scaling is applied before cross validation

[weishen89]

Scaling should applied separately on the training set and the test set (using the mean and sd from the training set).
The current code provides an over-optimistic result.