Andrew Geng, Pin-Yu Chen
This repo contains the reference source code in PyTorch for Reprogrammer. For more details please check our paper Model Reprogramming Outperforms Fine-tuning on Out-of-distribution Data in Text-Image Encoders (SatML 24).
The code is built with the following libraries:
- To train reprogrammer, run
python -u full_reprogramming.py --name=reprogrammer --in-dataset=CIFAR-10 --mr-resolution=32 --gpu=0 >> ./reprogrammer.out
- To evaluate reprogrammer OOD Generalization, run
python -u evaluate_robustness.py --name=reprogrammer --in-dataset=CIFAR-10 --ood-dataset=CIFAR10.1 --image-resolution=32 --gpu=0 >> robustness_reprogrammer.out
- To evaluate reprogrammer OOD Generalization, run
python -u evaluate_ood.py --name=reprogrammer --in-dataset=CIFAR-10 --gpu=0 >> ./ood_detection_reprogrammer.out
If you find our code useful, please consider citing:
@inproceedings{
geng2024rp,
title={Model Reprogramming Outperforms Fine-tuning on Out-of-distribution Data in Text-Image Encoders},
author={Andrew Geng and Pin-Yu Chen},
booktitle={SatML},
year={2024}
}