1 function example_djm_partitions()
   2   %{
   3   %%%%%%%%%%%%%%%%%%%%%
   4   djm: This shows a simple group analysis for the vt mask of the AK6 dataset.
   5   It illustrates three types of partition: single split-half (e.g. oddeven), nfold
   6   (aka leave-one-out), and repeated split-half.
   7   %%%%%%%%%%%%%%%%%%%%%
   8   %}
   9 
  10   %% Define data
  11   config=cosmo_config();
  12   subs={'s01','s02','s03','s04','s05','s06','s07','s08'};
  13   partitions={'odd_even', 'nfold', 'full_split_half'};
  14 
  15   rng(0); % reset random number generator for reproducability
  16 
  17   figure(9); clf(9);
  18 
  19   result=nan( numel(subs), numel(partitions) ); % djm, preallocate array
  20   for s=1:numel(subs)
  21     fprintf('.')
  22     data_path=fullfile(config.tutorial_data_path,'ak6',subs{s});
  23     data_fn=fullfile(data_path,'glm_T_stats_perrun.nii');
  24     mask_fn=fullfile(data_path,'vt_mask.nii');
  25     ds=cosmo_fmri_dataset(data_fn,'mask',mask_fn,...
  26     'targets',repmat(1:6,1,10),... % 6 types of animal
  27     'chunks',floor(((1:60)-1)/6)+1); % 10 runs
  28 
  29     % remove constant features (due to liberal masking)
  30     ds=cosmo_remove_useless_data(ds);
  31 
  32     % Assign a function handle to the cosmo_crossvalidation_measure
  33     measure=@cosmo_crossvalidation_measure;
  34 
  35     % Make a struct containing the arguments for the measure:
  36     args=struct();
  37     args.classifier=@cosmo_classify_nn;
  38 
  39     for p=1:numel(partitions)
  40 
  41       switch partitions{p}
  42         case 'odd_even'
  43           args.partitions=cosmo_oddeven_partitioner(ds,'half');
  44         case 'nfold'
  45           args.partitions=cosmo_nfold_partitioner(ds);
  46         case 'full_split_half'
  47           args.partitions=cosmo_nchoosek_partitioner(ds,0.5);
  48         otherwise
  49           error('unknown partition type')
  50       endswitch
  51       cosmo_check_partitions(args.partitions,ds);
  52 
  53       % Apply the measure to ds, with args as second argument.
  54       tic
  55       ds_accuracy=measure(ds,args);
  56 
  57       % for subject 1, plot partitions, and how long this scheme took
  58       if s==1
  59         subplot(2,2,p)
  60         plot_partitions(args.partitions);
  61         title(sprintf('%s; tooks %.1f s per subject',partitions{p},toc),'interpreter','none')
  62         drawnow
  63       endif
  64 
  65       % store the result as accuracy minus chance (so positve values indicate information present)
  66       result(s,p)=ds_accuracy.samples-1/6;
  67 
  68     end % next partition scheme
  69   end % next subject
  70 
  71   % Run ttest versus chance (now zero)
  72   [h, p, ci, stat]=ttest(result);
  73 
  74   % plot results with 95% confidence intervals
  75   subplot(2,2,4)
  76   errorbar(mean(result),range(ci)./2,'o')
  77   set(gca,'xtick',1:numel(partitions),'xticklabel',partitions)
  78   ylabel('accuracy minus chance')
  79   ylim([-0.2 0.5]);
  80   hold on
  81   plot(xlim,[0 0],'--')
  82   drawnow
  83 
  84   return
  85 
  86   function plot_partitions(partitions)
  87     % djm: plot partitions like in documentation
  88     maxs=max(cellfun(@max,[ partitions.train_indices partitions.test_indices]));
  89     nc=numel( partitions.test_indices);
  90     out=zeros(maxs,nc);
  91     for c=1:nc
  92       out(partitions.train_indices{c},c)=1;
  93       out(partitions.test_indices{c},c)=2;
  94     endfor
  95     imagesc(out);
  96     if nc<21
  97       set(gca,'xtick',1:nc)
  98     else
  99       set(gca,'xtick',[1 nc])
 100     endif
 101     xlabel('classifiers')
 102     ylabel('samples')
 103     colormap([1 1 1; 0 1 0; 1 0 0]);
 104     caxis([0 2])
 105     return