[Nipy-devel] RE : Fwd: 1st/2nd level analysis modules in python

Alexis Roche alexis.roche@gmail....
Tue Nov 10 02:30:22 CST 2009

Salut Philippe,

Thanks for the feedback.

>  I just like to point out that your choice about the storage of the voxelwise covariance matrix
> usually generates MEMORY errors as soon as the number of regressors exceeds 20 columns in the
> design matrix [glm.py in the neurospin sub-branch].

Interesting. Did you run it on a standard PC? Can you specify its
memory performance?

> This afternoon, Bertrand told me that there exists an alternative strategy,
> which consists in storing separately the variance matrix and the autoregressive part of the
> covariance matrix. The latter could be compressed using for instance a discrete grid on which
> you could tabulate this autoregressive part. So my question is: have you any idea about when
> this could be done and available in the next release?

I believe the approach you are referring to has already been
implemented in nipy by Jonathan Taylor. Regarding my own algorithm, I
have other strategies in mind that are simpler in that they are purely
computer-based, I just need to make some time to implement them. No
precise idea when this could be done, hopefully quite soon.



More information about the Nipy-devel mailing list