

[drammock]

the plots are not created within a plt.ion() context so they're blocking execution. A solution would be to generate (but not show) the figures in the for-loop over subsets of components, then show all of them at the end after the for-loop completes. Getting this stuff to work right in all contexts (notebook, ipython, spyder, pycharm) is tricky; don't hesitate to ask for help if you get stuck.

[mscheltienne]

I have to say, this blocking/event-loop/backend aspect of matplotlib always eluded me. Couple of additional points:

• This issue did not appear because of a change we made. Testing the same code snippet in python/IPython/Spyder with MNE 1.2 results in the same behavior while it was correctly displaying all figures a couple of months ago. I tried rolling back some dependencies, including matplotlib, without luck. Not sure which one is responsible.
• One solution is to provide the argument block=False here:

  mne-python/mne/viz/topomap.py

  Line 1742 in 1e6aa7d

  plt_show(show)

  
  After this change, it does correctly open all figures BUT:
  • from a regular python interpreter the call becomes non-blocking, thus the code snippet above by itself would run and end immediately. Maybe that's how it should be?
  • from an IPython console or from a regular python interpreter (blocked with an input() at the end), opening the properties plot by clicking on the topographies works but with the lovely QCoreApplication::exec: The event loop is already running. This warning is not present in Spyder.

---

And for your proposition @drammock

show all of them at the end after the for-loop completes

I don't see an easy way to do it without adding an argument to flag if we are in a single execution (picks is not None) or in a recursive execution (initial call with picks=None) and I don't like this solution for a public function.
EDIT: I will rewrite this afternoon the function structure to remove the recursion, see if that helps.

---

Matplotlib backend used for testing: QtAgg with PyQt5.

[mscheltienne]

More discoveries! In #11696 I removed the recursion in plot_ica_components and now gather all figures and show only once at the end. I don't see a downside in this approach, and it solves part of the issue. All the figures open at once.

But:

• In a python interpreter, works great.
• In an IPython console, works great.
• In a Spyder console or in a jupyter notebok, closing the figure hangs the interpreter 😮‍💨 (similar to Matplotlib plot in jupyterlab using Qt hangs except if an inline plot was produced before  #11510 maybe?)

Speedy screencast to demonstrate (too bad it doesn't capture the mouse..)

Screencast.from.05-15-2023.02.46.05.PM.webm

And now the discovery part. It works on user-provided axes!

from matplotlib import pyplot as plt
from mne.datasets import sample
from mne.io import read_raw_fif
from mne.preprocessing import ICA


directory = sample.data_path() / "MEG" / "sample"
raw = read_raw_fif(directory / "sample_audvis_raw.fif", preload=False)
raw.pick_types(eeg=True)
raw.crop(0, 10)
raw.load_data()
raw.filter(1., 40.)
ica = ICA(n_components=5)
ica.fit(raw)

f, ax = plt.subplots(1, 2)
ica.plot_components(inst=raw, picks=[0, 2], axes=ax)

This works in all of those 4 tests and does not hang the interpreter in Spyder and Jupyter. Something is not behaving around here..

mne-python/mne/viz/utils.py

Lines 470 to 527 in 1e6aa7d

	def _prepare_trellis(
	n_cells,
	ncols,
	nrows="auto",
	title=False,
	colorbar=False,
	size=1.3,
	sharex=False,
	sharey=False,
	):
	from matplotlib.gridspec import GridSpec
	from ._mpl_figure import _figure
	if n_cells == 1:
	nrows = ncols = 1
	elif isinstance(ncols, int) and n_cells <= ncols:
	nrows, ncols = 1, n_cells
	else:
	if ncols == "auto" and nrows == "auto":
	nrows = math.floor(math.sqrt(n_cells))
	ncols = math.ceil(n_cells / nrows)
	elif ncols == "auto":
	ncols = math.ceil(n_cells / nrows)
	elif nrows == "auto":
	nrows = math.ceil(n_cells / ncols)
	else:
	naxes = ncols * nrows
	if naxes < n_cells:
	raise ValueError(
	"Cannot plot {} axes in a {} by {} "
	"figure.".format(n_cells, nrows, ncols)
	)
	if colorbar:
	ncols += 1
	width = size * ncols
	height = (size + max(0, 0.1 * (4 - size))) * nrows + bool(title) * 0.5
	height_ratios = None
	fig = _figure(toolbar=False, figsize=(width * 1.5, 0.25 + height * 1.5))
	gs = GridSpec(nrows, ncols, figure=fig, height_ratios=height_ratios)
	axes = []
	if colorbar:
	# exclude last axis of each row except top row, which is for colorbar
	exclude = set(range(2 * ncols - 1, nrows * ncols, ncols))
	ax_idxs = sorted(set(range(nrows * ncols)) - exclude)[: n_cells + 1]
	else:
	ax_idxs = range(n_cells)
	for ax_idx in ax_idxs:
	subplot_kw = dict()
	if ax_idx > 0:
	if sharex:
	subplot_kw.update(sharex=axes[0])
	if sharey:
	subplot_kw.update(sharey=axes[0])
	axes.append(fig.add_subplot(gs[ax_idx], **subplot_kw))
	return fig, axes, ncols, nrows

Probably the MNEFigure part since bypassing _figure also works.

[mscheltienne]

mne-python/mne/viz/_mpl_figure.py

Lines 78 to 85 in 1e6aa7d

	plt.ion()
	BACKEND = get_backend()
	# This ↑↑↑↑↑↑↑↑↑↑↑↑↑ does weird things:
	# https://github.com/matplotlib/matplotlib/issues/23298
	# but wrapping it in ion() context makes it go away (can't actually use
	# `with plt.ion()` as context manager, though, for compat reasons).
	# Moving this bit to a separate function in ../../fixes.py doesn't work.
	plt.ioff()

What is the purpose of those lines? If we remove them, everything works fine in python, ipython, spyder and jupyter (still needs to be tested in VSCode, PyCharm).