Transfer classification with time-frequency data without features

[marenwolf]

Hi Matthias, when I try to do cross decoding for time-frequency data (4D, samples x channels x frequencies x timepoints) with one dataset for training and one dataset for testing and specifying no feature dimension, I get the following error:

Error using reshape
Size vector must have at least two elements.

Error in squeeze1 (line 19)
X = reshape(X, sz);

Error in mv_classify (line 483)
            Xtest_ix = squeeze1(Xtest(sample_skip{:}, ix_test{:}, feature_skip{:}));

This is the code I used:

cfg =  [];
cfg.classifier = 'lda';
cfg.metric     = {'auc', 'accuracy'};
cfg.sample_dimension    = 1;
cfg.feature_dimension = [];
cfg.dimension_names = {'samples', 'channels', 'frequencies', 'timepoints'};
cfg.preprocess      = 'undersample';

[perf_cross, cross_result] = mv_classify(cfg, data1, clabels1, data2, clabels2);

The same code works with an older version of the MVPAlight toolbox that I downloaded on 25.03.2022.

Including these lines of codes in the squeeze1 function lets me run the analysis for now, but I am not sure if everything else then still works as intended:

if length(sz) == 1
    sz = [sz 1];
end

Do you have any suggestions on how to fix it?

Thank you! Maren

[treder]

Hi @marenwolf Apologies I didn't get to it earlier on - it's indeed a bug and your code fixes it, thank you ! I did some more testing and there's no unwanted side effects from what I see so I've committed the change.