Basic ViT implementation in PyTorch. Based on the paper "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" by Dosovitskiy et al. (2021) and Better plain ViT baselines for ImageNet-1k by Beyer et al. (2022).
All code is contained in the vit-implementation.ipynb file.
Achieves 80.74% accuracy in CIFAR-10 with 100 epochs, 8 layers, 8 heads, and 256 embedding dimensions. Disclaimer: Not much hyperparameter tuning was done.