[pymvpa] SVD did not converge

[andrea bertana]

Dear PyMVPA experts,

I'm trying to use Hyperalignment() procedure to align different subjects'
brain.

I am mainly referring to the example described in this webpage -
http://dev.pymvpa.org/examples/hyperalignment.html

However, when I try to compute the common space on the training set (10
participant, single - participant matrix: 301 time-point x 1000 voxels ) I
get the following message:

LinAlgError, 'SVD did not converge'

To the best of my knowledge, singular value decomposition could fail for
several reasons, here is what I checked:

   -

   Matrices have no Nan's.
   -

   Standard deviation for each voxel over time is > 0, meaning that there
   are no zero or constant voxels.
   -

   By aligning each pair of participants independently, I couldn't find any
   specifically 'compromised' participant, as far as no clear pattern emerged
   (depending on the pairs the SVD could-couldn't converge).


Further informations:

   -

   Data come from Human Connectome Project (HCP)- Working memory task
   - We tried also with another task (motor) that comes from HCP using the
   same subjects but we get a same error
   -

   The ds_all container with participants' information (samples and
   features):
   http://nilab.cimec.unitn.it/people/vittorioiacovella/ds_all.pickle
   -

   Please, find below the code I am using for this task.


import numpy as np

from mvpa2.suite import *

import pickle as pk

import sys

#Insert all the part of loading ds_all

path=sys.argv[1]

ds_all = pk.load(open(path))

### DEFINE THE CLASSIFIER

## use same linear support vector machine

clf = LinearCSVMC()

# STARTING THE HYPERALIGNMENT

verbose(2, "between-subject (hyperaligned)...", cr=False, lf=False)

hyper_start_time = time.time()

bsc_hyper_results = []

#cross-validation over subjects

cv = CrossValidation(clf, NFoldPartitioner(attr='subject'),

                errorfx=mean_match_accuracy)



    # HYPERALIGNMENT

    # - Leave-one-run-out for hyperalignment training

nruns = 1,2

for test_run in nruns:

   #Split in training and testing set:

   # For leave one run out, use this:

   ds_train = [sd[sd.sa.chunks != test_run,:] for sd in ds_all]

   ds_test = [sd[sd.sa.chunks == test_run,:] for sd in ds_all]

   #Defining hyper-function and computing hyperalignment parameters.

   hyper = Hyperalignment()

   hypmaps = hyper(ds_train)

   #Applying hyperalignment parameters on the test set (the run left out)

   ds_hyper = [ hypmaps[i].forward(sd) for i, sd in enumerate(ds_test)]

   # zscore each subject individually after transformation for optimal
performance

   zscore(ds_hyper, chunks_attr='subject')

   #Encoding by simple avaraging each block after all subjects data are in
common space.

   averager = mean_group_sample(['targets', 'block', 'chunks'])

   ds_hyper_encoded = [sd.get_mapped(averager) for sd in ds_hyper]

   ds_hyper = vstack(ds_hyper_encoded)

   #Store results of classification for different partitions.

   res_cv = cv(ds_hyper)

   #Append results on a list (each element comes from one run)

   bsc_hyper_results.append(res_cv)

    #Final results

bsc_hyper_results = hstack(bsc_hyper_results)

#Showing final results.

verbose(2, "done in %.1f seconds" , (time.time() - hyper_start_time,))

verbose(2, "between-subject (hyperaligned): %.2f +/-%.3f" \

    , (np.mean(bsc_hyper_results),

      np.std(np.mean(bsc_hyper_results, axis=1)) / np.sqrt(nsubjs - 1)))


Thank you.

Best,
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20140402/aaf65784/attachment-0001.html>