30% trainning, unsupervised with linear classfier