MRG, ENH: make subplots_adjust work with {raw,epochs}.plot() and ica.plot_sources()

[drammock]

closes #6647

• raw
• epochs
• ICA

[drammock]

Here's an example of subplots_adjust working on a raw.plot() figure: https://14722-1301584-gh.circle-artifacts.com/0/dev/auto_tutorials/preprocessing/plot_15_handling_bad_channels.html#interpolation-in-mne-python

[codecov]

Codecov Report

Merging #6670 into master will increase coverage by 0.09%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #6670      +/-   ##
==========================================
+ Coverage   89.39%   89.48%   +0.09%     
==========================================
  Files         415      420       +5     
  Lines       75230    75589     +359     
  Branches    12366    12383      +17     
==========================================
+ Hits        67249    67638     +389     
+ Misses       5149     5142       -7     
+ Partials     2832     2809      -23

[lgtm-com]

This pull request introduces 5 alerts when merging 3a3666f into 0ad37f0 - view on LGTM.com

new alerts:

• 3 for Unhashable object hashed
• 2 for Unused import

[cbrnr]

Any chance you can also implement a zen mode (i.e. get rid of the buttons and the timeline overview)?

[drammock]

Yes, but I'd like to do that in a separate PR.

…

On August 17, 2019 2:52:37 AM PDT, Clemens Brunner ***@***.***> wrote: Any chance you can also implement a zen mode (i.e. get rid of the buttons and the timeline overview)?

[drammock]

CIs are happy except for Circle (complaining about dipy link, see #6671). Ready for review and merge. Someone with a hi-dpi screen (@larsoner?) should try generating and resizing the raw.plot() window, then calling subplots_adjust(), to make sure everything looks normal.

[larsoner]

@cbrnr also feel free to try. I'll give it a shot tomorrow

[cbrnr]

I tried raw.plot on my hidpi screen and it works perfectly! +1 for merge (maybe @larsoner can you test this for epochs and/or ICA?).

[larsoner]

On my HiDPI macOS screen things do not look correct. The following code:

import mne
path = mne.datasets.sample.data_path()
raw = mne.io.read_raw_fif(path + '/MEG/sample/sample_audvis_raw.fif')
raw.plot()
events = mne.find_events(raw)
epochs = mne.Epochs(raw, events)
epochs.plot()

Shows this (just the raw plot here but epochs has the same spacing issues) on master, which has no overlap between elements (correct):

But on this PR it looks like:

I am not sure why I have these overlap problems but @cbrnr did not (unless you missed them?).

[larsoner]

If you want to test on a non-HiDPI screen as long as you're on the Qt5Agg backend you can do (here on my Linux non-HiDPI machine):

QT_SCALE_FACTOR=2 python -ui ~/Desktop/plot.py

I get on master:

But on this PR:

[drammock]

thanks, I think I know what's wrong. Will fix today.

…

On August 19, 2019 7:47:48 AM PDT, Eric Larson ***@***.***> wrote: If you want to test on a non-HiDPI screen as long as you're on the Qt5Agg backend you can do (here on my Linux non-HiDPI machine): ``` QT_SCALE_FACTOR=2 python -ui ~/Desktop/plot.py ``` I get on `master`:  But on this PR: 

[drammock]

@larsoner @cbrnr the changes just pushed work locally (show correct button alignment / no overlap) when testing with QT_SCALE_FACTOR=2 (or 3 or 4). Please double-check, but assuming it's good (and the CIs are happy) this should be ready for merge.

[drammock]

Still looks OK on Circle...

https://14746-1301584-gh.circle-artifacts.com/0/dev/auto_tutorials/preprocessing/plot_15_handling_bad_channels.html#interpolation-in-mne-python

[larsoner]

The gaps at the top and bottom seem incorrect. master:

This PR:

[larsoner]

Thanks for the deep dive into mpl @drammock. Works on my macOS HiDPI screen and your Linux tests look conclusive. Example also looks good:

https://14782-1301584-gh.circle-artifacts.com/0/dev/auto_tutorials/preprocessing/plot_15_handling_bad_channels.html

+1 for merge from me

[cbrnr]

Unfortunately now it doesn't work for me - but only if I'm on a scaled resolution and using the osx backend (qt5 works fine). After resizing the window (subplots_adjust works fine), the margins are wrong:

import mne
path = mne.datasets.sample.data_path()
raw = mne.io.read_raw_fif(path + '/MEG/sample/sample_audvis_raw.fif')
raw.plot()

Initially, this looks fine:

After manually resizing the window:

[larsoner]

I also see this when running with MPLBACKEND=macosx. This is almost certainly a matplotlib bug in that backend (it tends to be buggy IIRC). That backend is also just about unusable for me it's so slow. I wonder if we should really support it.

@drammock is there something I could try adding as a resize event handler (for the macosx backend only) that might force it to update the params?

[drammock]

I'm really not sure what to do here. I can't tell from the screenshots whether resizing is not happening at all, or if it's happening but not to the correct degree.

[cbrnr]

Let's not drop osx support. This is still the default backend and I don't find it slower than qt5 (what's really slowing my Mac down to a point where it's almost unusable is scaled resolution aka HiDPI, but that's another story).

[drammock]

I have arranged access to a Mac for next week so I'll try to debug on Monday. I tried today on KC's Mac desktop but could not reproduce (his resolution may not be high enough? It is an "Apple Cinema display" 2560 x 1440)

[larsoner]

The Qt5Agg backend is fine, did you do it with MPLBACKEND=MacOSX python -ui test.py or so?

[drammock]

I tested only with the default backed but had to run to a meeting before I could verify whether it was defaulting to qt or macosx. Will continue testing tomorrow.

[cbrnr]

@drammock you need to set a scaled resolution, it works with the native resolution. I'm not sure this works with 2560x1440, but try setting it to a scaled (lower) resolution. I have a 4K display and select the second resolution to the right in the Displays settings page.

[drammock]

I'm really struggling to reproduce this. Screenshot:

[drammock]

running in a plain python terminal instead of ipython, the mouse cursor never turns to double-headed arrow so it looks like resizing doesn't work --- note also the grayed out close/minimize buttons even though the window is on top --- but if I click and drag the window edge anyway and then go back to the terminal and start typing, the window size does belatedly change (and behavior of the margins is as expected). Before resize:

After subplots_adjust and resize:

[drammock]

Running as a script I also get expected behavior:

[drammock]

Here are the test system details:

(mnedev) D-10-19-149-242:~ labshared$ python -c 'import mne; mne.sys_info()'
Platform:      Darwin-17.4.0-x86_64-i386-64bit
Python:        3.7.4 (default, Aug 13 2019, 15:17:50)  [Clang 4.0.1 (tags/RELEASE_401/final)]
Executable:    /Users/labshared/miniconda3/envs/mnedev/bin/python
CPU:           i386: 12 cores
Memory:        16.0 GB

mne:           0.19.dev0
numpy:         1.16.4 {blas=mkl_rt, lapack=mkl_rt}
scipy:         1.3.1
matplotlib:    3.1.0 {backend=MacOSX}

sklearn:       0.21.2
numba:         0.45.1
nibabel:       2.5.0
cupy:          Not found
pandas:        0.25.0
dipy:          1.0.0
mayavi:        4.7.1 {qt_api=pyqt5, PyQt5=5.9.2}
pyvista:       0.22.1
vtk:           8.1.2

[drammock]

Same results at 2048x1152 resolution:

...and at 1280x720:

[drammock]

for completeness here are the Qt5Agg screenshots on the same system. Default scaling:

With QT_SCALE_FACTOR=2:

[cbrnr]

I think your monitor might not be recognized as a retina/HiDPI screen because its native resolution is too low. To reproduce the problem, you need to have a screen that macOS recognized as retina/HiDPI capable. Here's what my display preferences window looks like - notice the five presets for different scaling factors:

I think this is only shown if the monitor is at least 4K, and these settings seem to influence how scaling is performed. In your case, I think macOS performs regular non-HiDPI scaling, and therefore you don't see the problem. I can simulate this even on my HiDPI screen by alt-clicking on the "Scaled" radio button and selecting a resolution that says "(low resolution)":

Now this is exactly the same 3360 x 1890 resolution I normally use, but in low resolution mode (aka non-HiDPI) I don't see the issue with overlapping elements either.

[cbrnr]

PS: What this means is that this issue occurs on MacBooks with retina displays (so basically all of them) with the default osx backend (so no external display required). This is even worse, because there is no way to select the native (full) resolution for internal displays (because for whatever reason, even the maximum resolution is a scaled one).

[cbrnr]

PPS: Here's the code snippet I used (copied from your screenshots):

import os
import matplotlib.pyplot as plt
import mne


plt.get_backend()
sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = os.path.join(sample_data_folder, 'MEG', 'sample',
                                    'sample_audvis_raw.fif')
raw = mne.io.read_raw_fif(sample_data_raw_file)
raw.crop(0, 60).load_data()
fig = raw.plot()  # works
fig.subplots_adjust(top=0.9)  # works
# now resize the window - destroys figure layout

[drammock]

@cbrnr @larsoner I've pushed a fix that I think will work. Please give it a try. (shout out to @ktavabi for letting me use his machine to debug).

[lgtm-com]

This pull request introduces 3 alerts when merging 2f1b9b6 into b7dd664 - view on LGTM.com

new alerts:

• 3 for Unhashable object hashed

[cbrnr]

Nice! I can confirm that this now works with all combinations of osx/qt5 backends and retina/non-retina resolutions for me.

If the LGTM warnings are fixed (by e.g. using tuples instead of lists), +1 for merge!

[drammock]

@cbrnr those LGTM alerts are unrelated to this PR, and I would consider them false alarms. Moreover I don't think it will work to just change them to tuples. Two of the lists it's complaining about are being used to index numpy arrays, and lists/tuples are treated differently in that context. The third one is an instance of slice that is used to index a list. Frankly I'm not sure why those objects even need to be hashable in that context.

[drammock]

forgot to include the link; see here: https://lgtm.com/projects/g/mne-tools/mne-python/rev/pr-0e4df471613fdb5248495edb0782bef332768744

[agramfort]

Lets merge then

[cbrnr]

Thanks @drammock!

[larsoner]

BTW @drammock if you ever need to debug HiDPI on macOS on a non-HiDPI display, you can enable (usually ugly) HiDPI resolutions typically hidden from the user with something like:

$ sudo defaults write /Library/Preferences/com.apple.windowserver.plist DisplayResolutionEnabled -bool true

It's also possible clicking the "scaled" button while holding Alt/Option might do it (or be necessary).

I did something like this a couple of years ago when I only had access to a non-HiDPI macOS display and it worked. So hopefully next time you won't need to borrow @ktavabi's machine but any macOS one (where the owner doesn't mind some option tweaking) should work :)