ViT- an image is worth 16x16 words- transformers for image recognition at scale

 Google ICLR 2021

输入：图像

网络结构：

1. 图像拆patch (bs, num_patch, dim) to_patch_embedding
   1. Rearrange 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width
   2. norm
   3. Linear
   4. norm
2. 拼接类别token (bs, num_patch+1, dim)
3. 加位置embedding（矩阵点加）
4. transformer网络
   1. 堆叠att和ffn
      1. att：
         1. norm- selfAtt- Linear- dropout
      2. ffn：norm- Linear- act- dropout- Linear- dropout
         1. norm：LayerNorm
         2. act：GELU