35. Pyramid Vision Transformer : A Versatile Backbone for Dense Prediction without Convolutions