A Foolproof Way to Shrink Deep Learning Models