CvT

Why we included this project

CvT is Microsoft's official implementation of a vision transformer that folds convolutions into the standard ViT design, picking up the shift, scale, and distortion invariance CNNs are known for while keeping the global attention that makes transformers work. The paper's headline result is that this hybrid beats plain ViTs and ResNets on ImageNet-1k while using fewer parameters and FLOPs, and the repo ships the exact training and evaluation code behind those numbers, organized around YAML configs and a run script. Pretrained checkpoints for the CvT-13, CvT-21, and CvT-W24 variants are included, so you can reproduce the published results or fine-tune a model to your own dataset without reimplementing anything. It is a research codebase rather than a pip-installable library, so expect to work directly in PyTorch and adapt the training loop to your setup. If you are comparing efficient vision backbones or want a solid starting point for classification experiments, the pretrained weights and clean experiment structure make it worth a look.