Reduced the large memory requirements of ml_trainproximal.

Author: chkothe

code/machine_learning/ml_trainproximal.m

@@ -806,10 +806,9 @@
 
 function M = operator_to_matrix_cached(op,n)
 vec = @(x)x(:);
-M = [];
+M = sparse([]);
 w = zeros(n,1);
-for c=n:-1:1
+for c=[n 1:n-1]
     v = w; v(c) = 1;
     M(:,c) = vec(op(v));
 end
-M = sparse(M);

dependencies/rica/rica.m

@@ -115,10 +115,10 @@
 Mu = median(X,2);
 Sigma = hlp_diskcache('filterdesign',@cov_blockgeom,X',params.cov_blocksize);
 %% Optionally reject bad samples
-if params.cov_rejection
+if params.cov_rejection && exist('reject_samples_robcov','file')
     X = reject_samples_robcov(X,params.cov_rejection,Mu,Sigma); end
 %% Whiten the data
-X = bsxfun(@minus,X,mu);
+X = bsxfun(@minus,X,Mu);
 mix = sqrtm(Sigma);
 sphere = inv(mix);
 X = sphere*X; %#ok<*MINV>
@@ -131,7 +131,7 @@
 
 params.sphere = sphere;
 params.mix = mix;
-params.rcov = rcov;
+params.rcov = Sigma;
 
 %% Select cost function
 if strcmp(params.dict_criterion,'reconstruction')

userscripts/tutorial_erp1.m

@@ -140,7 +140,7 @@
     'Prediction',{'MachineLearning',{'Learner',{'dal',2.^(10:-0.125:1),'scaling','std','regularizer','glc','shape',[64 NaN]}}}};
 % again wavelets, using group-sparse logistic regression and an elastic-net like extra penalty (using the proximal classifier)
 % cross-validdation takes ca. 8hrs; demo only for syntax
-approaches.waveletsprox = {'DataflowSimplified' 'SignalProcessing',{'Resampling',100,'EpochExtraction',[0 0.8],'WaveletTransform','on'}, ...
+approaches.waveletsprox = {'DataflowSimplified' 'SignalProcessing',{'Resampling',128,'EpochExtraction',[0 0.8],'WaveletTransform','on'}, ...
     'Prediction',{'FeatureExtraction',{'GroupInto','matrix'}, ...
         'MachineLearning',{'Learner',{'proximal' 'Regularizers',{'Term1','l1/l2','Term2','l2'}}}}};