Learning Disentangled Representation by Exploiting Pretrained Generative Models: A Contrastive Learning View