[pymvpa] Bug in regression classifiers

[Per B. Sederberg]

Just a follow-up.  I have not solved the bug (and I'm not sure I can
because it's deep in Yarik/Michael code), but I have created a unit
test for it in test_clf.py:

testFeatureSelectionClassifierWithRegression

So, that's a place to start!

Thanks,
Per


At Sat, 8 Nov 2008 15:36:24 -0500,
Per B. Sederberg wrote:
> 
> Hi Folks:
> 
> I've sort of tracked down a bug in regression classifiers when used as
> part of a feature selection classifier wherein the code that saves the
> results as the values and then converts the results into predictions
> gets called twice, which means that the results (which are values for
> regressions) get turned into predictions and then copied to the
> values, so that there are no values left.  You can get reproduce the
> error using this classifier:
> 
> # define some classifiers
> anova_keep = 1000
> tokeep = 25
> 
> # ANOVA feature selection
> anova_select = SensitivityBasedFeatureSelection(
>     OneWayAnova(),
>     FixedNElementTailSelector(anova_keep,mode='select',tail='upper'))
> 
> # LARS classifier
> lars = LARS(trace=False, max_steps=tokeep)
> lars.states.enable('values')
> #lars.states.disable('trained_labels')
> anova_lars = FeatureSelectionClassifier(
>     lars,
>     anova_select,
>     descr="LinSVM on %d(ANOVA)" % (tokeep))
> anova_lars.states.enable('values')
> 
> 
> I'm hoping someone knows a good way to fix this...
> 
> Thanks,
> Per