Regularized Mask Tuning: Uncovering Hidden Knowledge in Pre-trained Vision-Language Models

October 19, 2023 · View on GitHub

This repository provides a new type of tuning method, termed as regularized mask tuning, which masks the network parameters through a learnable selection.

Overview of methods supported by our method:

[Paper]

Code Coming Soon

Release training code.
Release pretrained models.

Experiments:

Method	ImageNet	Caltech101	FGVCAircraft	StanfordCars	Flowers102	OxfordPets	Food101	DTD	EuroSAT	UCF101	SUN397	Average	Gain
Zero-shot CLIP	66.73	92.94	24.72	65.32	71.34	89.21	86.06	44.39	47.60	66.75	62.50	65.23	-
R-AMT	73.07	97.00	58.47	85.93	98.17	93.80	87.47	74.57	91.80	86.93	76.40	83.96	+18.73
CoOP	72.01	95.47	43.29	82.91	96.93	91.92	84.33	69.21	86.05	82.25	74.58	79.90	-
CoOP+R-AMT	73.35	96.70	56.37	85.63	97.83	93.20	86.13	73.03	90.20	86.87	75.45	83.16	+3.26
TIP-Adapter	73.08	95.63	45.20	83.04	96.15	92.66	87.31	71.57	88.53	84.24	76.21	81.24	-
TIP-Adapter+R-AMT	74.28	96.97	61.07	86.27	97.80	94.07	87.43	74.77	91.50	86.93	76.97	84.37	+3.13

BibTeX

@inproceedings{RMT2023,
  title   = {Regularized Mask Tuning: Uncovering Hidden Knowledge in Pre-trained Vision-Language Models},
  author  = {Zheng, Kecheng and Wu, Wei and Feng$, Ruili and Zhu Kai and Liu, Jiawei and Zhao, Deli and Zha Zheng-Jun and Chen Wei and Shen, Yujun},
  booktitle = {ICCV},
  year    = {2023}
}

License

The project is under MIT License.