lucidrains/vit-pytorch

Integrates multiple Vision Transformer variants (NaViT, CaiT, Token-to-Token, CrossFormer, MobileViT, etc.) with self-supervised learning methods like masked autoencoders and DINO, enabling flexible research across architecture improvements and training paradigms. Supports variable-resolution batch processing through NaViT, 3D video inputs via ViVit, and knowledge distillation from convolutional teachers, all with clean PyTorch APIs for customizing patch size, depth, and attention heads.

24,988 stars. Used by 2 other packages. Actively maintained with 2 commits in the last 30 days. Available on PyPI.