add EsViT, by popular request, an alternative to Dino that is compati…

…ble with efficient ViTs with accounting for regional self-supervised loss

add EsViT, by popular request, an alternative to Dino that is compati…

…ble with efficient ViTs with accounting for regional self-supervised loss

fix mbconv residual block

fix maxvit - need feedforwards after attention

maxvit intent to build (lucidrains#211)

complete hybrid mbconv + block / grid efficient self attention MaxViT

let windowed tokens exchange information across heads a la talking he…

…ads prior to pointwise attention in sep-vit

cleanup sepvit

intent to build (lucidrains#210)

complete SepViT, from bytedance AI labs

do not layernorm on last transformer block for scalable vit, as there…

… is already one in mlp head

complete learnable memory ViT, for efficient fine-tuning and potentia…

…lly plays into continual learning