Home

Awesome

Self-regulating Prompts: Foundational Model Adaptation without Forgetting [ICCV 2023]

Self-regulating Prompts: Foundational Model Adaptation without Forgetting<br> Muhammad Uzair Khattak*, Syed Talal Wasim*, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, Fahad Shahbaz Khan

*Joint first authors

paper Website video slides

Official implementation of the paper "Self-regulating Prompts: Foundational Model Adaptation without Forgetting".

<hr />

PWC PWC PWC PWC PWC

<hr />

:rocket: News

<hr />

Highlights

main figure

<p align="justify"> <b> <span style="color: blue;">Left</span></b>: Existing prompt learning approaches for foundational Vision-Language models like CLIP rely on task-specific objectives that restrict prompt learning to learn a feature space suitable only for downstream tasks and consequently lose the generalized knowledge of CLIP (shown in <span style="color: purple;">purple</span></b>). Our self-regulating framework explicitly guides the training trajectory of prompts towards the closest point between two optimal solution manifolds (solid line) to learn task-specific representations while also retaining generalized CLIP knowledge (shown in <span style="color: green;">green</span>). <b><span style="color: blue;">Middle</span></b>: Averaged across 11 image recognition datasets, PromptSRC surpasses existing methods on the base-to-novel generalization setting. <b><span style="color: blue;">Right</span></b>: We evaluate our approach on four diverse image recognition benchmarks for CLIP and show consistent gains over previous state-of-the-art approaches. </p>

<p align="justify"> Abstract: Prompt learning has emerged as an efficient alternative for fine-tuning foundational models, such as CLIP, for various downstream tasks. Conventionally trained using the task-specific objective, i.e., cross-entropy loss, prompts tend to overfit downstream data distributions and find it challenging to capture task-agnostic general features from the frozen CLIP. This leads to the loss of the model's original generalization capability. To address this issue, our work introduces a self-regularization framework for prompting called PromptSRC (Prompting with Self-regulating Constraints). PromptSRC guides the prompts to optimize for both task-specific and task-agnostic general representations using a three-pronged approach by: (a) regulating {prompted} representations via mutual agreement maximization with the frozen model, (b) regulating with self-ensemble of prompts over the training trajectory to encode their complementary strengths, and (c) regulating with textual diversity to mitigate sample diversity imbalance with the visual branch. To the best of our knowledge, this is the first regularization framework for prompt learning that avoids overfitting by jointly attending to pre-trained model features, the training trajectory during prompting, and the textual diversity. PromptSRC explicitly steers the prompts to learn a representation space that maximizes performance on downstream tasks without compromising CLIP generalization. We perform experiments on 4 benchmarks where PromptSRC performs favorably well compared to the existing methods. Our code and pre-trained models are publicly available. </p>

Regularization Framework for Prompt Learning

We propose PromptSRC (Prompting with Self-regulating Constraints) which steers the prompts to learn a representation space that maximizes performance on downstream tasks without compromising CLIP generalization.

Key components of PromptSRC:

  1. Mutual agreement maximization: PromptSRC explicitly guides the prompts to jointly acquire both <i>task-specific knowledge</i> and <i>task-agnostic generalized knowledge</i> by maximizing the mutual agreement between prompted and features of the frozen VL model.
  2. Gaussian weighted prompt aggregation: We propose a weighted self-ensembling strategy for prompts over the training trajectory that captures complementary features and enhances their generalization abilities.
  3. Textual diversity: PromptSRC regulates prompts with textual diversity to mitigate sample diversity imbalance compared to the visual branch during training.

:ballot_box_with_check: Supported Methods

MethodPaperConfigsTraining Scripts
PromptSRCarXivlinklink
Independent V-L Prompting-linklink
MaPLeCVPR 2023linklink
CoOpIJCV 2022linklink
Co-CoOpCVPR 2022linklink
<hr />

Results

Results reported below show accuracy for base and novel classes for across 11 recognition datasets averaged over 3 seeds.

Effectiveness of PromptSRC in comparison with baseline Independent V-L Prompting

PromptSRC effectively maximizes supervised task performance (base classes) without compromising on CLIP's original generalization towards new unseen tasks (novel classes).

NameBase Acc.Novel Acc.HMEpochs
CLIP69.3474.2271.70-
Independent V-L Prompting84.2171.7977.5120
PromptSRC (ours)84.2676.1079.9720

PromptSRC in comparison with existing state-of-the-art

NameBase Acc.Novel Acc.HMEpochs
CLIP69.3474.2271.70-
CoOp82.6963.2271.66200
CoCoOp80.4771.6975.8310
ProDA81.5675.8376.65100
MaPLe82.2875.1478.555
PromptSRC (ours)84.2676.1079.9720

Installation

For installation and other package requirements, please follow the instructions detailed in INSTALL.md.

Data Preparation

Please follow the instructions at DATASETS.md to prepare all datasets.

Model Zoo

Vision-Language prompting methods

Name (configs)Model checkpoints
Independent V-L Promptinglink
PromptSRClink

Evaluation

Please refer to the EVAL.md for detailed instructions on using the evaluation scripts and reproducing the official results using our pre-trained models.

Training

Please refer to the TRAIN.md for detailed instructions on training PromptSRC and IVLP baseline from scratch.

<hr />

Citation

If you find our work, this repository, or pretrained models useful, please consider giving a star :star: and citation.

@InProceedings{Khattak_2023_ICCV,
    author    = {Khattak, Muhammad Uzair and Wasim, Syed Talal and Naseer, Muzammal and Khan, Salman and Yang, Ming-Hsuan and Khan, Fahad Shahbaz},
    title     = {Self-regulating Prompts: Foundational Model Adaptation without Forgetting},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {15190-15200}
}

Contact

If you have any questions, please create an issue on this repository or contact at uzair.khattak@mbzuai.ac.ae or syed.wasim@mbzuai.ac.ae.

Acknowledgements

Our code is based on MaPLe, along with Co-CoOp and CoOp repository. We thank the authors for releasing their code. If you use our model and code, please consider citing these works as well.