Hi Darren - just to add to Tim's email - you should definitely be using
3.1 and not 3.0 - there were some slight accuracy improvements in FLAME ,
and also 3.1 does include the "multi-threading"; in 3.1 the feat.tcl
script will process each slice separately (which gives the same results
because FLAME is voxelwise) if you have a sufficient number of subjects in
the analysis - you could adapt that loop easily to run the slices in
parallel on different machines.
Cheers, Steve.
On Wed, 3 Mar 2004, Tim Behrens wrote:
> Hey Darren -
>
> Whether you can split up the second level copes, depends on the design. In
> general, it's not a good idea. However you _can_ split up the first
> level copes so, in your case, you can run 9 times faster ( a single
> .gfeat for each first level cope ) and all should be good.
>
> You can also spread out the load spatially. At the flame stage (before
> Random fields/clusters etc.) everything is done voxel-by-voxel, so you can
> split them as much as you want and then merge them at the end before doing
> thresholding.
>
>
> I don't believe there is any multi-threaded code in FSL3.0 or 3.1 so I
> don't think any multi-processor stuff wil happen without you doing it
> yourself!
>
> wrt using stage-1 only. The "party line" advice is to use stage-1
> to check things out, but to use the total flame when you're
> publishing. Just depends on how patient you're feeling I guess. I think
> the results will be pretty similar...
>
>
> Hope this helps
>
> T
>
>
>
>
>
> -------------------------------------------------------------------------------
> Tim Behrens
> Centre for Functional MRI of the Brain
> The John Radcliffe Hospital
> Headley Way Oxford OX3 9DU
> Oxford University
> Work 01865 222782
> Mobile 07980 884537
> -------------------------------------------------------------------------------
>
> On Tue, 2 Mar 2004, Darren Schreiber wrote:
>
> > I am using FSL 3.0 to conduct a second level analysis of my "pictures
> > experiment" data. I have nine contrasts done on the first level and
> > twelve contrasts on the second level. Generating one cope1.gfeat
> > directory on the second level is taking about 24 hours of processing.
> > So, doing the whole analysis looks like it should take about nine days
> > (we are on day 3 now.)
> >
> > I am going to be doing the second level analysis of my "faces" data
> > later this week and was thinking of ways to speed it up. Since I am
> > running processes on a clustered machine, I was thinking that I could
> > generate twelve .fsf files to represent each of my second level
> > contrasts separately. Then I was thinking that I could run them all in
> > parallel and reduce the processing time to 24 hours or less total. Is
> > there anything that would prevent me from running multiple, parallel
> > second level contrasts?
> >
> > I'm also going to see if we can get FSL 3.1 on this machine. I vaguely
> > recall that 3.1 might have the capacity to split up processes into
> > threads that could be done in parallel automatically? Is that right?
> >
> > And, I was thinking of using just the 1st stage FLAME in FSL 3.1 due to
> > the speed up and small accuracy sacrifice. Is my thinking here good?
> >
> > As always, thanks for the help with this.
> >
> > Darren
> >
>
Stephen M. Smith DPhil
Associate Director, FMRIB and Analysis Research Coordinator
Oxford University Centre for Functional MRI of the Brain
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
|