The official implementation of HiDe-Prompt (NeurIPS 2023, Spotlight).

• Python 3.6+
  pip install -r requirements.txt

Our code has been tested on four datasets: CIFAR-100, ImageNet-R, 5-Datasets, and CUB-200:

• CIFAR-100
• Imagenet-R
• 5-Datasets (including SVHN, MNIST, CIFAR10, NotMNIST, FashionMNIST)
• CUB-200

We incorporated the following supervised and self-supervised checkpoints as backbones:

• Sup-21K VIT
• iBOT-21K
• iBOT
• MoCo v3
• DINO

Please download the self-supervised checkpoints and put them in the /checkpoints/{checkpoint_name} directory.

To reproduce the results mentioned in our paper, execute the training script in /training_script/{train_{dataset}_{backbone}.sh}.

If you encounter any issues or have any questions, please let us know.

This repository is developed based on the PyTorch implementation of Dual-Prompt.

If you find this code helpful to your work, please cite our paper:

@article{wang2023hide,
  title={Hierarchical Decomposition of Prompt-Based Continual Learning: Rethinking Obscured Sub-optimality},
  author={Wang, Liyuan and Xie, Jingyi and Zhang, Xingxing and Huang, Mingyi and Su, Hang and Zhu, Jun},
  journal={Advances in Neural Information Processing Systems},
  year={2023}
}