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