WIP: Decrease Brain memory consumption

[GuillaumeFavelier]

The first goal is to collect all the remaining vtk objects:

mprof run mem_brain.py && mprof plot

mem_brain.py:

import gc
import time
import numpy as np
import mne
rr, tris = mne.read_surface(mne.datasets.sample.data_path() + '/subjects/fsaverage/surf/lh.inflated')
tris = np.pad(tris, ((0, 0), (1, 0)), 'constant')
assert tris.shape[1] == 4
tris[:, 0] = 3
for _ in range(1):
    brain = mne.viz.Brain('fsaverage', 'both', 'inflated')
    brain.close()
    del brain
    time.sleep(0.1)
    gc.collect()
    print([o.__class__.__name__ for o in gc.get_objects() if o.__class__.__name__.startswith('vtk')], len(gc.get_objects()))

To remove:

['vtkXOpenGLRenderWindow', 'vtkGenericRenderWindowInteractor', 'vtkLookupTable', 'vtkOpenGLActor']

The second goal is to improve Brain memory consumption to help with #8379

cc @larsoner

[larsoner]

I wonder if vtkGenericRenderWindowInteractor cannot be GC'ed because of AddObserver, see http://vtk.1045678.n5.nabble.com/VTK-memory-management-td5715661.html

Then the LookupTable / Actor is probably another (separate) problem? It's probably going to take more code commenting-out and line tracking to sort it out...

[larsoner]

Now that Brain is public we should probably make as many attributes as possible private, so _plotter is better here

[GuillaumeFavelier]

And an automatic test would be to ensure that this list is actually empty.

[larsoner]

And an automatic test would be to ensure that this list is actually empty.

If you can get the list empty, I can add this test :)

What I'll really probably do is port the gc-testing parts of #8379 here and add the vtk-object-GC check, then we merge this, then I rebase #8379 ...

[GuillaumeFavelier]

Those 2 are added in BackgroundPlotter:

['vtkXOpenGLRenderWindow', 'vtkGenericRenderWindowInteractor']

And the rest is related to our use of add_mesh somehow:

['vtkLookupTable', 'vtkOpenGLActor']

[GuillaumeFavelier]

c2f56b5 gets rid of the ref to vtkLookupTable but one vtkOpenGLActor remains.

[larsoner]

See also pyvista/pyvista#958

[GuillaumeFavelier]

The remaining vtkOpenGLActor is a true mystery.

It appears in this code:

import gc
import time
import numpy as np
import mne
rr, tris = mne.read_surface(mne.datasets.sample.data_path() + '/subjects/fsaverage/surf/lh.inflated')
tris = np.pad(tris, ((0, 0), (1, 0)), 'constant')
assert tris.shape[1] == 4
tris[:, 0] = 3
for _ in range(1):
    brain = mne.viz.Brain('fsaverage', 'both', 'inflated')
    brain.close()
    del brain
    time.sleep(0.1)
    gc.collect()
    print([o.__class__.__name__ for o in gc.get_objects() if o.__class__.__name__.startswith('vtk')], len(gc.get_objects()))

But not in this one:

import gc
import time
import numpy as np
import pyvista
import mne
rr, tris = mne.read_surface(mne.datasets.sample.data_path() + '/subjects/fsaverage/surf/lh.inflated')
tris = np.pad(tris, ((0, 0), (1, 0)), 'constant')
assert tris.shape[1] == 4
tris[:, 0] = 3
for _ in range(1):
    brain = mne.viz.Brain('fsaverage', 'both', 'inflated')
    brain.close()
    del brain
    time.sleep(0.1)
    gc.collect()
    print([o.__class__.__name__ for o in gc.get_objects() if o.__class__.__name__.startswith('vtk')], len(gc.get_objects()))

[GuillaumeFavelier]

@larsoner on my machine, using this script:

import gc
import time
import numpy as np
import mne
rr, tris = mne.read_surface(mne.datasets.sample.data_path() + '/subjects/fsaverage/surf/lh.inflated')
tris = np.pad(tris, ((0, 0), (1, 0)), 'constant')
assert tris.shape[1] == 4
tris[:, 0] = 3
for _ in range(1):
    brain = mne.viz.Brain('fsaverage', 'both', 'inflated')
    brain.close()
    del brain
    time.sleep(0.1)
    gc.collect()
    print([o.__class__.__name__ for o in gc.get_objects() if o.__class__.__name__.startswith('vtk')], len(gc.get_objects()))

gives:

['vtkOpenGLActor']

And strangely enough if I import pyvista, it disappears. I don't have an explanation for now.

Can you confirm that you obtain a similar list?

[larsoner]

@GuillaumeFavelier I incorporated this into #8379 except for d597e80. Feel free to cherry-pick that one and push to my PR. Travis should fail there soon with some VTK objects linger errors, you should be able to replicate locally on my branch with just:

pytest mne/viz/_brain -x

[GuillaumeFavelier]

I'll work directly on your branch then 👍