DynamicViT: Efficient Vision Transformers with Dynamic Token S | Data Science Digest

DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification

In this research, Yongming Rao et al. propose a dynamic token sparsification framework to prune redundant tokens progressively and dynamically based on the input. A lightweight prediction module can estimate the importance score of each token given the current features. The module is added to different layers to prune redundant tokens hierarchically.

Web Page — https://bit.ly/3dgESlq
Paper — https://bit.ly/3xWf5XJ
Code — https://bit.ly/3jlUtUK