[pymvpa] Searchlight / feature selection crash

[Johan Carlin]

Hi guys,
I keep getting a strange crash with my searchlight analysis. This only seems
to happen for certain subjects out of a large sample, and only happens when
I combine the searchlight with sensitivity-based feature selection. The
searchlight starts to run, and then crashes with:

TypeError                                 Traceback (most recent call last)

/imaging/jc01/Becky01/MVPA/B01_MVPA.py in <module>()
    121                         print 'run finished without errors.'
    122                 except: print 'run finished with some errors.
Re-run?'
--> 123         else: map(sl_run,tasklist)
    124
    125

/imaging/jc01/UtilityScripts/Python/JohansMVPA.py in sl_run(T)
   1078         print 'running %s' % T['out']
   1079         set_trace()
-> 1080         sl_map = sl(ds)
   1081 # 1-(1/nconds) - error
   1082         sl_map_r = [(1.-(1./len(T['conds']))) - i for i in sl_map] #
chance = 0

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/measures/base.pyc
in __call__(self, dataset)
    103
    104
--> 105
    106
    107

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/measures/searchlight.pyc
in _call(self, dataset)
    106
    107
--> 108
    109
    110

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/measures/base.pyc
in __call__(self, dataset)
    103
    104
--> 105
    106
    107

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/algorithms/cvtranserror.pyc
in _call(self, dataset)
    171
    172
--> 173
    174
    175

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/clfs/transerror.pyc
in __call__(self, testdataset, trainingdataset)
   1281
   1282
-> 1283
   1284
   1285

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/clfs/transerror.pyc
in _precall(self, testdataset, trainingdataset)
   1237
   1238
-> 1239
   1240
   1241

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/clfs/base.pyc
in train(self, dataset)    352
    353
--> 354
    355
    356

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/clfs/meta.pyc
in _train(self, dataset)
   1056
   1057
-> 1058
   1059
   1060

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/featsel/base.pyc
in __call__(self, dataset, testdataset)
    106
    107
--> 108
    109
    110

/imaging/jc01/PythonInstall/lib/python2.5/site-packages/mvpa/featsel/helpers.pyc
in __call__(self, seq)
    400
    401
--> 402
    403
    404

TypeError: object of type 'numpy.float64' has no len()
WARNING: Failure executing file: <B01_MVPA.py>

----

These empty code snippets are a bit hard to decipher. :) Relevant bits of
code:

# Configure classifier, crossvalidation, searchlight
    clfcmd = eval(T['clf']+'()') # CSVMC
    if T['FS']:
        clfcmd = FeatureSelectionClassifier(clfcmd,
            SensitivityBasedFeatureSelection(
                clfcmd.getSensitivityAnalyzer(transformer=abs),
                FixedNElementTailSelector(T['FS_limit'],
mode='select',tail='upper'))) # select 50 features

    cv = CrossValidatedTransferError(TransferError(clfcmd),
             NFoldSplitter(), #OddEvenSplitter(),
             enable_states = ['results'])

    sl = Searchlight(cv,radius=T['radius']) # 10 mm

# Rescore labels
    new_labels = []
    for l in T['labels']:
        if not l in T['conds']: new_labels.append(0.5)
        else: new_labels.append(l)

# Load data
    ds = NiftiDataset(samples=T['sub4D'],
        labels=new_labels,
        chunks=T['chunks'],
        mask=T['submask'])

# Remove irrelevant labels
    zerocheck = array([v != 0.5 for v in new_labels],dtype='bool')
    ds = ds.selectSamples(zerocheck)

# Preprocessing
    ds = removeInvariantFeatures(ds)
    zscore(ds,perchunk=False)

# Run, rescale, and save
    print 'running %s' % T['out']
    #set_trace()
    sl_map = sl(ds)

--------

For the analysis that crashes, the dataset looks like this:

 ds
Out[87]: <Dataset / float32 8 x 21583 uniq: 4 chunks 2 labels>
(Pdb) ds.C
Out[87]: array([ 1.,  2.,  3.,  4.,  1.,  2.,  3.,  4.])
(Pdb) ds.L
Out[87]: array([  60.,   60.,   60.,   60.,  300.,  300.,  300.,  300.])
(Pdb) shape(ds.S)
Out[87]: (8, 21583)

-------

Once again, the exact same classification problem works fine for all other
subjects. I assume the problem comes in with feature selection because a
previous run without it finished with no errors.

I'm a bit stumped by this. Let me know if anyone has any ideas!

Johan




-- 
Johan Carlin
Graduate Student
MRC Cognition & Brain Sciences Unit
15 Chaucer Road
Cambridge CB2 7EF
UK

johan.carlin at mrc-cbu.cam.ac.uk
+44 (0)1223 355294 ext 593
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20090902/41716f70/attachment.htm>