Bug fixes V 1.02
  • added sensitivity and specificity to @loss functions


  • fixed bug in @feat_sel object (@chain construction in feat_sel.m didn't always work)


  • fixed printing output bugs with @rfe and @l0


  • implemented new 'from_data' kernel which allows you to put the kernel as the X component of the data, e.g d=gen(toy); d.X=d.X*d.X'; train(svm(kernel('from_data')),d);


  • fixed bug with custom kernel using svmlight (note: this hasn't been compiled for windows yet, only linux)


  • for custom kernel can now have empty x in supplied data, e.g d=gen(toy); get_mean(train(cv(svm(kernel('custom',d.X*d.X'))),data([],d.Y)))


  • @svm testing now tests in batches to avoid memory load of computing kernel


  • fixed bugs in @one_class_svm with poly and rbf kernels


  • sign is now taken automatically with class_loss if given unsigned output (but a warning is produced)


  • fixed bug not allowing test with third argument, e.g test(a,d,'class_loss')


  • fixed bug with indexing alphas, e.g command like: [r a]=train(svm,toy); a.alpha(1:5)