The
  authors of SegFormer first pre-trained the Transformer encoder on ImageNet-1k to classify images.