In　recent　years,　the　deep　neural　network　has　achieved　astonishing　results　in　the　automatic　Whole　slide　images(WSI)　processing.　The　current　mainstream　approach　is　inseparable　from　a　large　number　of　manual　annotations.　However,　labeling　such　giant　images　with　billions　of　pixels　is　very　labor-intensive.　So　the　shortage　of　annotation　has　become　a　bottleneck　in　developing　Whole　slide　image　diagnostic　models.　Therefore,　we　propose　a　new　self-supervised　learning(SSL)　network　to　solve　the　problem　of　insufficient　annotation.　In　our　work,　massive　semantic　information　can　be　extracted　from　a　large　number　of　WSI,　which　significantly　gets　rid　of　our　dependence　on　the　label.　At　the　same　time,　the　results　are　further　refined　by　a　silhouette-coefficient-based　recursive　Spectral　Clustering　Bipartition,　which　significantly　improved　the　classification　accuracy.　Moreover,　our　framework　is　highly　transferable　and　can　take　on　many　downstream　tasks　in　pathology.　Our　final　results　are　verified　on　the　NCT-CRC-100K　and　MSHIR　datasets.　Our　code　is　available　at　https://github.com/Hongbo-Chu/generative-contrastive　©　2023　ACM.