AN IMAGE IS WORTH 16X16 WORDS- TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE
https://www.youtube.com/watch?v=j3VNqtJUoz0&t=16s
核心思想：
- 将图像分为patches, 线性映射, 再加上图片的position embeding来输入transformer encoder
- 额外使用一个cls token用于占位（ViT的输出就是这个cls input token对应的output token）
AN IMAGE IS WORTH 16X16 WORDS- TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE