ICCV2021, Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet - View it on GitHub