This repository contains the code for our paper Enhancing Vision-Language Few-Shot Adaptation with Negative Learning.

We test our codebase with PyTorch 1.12.1 with CUDA 11.6. Please install corresponding PyTorch and CUDA versions according to your computational resources. Then install the rest of required packages by running pip install -r requirements.txt.

Please follow DATASET.md to download all the 11 datasets we used for experiments. We adapt this from CoOp.

You can extract the features by running CUDA_VISIBLE_DEVICES=0 python extract_features.py.

After running, you can get all the image features from tran/val/test set, as well as the positive/negative textual features in caches/[dataset_name].

You can simply run CUDA_VISIBLE_DEVICES=0 python main.py --config configs/[dataset_name].yaml --shot [shot_number] to train and test the SimNL model.

Here, dataset_name should be one of [caltech101, dtd, eurosat, fgvc, food101, imagenet, oxford_flowers, oxford_pets, stanford_cars, sun397, ucf101], and shot_number is chosen from 1/2/4/8/16.

Our codebase is adapted from Tip-Adapter, CLIP, APE, and CuPL. We thank the authors for releasing their code!

If you have any questions, please contact at [email protected].

If you find this code useful, please consider citing our work:

@article{zhang2024enhancing,
  title={Enhancing Vision-Language Few-Shot Adaptation with Negative Learning},
  author={Zhang, Ce and Stepputtis, Simon and Sycara, Katia and Xie, Yaqi},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={5905-5915}
  year={2025}
}