ViT

Overview

From-scratch implementation of ViT-B/16 (Vision Transformer). ViT applies a standard transformer encoder directly to sequences of 16×16 image patches, treating each patch as a token — demonstrating that CNNs are not required for competitive vision performance. Based on An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Dosovitskiy et al., 2021).

Architecture

• Patch size: 16×16 pixels
• Patch embeddings via linear projection
• Learnable [CLS] token for classification
• Standard transformer encoder (multi-head self-attention, MLP, LayerNorm)
• Classification head on [CLS] token output

Training

Results

Paper

An Image is Worth 16x16 Words — Dosovitskiy et al., 2021