I think Dr. Lin said that you should use the same factor
for scaling training and testing data, i.e.,
svm-scale -s range ijcnn1 > a
svm-scale -r range ijcnn1.t > b
Is that true?