Cascade Prompt Learning for Vision-Language Model Adaptation
Ge Wu*, Xin Zhang*, Zheng Li, Zhaowei Chen, Jiajun Liang, Jian Yang Xiang Li *Joint first authors
(September 2024) Training codes of CasPL PromptSRC will be released.
(July 2024) Our work is accepted to ECCV 2024!
- We propose a novel cascade prompt learning framework consisting of both boosting and adapting prompt phases. To our best knowledge, CasPL is the first to introduce cascaded diverse prompts with multiple phases for VLMs, which is a brand new learning paradigm for fine-tuning VLMs.
- We demonstrate that the boosting prompts can distill domain-general knowledge from the senior teacher over massive unlabeled domain images, leading to superior recognition performance and efficient inference.
- As a plug-and-play framework, CasPL can be seamlessly integrated into any existing prompt learning approaches, with negligible parameters (boosting prompt tokens, <0.1%) and ignorable additional inference cost introduced.
Base-to-Novel: Results reported below show accuracy for base and novel classes for across 11 recognition datasets averaged over 3 seeds.
| Name | Base Acc. | Novel Acc. | HM |
|---|---|---|---|
| CLIP | 69.34 | 74.22 | 71.70 |
| CoOp | 82.69 | 63.22 | 71.66 |
| CoOp CasPL | 84.78 | 74.49 | 79.30 ( 7.64) |
| CoCoOp | 80.47 | 71.69 | 75.83 |
| CoCoOp CasPL | 83.63 | 78.12 | 80.78 ( 4.95) |
| MaPLe | 82.28 | 75.14 | 78.55 |
| MaPLe CasPL | 84.48 | 79.59 | 81.96 ( 3.41) |
| PromptSRC | 84.26 | 76.10 | 79.97 |
| PromptSRC CasPL | 86.11 | 79.54 | 82.69 ( 2.72) |
Few-shot: Results reported below show accuracy across 11 recognition datasets averaged over 3 seeds.
| Name | 1 shot | 2 shot | 4 shot | 8 shot | 16 shot |
|---|---|---|---|---|---|
| Linear probe CLIP | 45.83 | 57.98 | 68.01 | 74.47 | 78.79 |
| CoOp | 67.56 | 70.65 | 74.02 | 76.98 | 79.89 |
| CoCoOp | 66.79 | 67.65 | 71.21 | 72.96 | 74.90 |
| MaPLe | 69.27 | 72.58 | 75.37 | 78.89 | 81.79 |
| PromptSRC | 72.32 | 75.29 | 78.35 | 80.69 | 82.87 |
| PromptSRC CasPL | 75.91 | 77.94 | 80.45 | 82.22 | 84.49 |
For installation and other package requirements, please follow the instructions detailed in INSTALL.md.
Please follow the instructions at DATASETS.md to prepare all datasets.
python train_script/caspl_promptsrc/train_base_to_new.py
train_base_to_new.py contains two-stage training, and the other experimental files also contain two-stage training:
bash scripts/ivlp/base2new_train_ivlp_teacher.sh #the first stage to get the boosting prompt
bash scripts/caspl/promptsrc/base2new_train_promptsrc_student.sh #the second stage to train the adapting prompt
python train_script/caspl_promptsrc/train_few_shot.py
python train_script/caspl_promptsrc/train_cross_dataset.py
If you find our work, this repository, or pretrained models useful, please consider giving a star and citation.
If you have any questions, please create an issue on this repository or contact at [email protected].
Our code is based on PromptSRC, 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.