BiFormer

Why we included this project

BiFormer is a CVPR 2023 vision transformer built around bi-level routing attention: instead of attending over the whole image, the model first picks a small set of relevant regions and then runs attention only inside them. That keeps accuracy high while cutting compute, which is the main reason people building efficient backbones reach for it. The repo ships pretrained ImageNet-1K checkpoints in several sizes plus working code for object detection and semantic segmentation, so you can drop the backbone into downstream tasks rather than reimplementing it. It builds on timm and documents training and evaluation for both slurm clusters and single machines. This is a research release rather than a maintained library, so expect to adapt it for production, but the attention module is a useful starting point for custom vision models.