This is an official implementation of CvT: Introducing Convolutions to Vision Transformers. We present a new architecture, named Convolutional vision Transformers (CvT), that improves Vision Transformers (ViT) in performance and efficienty by introducing convolutions into ViT to yield the best of both disignes. This is accomplished through two primary modifications: a hierarchy of Transformers containing a new convolutional token embedding, and a convolutional Transformer block leveraging a convolutional projection. These changes introduce desirable properties of convolutional neural networks (CNNs) to the ViT architecture (e.g. shift, scale, and distortion invariance) while maintaining the merits of Transformers (e.g. dynamic attention, global context, and better generalization). We validate CvT by conducting extensive experiments, showing that this approach achieves state-of-the-art performance over other Vision Transformers and ResNets on ImageNet-1k, with fewer parameters and lower FLOPs. In addition, performance gains are maintained when pretrained on larger dataset (e.g. ImageNet-22k) and fine-tuned to downstream tasks. Pre-trained on ImageNet-22k, our CvT-W24 obtains a top-1 accuracy of 87.7% on the ImageNet-1k val set. Finally, our results show that the positional encoding, a crucial component in existing Vision Transformers, can be safely removed in our model, simplifying the design for higher resolution vision tasks.

Main results

Models pre-trained on ImageNet-1k

Model Resolution Param GFLOPs Top-1
CvT-13 224x224 20M 4.5 81.6
CvT-21 224x224 32M 7.1 82.5
CvT-13 384x384 20M 16.3 83.0
CvT-32 384x384 32M 24.9 83.3

Models pre-trained on ImageNet-22k

Model Resolution Param GFLOPs Top-1
CvT-13 384x384 20M 16.3 83.3
CvT-32 384x384 32M 24.9 84.9
CvT-W24 384x384 277M 193.2 87.6

You can download all the models from our model zoo.

Quick start


Assuming that you have installed PyTroch and TorchVision, if not, please follow the officiall instruction to install them firstly.
Intall the dependencies using cmd:

python -m pip install -r requirements.txt --user -q

The code is developed and tested using pytorch 1.7.1. Other versions of pytorch are not fully tested.

Data preparation

Please prepare the data as following:

    | |-class1
    | | |-img1.jpg
    | | |-img2.jpg
    | | |-...
    | |-class2
    | | |-img3.jpg
    | | |-...
    | |-class3
    | | |-img4.jpg
    | | |-...
    | |-...
      | |-img5.jpg
      | |-...
      | |-img6.jpg
      | |-...
      | |-img7.jpg
      | |-...


Each experiment is defined by a yaml config file, which is saved under the directory of experiments. The directory of experiments has a tree structure like this:

| |-{ARCH_A}
| |-{ARCH_B}
| |-{ARCH_A}
| |-{ARCH_B}
| |-{ARCH_A}
| |-{ARCH_B}

We provide a script for running jobs in local machine.

Usage: [run_options]
  -g|--gpus <1> - number of gpus to be used
  -t|--job-type <aml> - job type (train|test)
  -p|--port <9000> - master port
  -i|--install-deps - If install dependencies (default: False)

Training on local machine

bash -g 8 -t train --cfg experiments/imagenet/cvt/cvt-13-224x224.yaml

You can also modify the config paramters by the command line. For example, if you want to change the lr rate to 0.1, you can run the command:

bash -g 8 -t train --cfg experiments/imagenet/cvt/cvt-13-224x224.yaml TRAIN.LR 0.1


  • The checkpoint, model, and log files will be saved in OUTPUT/{dataset}/{training config} by default.

Testing pre-trained models

bash -t test --cfg experiments/imagenet/cvt/cvt-13-224x224.yaml TEST.MODEL_FILE ${PRETRAINED_MODLE_FILE}


If you find this work or code is helpful in your research, please cite:

  title={Cvt: Introducing convolutions to vision transformers},
  author={Wu, Haiping and Xiao, Bin and Codella, Noel and Liu, Mengchen and Dai, Xiyang and Yuan, Lu and Zhang, Lei},
  journal={arXiv preprint arXiv:2103.15808},