[FieldTrip] ft_sourcestatistics startup memory problem
Akiko Ikkai
akiko.ikkai at gmail.com
Sat Oct 20 00:10:07 CEST 2012
Hi,
So I have followed Johanna and Stephan's advise to create template_grid and
each subject's grid from MNI image, which is now 1cm spacing instead of
.5cm. These steps seem to run, and I'm able to create a common spatial
filter to run beamformer using DICS. However, I get an error when I apply
the commom filter to each condition.
I had to tweak a few things on example on the page
http://fieldtrip.fcdonders.nl/example/create_single-subject_grids_in_individual_head_space_that_are_all_aligned_in_mni_space?s[]=template&s[]=grid
, since my data is EEG and electrode locations vary between subjects,
such
that:
cfg = [];
cfg.grid.xgrid = -20:1:20;
cfg.grid.ygrid = -20:1:20;
cfg.grid.zgrid = -20:1:20;
cfg.grid.tight = 'yes';
cfg.inwardshift = -1.5;
cfg.vol = template_vol;
*cfg.elec = ft_data.elec; % individual sub's electrode locations: had to
add this (got error otherwise)*
template_grid = ft_prepare_sourcemodel(cfg);
then
cfg = [];
cfg.grid.warpmni = 'yes';
cfg.grid.template = template_grid;
cfg.grid.nonlinear = 'yes';
cfg.mri = mri_aligned;
*cfg.elec = ft_data.elec; **% individual sub's electrode locations: had to
add this (got error otherwise)*
grid = ft_prepare_sourcemodel(cfg);
These create grids in MNI space, both for template and individual fine. To
create a common spatial filter, I run the following command, which runs:
cfg = [];
cfg.grid.pos = grid.pos;
cfg.grid.dim = grid.dim;
cfg.grid.inside = grid.inside;
cfg.grid.outside = grid.outside;
cfg.inwardshift = -1.5;
cfg.vol = vol_cm;
cfg.channel = {'all'};
cfg.frequency = 10.5;
cfg.method = 'dics';
cfg.dics.projectnoise = 'yes';
cfg.dics.keepfilter = 'yes';
cfg.dics.feedback = 'no';
source_common = ft_sourceanalysis(cfg, freq_common);
however, when I apply this spatial filter to a task condition, I get an
error message
cfg = [];
cfg.elec = freq_common.elec;
cfg.grid.pos = source_common.pos;
cfg.grid.filter = source_common.avg.filter;
cfg.inwardshift = -1.5;
cfg.vol = vol_cm;
cfg.channel = {'all'};
cfg.frequency = 10.5;
cfg.method = 'dics';
cfg.dics.projectnoise = 'yes';
cfg.dics.keepfilter = 'yes';
cfg.dics.feedback = 'no';
source_itemL = ft_sourceanalysis(cfg, freq_itemL);
source_itemL.unit = 'cm';
??? Error using ==> mtimes
Inner matrix dimensions must agree.
Error in ==> beamformer_dics at 316
csd = filt * Cf * ctranspose(filt); % Gross
eqn. 4
and 5
Error in ==> ft_sourceanalysis at 588
dip(i) = beamformer_dics(grid, sens, vol, [], squeeze(Cf(i,:,:)),
optarg{:});
The error occurs when dip.pos = 40 (i.e. it runs fine 1-39). I can't spot
where exactly the error is originating from. Am I missing some steps when I
create template grids that I feed into ft_sourceanalysis?
Thank you in advance for your help! Akiko
On Thu, Oct 18, 2012 at 4:29 PM, Akiko Ikkai <akiko.ikkai at gmail.com> wrote:
> Hi Everyone,
>
> Thank you very much for your advise. I just came back from SfN, and just
> found your replies, so I will test these options and report back what I
> find.
>
> Thank you again! Akiko
>
>
> On Mon, Oct 8, 2012 at 6:14 AM, Stephen Whitmarsh <
> stephen.whitmarsh at gmail.com> wrote:
>
>> Hi Akiko!
>>
>> Just to chime in - your source grid is very, very large indeed! ALthough
>> I started with a very fine grid, at one point I also had to downside it
>> (going to 5mm), and had to stop using interpolated source data for stats
>> and the like. Johanna's suggestion will certainly do the trick and it will
>> speed up your analysis enormously as well. You then only need to do
>> interpolate for plotting purposes.
>>
>> all the best,
>> Stephen
>>
>>
>> On 8 October 2012 12:02, Johanna Zumer <johanna.zumer at donders.ru.nl>wrote:
>>
>>> Hi Akiko,
>>>
>>> In addition to Saskia's comment (which is very useful!) remember also
>>> that if you create the subject's grid from the warped MNI template grid
>>> (explained here
>>> http://fieldtrip.fcdonders.nl/example/create_single-subject_grids_in_individual_head_space_that_are_all_aligned_in_mni_space?s[]=template&s[]=grid)
>>> then you can keep each subject's source data in a 'source' structure with
>>> .pos field and still do group level statistics, without need to convert to
>>> 'volume' structure which upsamples the spatial resolution perhaps
>>> artificially too high.
>>>
>>> Cheers,
>>> Johanna
>>>
>>>
>>> 2012/10/7 Akiko Ikkai <akiko.ikkai at gmail.com>
>>>
>>>> Hi Saskia,
>>>>
>>>> Thank you for the quick advise! Removing cfg definitely works well.
>>>> Each of the group data is now 1.3G (I also used "single" to convert
>>>> everything into single precision), which is definitely manageable.
>>>> Computation time has also been reduced to 3 times less now.
>>>>
>>>> Thanks! Akiko
>>>>
>>>>
>>>> On Sun, Oct 7, 2012 at 2:16 PM, Saskia Haegens <shaegens at gmail.com>wrote:
>>>>
>>>>> Hi Akiko,
>>>>>
>>>>> In my experience with grandavg source structs, sometimes the cfg
>>>>> (that's attached to the data struct) becomes very large and can
>>>>> consume a considerable amount of memory. I'm not sure if that's the
>>>>> case/problem here, but might be worth checking and removing the cfg.
>>>>> You could even use checkconfig to cleanup your cfg with:
>>>>> data.cfg = ft_checkconfig(data.cfg, 'checksize', 'yes')
>>>>> Hope this helps!
>>>>>
>>>>> Best,
>>>>> Saskia
>>>>>
>>>>> On Sun, Oct 7, 2012 at 11:36 AM, Akiko Ikkai <akiko.ikkai at gmail.com>
>>>>> wrote:
>>>>> > Dear Fieldtrip users,
>>>>> >
>>>>> > I've been trying to run group stats on my EEG source data, which
>>>>> contains 14
>>>>> > subjects' normalized beamformer data, and having serious swap memory
>>>>> issue
>>>>> > (not Matlab memory issue, but OS swap memory).
>>>>> >
>>>>> > I'm trying to contrast 2 conditions (within subject design). Each
>>>>> subject's
>>>>> > normalized beamformer data (1 condition) is
>>>>> >
>>>>> > source_lTMI_intNorm =
>>>>> >
>>>>> > anatomy: [181x217x181 double]
>>>>> >
>>>>> > inside: [181x217x181 logical]
>>>>> >
>>>>> > avg: [1x1 struct]
>>>>> >
>>>>> > transform: [4x4 double]
>>>>> >
>>>>> > dim: [181 217 181]
>>>>> >
>>>>> > cfg: [1x1 struct]
>>>>> >
>>>>> >
>>>>> >>whos
>>>>> >
>>>>> >
>>>>> >
>>>>> > Name Size Bytes Class
>>>>> Attributes
>>>>> >
>>>>> > source_lTMI_intNorm 1x1 520114791 struct
>>>>> >
>>>>> >
>>>>> > therefore, when I open all subjects' data ("data1group" and
>>>>> "data2group"),
>>>>> > it's huge...
>>>>> >
>>>>> > Name Size Bytes Class Attributes
>>>>> >
>>>>> > data1group 1x14 5909429438 cell
>>>>> >
>>>>> > data2group 1x14 6705652782 cell
>>>>> >
>>>>> >
>>>>> > data1group & data2group are both 1x14 struct (1 cell/subject).
>>>>> Therefore,
>>>>> >
>>>>> >>data1group{1}
>>>>> >
>>>>> > anatomy: [181x217x181 double]
>>>>> >
>>>>> > inside: [181x217x181 logical]
>>>>> >
>>>>> > avg: [1x1 struct]
>>>>> >
>>>>> > transform: [4x4 double]
>>>>> >
>>>>> > dim: [181 217 181]
>>>>> >
>>>>> > cfg: [1x1 struct]
>>>>> >
>>>>> > So, when I try to run
>>>>> >
>>>>> > cfg=[];
>>>>> >
>>>>> > cfg.dim = data1group{1}.dim;
>>>>> >
>>>>> > cfg.method = 'montecarlo';
>>>>> >
>>>>> > cfg.statistic = 'depsamplesT';
>>>>> >
>>>>> > cfg.parameter = 'avg.pow';
>>>>> >
>>>>> > cfg.correctm = 'cluster';
>>>>> >
>>>>> > cfg.numrandomization = 100;
>>>>> >
>>>>> > cfg.alpha = 0.05;
>>>>> >
>>>>> > cfg.tail = 0;
>>>>> >
>>>>> > nsubj=length(data1group);
>>>>> >
>>>>> > cfg.design(1,:) = [1:nsubj 1:nsubj];
>>>>> >
>>>>> > cfg.design(2,:) = [ones(1,nsubj) ones(1,nsubj)*2];
>>>>> >
>>>>> > cfg.uvar = 1;
>>>>> >
>>>>> > cfg.ivar = 2;
>>>>> >
>>>>> > stat = ft_sourcestatistics(cfg, data1group{:}, data2group{:});
>>>>> >
>>>>> > stat.anatomy = data1group{1}.anatomy;
>>>>> >
>>>>> >
>>>>> > my computer (os 10.6.8, 6G memory) runs out of swap memory (startup
>>>>> > memory?), which forces me to quit Matlab. I'm running above
>>>>> processes in a
>>>>> > function, so I'm not running into Matlab memory error.
>>>>> >
>>>>> > Could someone help me how it could run more efficiently? I guess
>>>>> > cfg.inputfile is not available for ft_sourcestatistics, so I have to
>>>>> > eventually load 2 group data in Matlab workspace...?
>>>>> >
>>>>> > Thank you in advance! Akiko
>>>>> >
>>>>> > --
>>>>> > Akiko Ikkai, Ph.D.
>>>>> > Postdoctoral Fellow
>>>>> > Department of Psychological and Brain Sciences
>>>>> > Johns Hopkins University
>>>>> > Ames Hall, 3400 N. Charles St.
>>>>> > Baltimore, MD 21218
>>>>> >
>>>>> >
>>>>> >
>>>>> > _______________________________________________
>>>>> > fieldtrip mailing list
>>>>> > fieldtrip at donders.ru.nl
>>>>> > http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>>>> _______________________________________________
>>>>> fieldtrip mailing list
>>>>> fieldtrip at donders.ru.nl
>>>>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Akiko Ikkai, Ph.D.
>>>> Postdoctoral Fellow
>>>> Department of Psychological and Brain Sciences
>>>> Johns Hopkins University
>>>> Ames Hall, 3400 N. Charles St.
>>>> Baltimore, MD 21218
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> fieldtrip mailing list
>>>> fieldtrip at donders.ru.nl
>>>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>>>
>>>
>>>
>>> _______________________________________________
>>> fieldtrip mailing list
>>> fieldtrip at donders.ru.nl
>>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>>
>>
>>
>> _______________________________________________
>> fieldtrip mailing list
>> fieldtrip at donders.ru.nl
>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>>
>
>
>
> --
> Akiko Ikkai, Ph.D.
> Postdoctoral Fellow
> Department of Psychological and Brain Sciences
> Johns Hopkins University
> Ames Hall, 3400 N. Charles St.
> Baltimore, MD 21218
>
>
>
--
Akiko Ikkai, Ph.D.
Postdoctoral Fellow
Department of Psychological and Brain Sciences
Johns Hopkins University
Ames Hall, 3400 N. Charles St.
Baltimore, MD 21218
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20121019/a9a96637/attachment-0001.html>
More information about the fieldtrip
mailing list