Less Attention is More: Prompt Transformer for Generalized Category Discovery (CVPR2025)

Less Attention is more: Prompt Transformer for Generalized Category Discovery (CVPR 2025)
By Wei Zhang, Baopeng Zhang*, Zhu Teng, Wenxin Luo, Junnan Zou, Jianping Fan

🏃 Running

Dependencies

We provide torch version and paddle version in this repo.

• Torch: The code was trained on python3.10 pytorch2.0.0 and CUDA11.7. You can install dependencies by

pip install -r requirements.txt

• Paddle: This version was trained on python3.10.9 paddlepaddle-gpu2.6.2.post112

Config

Set paths to datasets and desired log directories in config.py.

Datasets

• We trained on three generic datasets: CIFAR-10, CIFAR-100, ImageNet-100.
• We also used four fine-grained benchmarks: StanfordCars, CUB-200, FGVC-Aircraft and Herbarium19.

Scripts

Train the model (taking the Stanford Cars benchmark as an example).

bash scripts/run_cars.sh

Pretrain model

Download the pretrain models and put them in pretrain_model folder. Changing the path (such as /pretrain_model/scars.pt) in train.py if you use these models.

💡 Citing this work

If you find this repo useful for your research, please consider citing our paper:

@inproceedings{zhang2025less,
  title={Less Attention is More: Prompt Transformer for Generalized Category Discovery},
  author={Zhang, Wei and Zhang, Baopeng and Teng, Zhu and Luo, Wenxin and Zou, Junnan and Fan, Jianping},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={30322--30331},
  year={2025}
}

💜 Acknowledge

The code is largely built on : LegoGCD and SPTNet. Many thanks to their briliant works.