TVT: Transferable Vision Transformer for Unsupervised Domain Adaptation

Nov 27, 2021 1 min read

TVT

Code of TVT: Transferable Vision Transformer for Unsupervised Domain Adaptation

Pretrained ViT

Download ViT-B_16.npz and put it at checkpoint/ViT-B_16.npz

Datasets:

Download data and replace the current data/
Download images from Office-31, Office-Home, VisDA-2017 and put them under data/. For example, images of Office-31 should be located at data/office/domain_adaptation_images/

Training:

All commands can be found in script.txt. An example:

python3 main.py --train_batch_size 64 --dataset office --name wa \
--source_list data/office/webcam_list.txt --target_list data/office/amazon_list.txt \
--test_list data/office/amazon_list.txt --num_classes 31 --model_type ViT-B_16 \
--pretrained_dir checkpoint/ViT-B_16.npz --num_steps 5000 --img_size 256 \
--beta 0.1 --gamma 0.01 --use_im --theta 0.1

Citation:

@article{yang2021tvt,
  title={TVT: Transferable Vision Transformer for Unsupervised Domain Adaptation},
  author={Yang, Jinyu and Liu, Jingjing and Xu, Ning and Huang, Junzhou},
  journal={arXiv preprint arXiv:2108.05988},
  year={2021}
}

Our code is largely borrowed from CDAN and ViT-pytorch

John was the first writer to have joined pythonawesome.com. He has since then inculcated very effective writing and reviewing culture at pythonawesome which rivals have found impossible to imitate.