Vision Transformer

Vision Transformer (ViT) model pre-trained on ImageNet-21k (14 million images, 21,843 classes) at resolution 224x224. It was introduced in the paper An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale and further enhanced in the follow-up paper How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers. The weights were converted from the B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0.npz file in GCS buckets presented in the original repository.