This repository is the official implementation of the BMVC2024 paper Prompt Generation Networks for Input-Space Adaptation of Frozen Vision Transformers by Jochem Loedeman, Maarten Stol, Tengda Han and Yuki M Asano.

To install python dependencies, make sure that poetry is installed and execute the following in the project root directory:

poetry install

See data/README.md

Download the full checkpoint for DINO ViT-S/16 from here and insert it as pgn/pgn_models/dino/dino_deitsmall16_pretrain_full_checkpoint.pth.

To train/test with the CLIP backbone, run

poetry run train_clip
poetry run test_clip

To train/test with either DINO or supervised ViT, specify the backbone with --vision_model_type and run

poetry run train_visionmodel
poetry run test_visionmodel

For all available command line arguments, see pgn/scripts.

Pretrained PGNs are supplied in pretrained_pgns/. To use them in the context of this repository, specify the desired model by setting the --pgn_path argument in the test scripts.

If you find this repository is useful for your project, please consider citing our paper:

@article{Loedeman2022prompt,
    author       = "Jochem Loedeman and Maarten Stol and Tengda Han and Yuki M Asano",
    title        = "Prompt Generation Networks for Input-based Adaptation of Frozen Vision Transformers",
    journal      = "arxiv preprint arxiv:2210.06466",
    year         = "2022",
}