The optimization goal of the SVM is to maximize the distance between the positive and negative hyperplanes. But before optimizing, how sklearn first find the positive and negative support vectors and therefore the margin hyperplanes?