Browse Teach Curate

Vision Transformer (ViT)

The architecture that applies a standard Transformer encoder to images by splitting an image into fixed-size patches, linearly projecting each patch into a token embedding, adding positional encodings and a class token, and passing the sequence through self-attention layers. Covers patch embedding, why ViTs need large-scale pre-training, and how the pre-trained encoder is reused for downstream tasks.

jeremy Video

Vision Transformer Quick Guide - Theory and Code in (almost) 15 min

The Vision Transformer extends the encoder-only architecture of sequential transformers to visual data by treating fixed-size image patches and a learned class token (CLS) as tokens within a sequence…