BUG: VectorSourceEstimate.plot() in v20 does not show the surfaces

[yh-luo]

Describe the bug

VectorSourceEstimate.plot() is not working properly in v20.

Steps to reproduce

I used the sample dataset to compute some VectorSourceEstimate files to play around. Trouble shootings in Advanced setup of MNE-Python did not fix the problem.

import os.path as op
import mne
from scripts.config import baseline, meg_dir, subjects_dir

os.environ["SUBJECTS_DIR"] = subjects_dir
stc_aud = mne.read_source_estimate(
    op.join(meg_dir, 'aud_left_minus_right-grand_average-ave-stc'))
stc_aud.plot(hemi='both', initial_time=0.1)

Expected results

In MNE v19, it works without problems

Actual results

In MNE v20, the surfaces disappear after plotting.

The problem persists in master.

Additional information

If required, I can upload the source estimate file.

mne.sys_info

Platform:      Linux-5.6.8-200.fc31.x86_64-x86_64-with-fedora-31-Thirty_One
Python:        3.7.7 (default, Mar 13 2020, 10:23:39)  [GCC 9.2.1 20190827 (Red Hat 9.2.1-1)]
Executable:    /home/yuhan/env/mne/bin/python
CPU:           x86_64: 4 cores
Memory:        15.5 GB

mne:           0.20.3
numpy:         1.17.3 {blas=openblas, lapack=openblas}
scipy:         1.3.2
matplotlib:    3.1.1 {backend=Qt5Agg}

sklearn:       0.21.3
numba:         0.46.0
nibabel:       2.5.1
cupy:          Not found
pandas:        0.25.3
dipy:          1.0.0
mayavi:        4.7.1 {qt_api=pyqt5, PyQt5=5.13.2}
pyvista:       0.24.1
vtk:           8.1.2

[larsoner]

This could be garbage collection, if you do brain = stc.plot(....) to keep a reference to the object, does it help?

[yh-luo]

@larsoner No, brain = stc_aud.plot(hemi='both', initial_time=0.1) shows the same behavior. Also, the bug exists in both Python interpreter and IPython.

[yh-luo]

I used Visualize source time courses to further test SourceEstimate.plot(), VolSourceEstimate.plot() and VectorSourceEstimate.plot(), only VectorSourceEstimate has this problem.

[yh-luo]

I think depth peeling could be the issue. The surfaces are shown correctly (but ugly) when brain_alpha=1. I tried to comment out the below part to avoid depth peeling and the problem is fixed too.

https://github.com/mne-tools/mne-python/blob/5f4c9046fb07cf75a316d56d809e4b7266753ceb/mne/viz/_3d.py#L2503-L2507

[GuillaumeFavelier]

I can't reproduce this one.

Depth peeling for the mayavi backend was enabled in https://github.com/mne-tools/mne-python/pull/7030 to fix graphical issues with the glass brain. But apparently, this workaround is not working for every configuration.

Choosing the less painful solution, between graphical issue and a brain that closes by itself: should we disable depth peeling @jasmainak, @larsoner ?

[yh-luo]

I'm closing this because it's actually a problem of Fedora with NVIDIA Optimus laptop. Others may encounter the same Mayavi rendering problem so the solution is below:

1. Install the proprietary nvidia driver
2. Configure graphics applications to render using the NVIDIA GPU. Set the environment variable __NV_PRIME_RENDER_OFFLOAD=1 and __GLX_VENDOR_LIBRARY_NAME=nvidia before running graphical applications. Or set them in .bashrc

   export __NV_PRIME_RENDER_OFFLOAD=1
   export __GLX_VENDOR_LIBRARY_NAME=nvidia

   I'm not sure if other Linux users have this problem with dual GPU laptops though.

[GuillaumeFavelier]

Thanks for sharing!

I'm not sure if other Linux users have this problem with dual GPU laptops though.

I used a slightly more complicated method on my Manjaro to get rid of those visual artifacts.