DynamicViT

Why we included this project

Vision transformers burn a lot of compute, and DynamicViT tackles that directly by learning to drop redundant image tokens as they move through the network, so later layers only process the patches that matter for a given input. The authors report cutting FLOPs by over 30% and raising throughput by more than 40% while keeping accuracy loss under 0.5%, and the repo ships pretrained ImageNet models for DeiT and LVViT backbones. The same sparsification idea carries over to ConvNeXt and Swin transformers, plus object detection and semantic segmentation, in the T-PAMI version of the paper. This is research code rather than a packaged library, so expect to adapt the training scripts to your own pipeline. If you are exploring efficient vision backbones or want to build on token sparsification, the PyTorch implementation is a solid reference to start from.