I replicated the Vision Transformer (ViT) using Python and PyTorch, absolutely from scratch. This project involves building the Vision Transformer architecture without relying on pre-built modules, providing a deeper understanding of how this state-of-the-art model works.
The Vision Transformer (ViT) divides an image into patches, applies a transformer to these patches, and then predicts labels. This implementation includes:
- Patch Embedding: Splits images into patches and embeds them.
- Transformer Encoder: Applies multi-head self-attention and feed-forward layers.
- Classification Head: Final dense layers for classification tasks.
Original paper: https://arxiv.org/abs/2010.11929
