[ENH] improve MEM performance #25

Edouard2laire · 2024-10-02T22:34:29Z

This PR:

Update the definition of the covariance matrix along the cortical surface
Improve the performance
improve memory usage when using parallel computing.
Improve the disk usage of the output

Reference:

* don't display progress bar for standalone * First implementation * clean code * aestetic changes * more aesthetic change * add spaces

* Improve parameter choice for l-curve * Improve kernel estimation using SVD trick to compute the norm

* Improve parameter choice for l-curve * Improve kernel estimation using SVD trick to compute the norm * better handle figures * Remove be_main_wmem

Optimize be_free_energy and be_mem_struct

Save wMEM results as factor decomposition

* Attempt to fix the covariance matrix * refactor changes * remove previous trick to be safer

* attempt to use hessian to speed up calculation * use hessian * Remove Hessian for now

* more refactor on memstruct * more refactor * more refactor * Fix memory issue for EEG/MEG * temporary move new implementation to be_free_energy2 * start support for custom mean / inactive variance * add unsaved change * remove unnecessary sum * minor cosmetic change * bug fix and clean code

best/mem/be_launch_mem.m

best/mem/be_minimize_free_energy.m

R was not a triangular matrix...

(python only support full)

* fix rMEM * some cleaning

This reverts commit b040ed1.

Edouard2laire and others added 28 commits October 2, 2024 18:10

improve MNE solver

c153b37

use scale 3 for nirs noise covariance

aee123a

disable noise covariance in MNE for now

83d47f9

keep all the power for nirs

11d434b

Par-for optimization (#8)

87e0bf2

* don't display progress bar for standalone * First implementation * clean code * aestetic changes * more aesthetic change * add spaces

MNE kernel (#9)

74767e9

* Improve parameter choice for l-curve * Improve kernel estimation using SVD trick to compute the norm

small changes (#10)

fe22b41

* Improve parameter choice for l-curve * Improve kernel estimation using SVD trick to compute the norm * better handle figures * Remove be_main_wmem

don't transpose the inversion matrix

06104bf

oups

9a50128

Move the normalization outside be_launch_mem

a9974ba

Move the conversion to time-series outside be_launch_mem

6b66e33

minor changes

495be0b

clean up output

ee42333

more clean-up on be_slice_obj

73ca08d

Merge pull request #11 from Edouard2laire/be_free_energy

a1a2807

Optimize be_free_energy and be_mem_struct

save sigma_s under obj rather than option

654648b

pre-compute G * active_var

b9dae0c

some cleaning

8c470b4

Even more optim

9db0652

last changes

f852199

fix call to be_fusion_of_modalities

6a8e3aa

cleam up MNE code

ac8778e

Save wMEM results as factor decomposition

02c634f

fix format of obj.data

dfccf7a

Merge pull request #15 from Edouard2laire/storage

05e4e51

Save wMEM results as factor decomposition

[bugfix] Fix the covariance matrix along the cortical surface (#17)

66ee9c9

* Attempt to fix the covariance matrix * refactor changes * remove previous trick to be safer

update optimization function (#16)

96240ab

* attempt to use hessian to speed up calculation * use hessian * Remove Hessian for now

Merge branch 'multifunkim:master' into wMEM-nirs2

05c4145

Edouard2laire marked this pull request as ready for review October 23, 2024 18:58

[refactor] call be_fusion_of_modality in MNE

da4683c

Edouard2laire and others added 2 commits November 1, 2024 17:57

Add option to use noise covariance in MNE. (default to false)

b95fa1e

Edouard2laire changed the title ~~[wMEM] improve performance for fNIRS data~~ [ENH] improve MEM performance Nov 26, 2024

Edouard2laire added 8 commits November 27, 2024 17:05

also reduce cMEM file size if only a part of the signal is localized

40b2b6e

Put back for loop (so much faster for nirs)

b20cbb9

Add datatype to pipeline options

36fb6b3

wavvelet_inverse_projection: use sparse matrix to avoid memory issue

cdc84ee

remove extra message from be_slice_obj

89a502a

bugfix when using inactive variance

294b563

Add be_pagemtimes for backward compatibility

1f6ac16

officially replace be_free_energy

7724734

Edouard2laire commented Dec 6, 2024

View reviewed changes

best/mem/be_launch_mem.m Outdated Show resolved Hide resolved

best/mem/be_launch_mem.m Outdated Show resolved Hide resolved

best/mem/be_minimize_free_energy.m Outdated Show resolved Hide resolved

Edouard2laire and others added 16 commits December 6, 2024 17:22

Remove be_launch_mem2

9ae6076

Apply comments

370214e

FIx QR decomposition for matlab < 2022a

58b8e63

R was not a triangular matrix...

be_pipelineoption; fix typo

f2e9f87

improve normalization (adpatative vs fixe)

3266cda

Fix figure display (reuse figure)

af42df1

Improve entropy drop display for cMEM

32fac39

Add a flag controling the output form (full or factor)

2b0fff4

(python only support full)

Cleanup option before saving

3aa4696

small refactor

7db63a1

fix rMEM code (part I / II) (#20)

b040ed1

* fix rMEM * some cleaning

WIP on rMEM

f785626

Add new option in the control panel

29e3cb6

minor update

d6feda5

Revert "fix rMEM code (part I / II) (#20)"

ab04ab9

This reverts commit b040ed1.

update version

561a975

Edouard2laire merged commit 9fc7629 into multifunkim:master Mar 12, 2025

Edouard2laire deleted the wMEM-nirs2 branch March 12, 2025 19:08

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[ENH] improve MEM performance #25

[ENH] improve MEM performance #25

Edouard2laire commented Oct 2, 2024 •

edited

Loading

[ENH] improve MEM performance #25

[ENH] improve MEM performance #25

Conversation

Edouard2laire commented Oct 2, 2024 • edited Loading

Edouard2laire commented Oct 2, 2024 •

edited

Loading