Skip to content

Latest commit

 

History

History
156 lines (118 loc) · 6.82 KB

README.md

File metadata and controls

156 lines (118 loc) · 6.82 KB
Raw

Subspace-Prompt-Learning

Official code for "Understanding and Mitigating Overfitting in Prompt Tuning for Vision-Language Models".

Chengcheng Ma1,2, Yang Liu3, Jiankang Deng4, Lingxi Xie4, Weiming Dong1, Changsheng Xu1.

1 Chinese Academy of Sciences Institute of Automation (CASIA)
2 University of the Chinese Academy of Sciences (UCAS)
3 Alibaba DAMO Academy
4 Huawei Technologies Co Ltd

TL;DR

We propose Subspace Prompt Tuning (SubPT) to mitigate the overfitting issue in the well-known prompt tuning method CoOp, and further propose Novel Feature Learner (NFL) to enhance the generalization ability onto novel categories beyond the training set.

SubPT is illustrated as: SubPT

The full picture of our method: Overview

Preparation

This repo is based on the toolbox Dassl.pytorch, and we add the model_subspace_backward_and_update function into Dassl.pytorch/dassl/engine/trainer.py to support subspace prompt tuning.

Before you go, please go to the ./Dassl.pytorch directory and make installation as follows.

# Create a conda environment
conda create -n subpt python=3.7

# Activate the environment
conda activate subpt

# Install dependencies
pip install -r requirements.txt

# Install torch (version >= 1.7.1) and torchvision
# Please make sure you have installed the gpu version due to the speed.
# For example:
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

# Install this library
cd Dassl.pytorch
python setup.py develop

Then go to the ./CoOp directory and run pip install -r requirements.txt to install a few more packages required by CLIP.

Follow DATASETS.md to install the datasets.

Usage

Few-shot classification on 11 datasets

Please go to the ./CoOp directory, and run "CoOp+SubPT" as follows.

############### Step 1. run CoOp ###############
# [SHOTS] and [EPOCH] are pairwise hyper-parameters in CoOp, specified as 
# (1 shot, 50 epoch)
# (2 shots, 100 epoch)
# (4 shots, 100 epoch) 
# (8 shots, 200 epoch) 
# (16 shots, 200 epoch)
# [EPOCH] is specially set as 50 for all shots for ImageNet.
cd scripts
bash coop.sh [SHOTS] [EPOCH] [DATASET]


############### Step 2. compute dominate eigenvectors representing the early-stage gradient flow ###############
# [FINISH] and [DIM] are the only two hyper-paramters in SubPT, corresponding to the $t_early$ and $r$ in our paper.
# [FINISH] and [DIM] are optionally specified as follows. [DIM] < [FINISH]. Other [DIM] values lead to similar results.
# (10, 5)  for 1 shot
# (20, 10) for 2 shot
# (30, 10) for 4 shot
# (40, 10) for 8 shot 
# (50, 10) for 16 shot
cd ..
python compute_eigenvector.py --ckpt_path [CKPT_PATH] --start 1 --finish [FINISH] --save_name \
  full_P/[DATASET]-CoOp-[SHOTS]shots-nctx16-seed1-start1-finish[FINISH]-dim[DIM].pth --n_components [DIM]


############### Step 3. re-run CoOp with SubPT ###############
# Note that [SHOTS] and [EPOCH] are in correspondence with Step 1, and [FINISH] and [DIM] are in correspondence with Step 2.
cd scripts
bash coop_sub.sh [SHOTS] [EPOCH] [FINISH] [DIM] [DATASET]

To run "CoOp+NFL" and "CoOp+SubPT+NFL", just replace coop.sh with coop_nfl.sh, and replace coop_sub.sh with coop_sub_nfl.sh. Before Step 1, please remember to pre-compute the text features with zero-shot CLIP and save them in the ./CoOp/text_features_nfl directory. (Hint: run zeroshot2.sh and add torch.save at here).

We kindly write a ./output/quick_view_all_acc.py script for you, in order to measure the classification accuracy.

Base-to-Novel Generalization on 11 datasets

Please go to the ./CoOp directory, and run "CoOp+SubPT" as follows.

############### Step 1. run CoOp ###############
# [SHOTS] and [EPOCH] are fixed as (4 shots, 100 epoch).
cd scripts
bash base2new_train_coop.sh [DATASET]


############### Step 2. compute dominate eigenvectors representing the early-stage gradient flow ###############
# [FINISH] and [DIM] are fixed as 30 and 10, respectively.
cd ..
python compute_eigenvector.py --ckpt_path [CKPT_PATH] --start 1 --finish 30 --save_name \
  full_P/b2n-[DATASET]-CoOp-4shots-nctx16-seed1-start1-finish30-dim10.pth --n_components 10


############### Step 3. re-run CoOp with SubPT ###############
# Note that [FINISH] and [DIM] are in correspondence with Step 2.
cd scripts
bash base2new_train_coop_sub.sh [DATASET]

After training, do evaluation as follows

cd scripts
# [SUB] is "base" or "new". [LOADEP] is 100, except 50 for ImageNet.
bash base2new_test_coop_sub.sh [SUB] [LOADEP] [DATASET]

Domain Generalization on 4 datasets

Please go to the ./CoOp directory, and run evaluation as follows.

cd scripts
# [TRAINER] can be CoOp, CoOp_sub, or CoOp_sub_nfl.
# [DATASET] can be imagenetv2, imagenet-sketch, imagenet-a, or imagenet-r.
bash eval.sh [TRAINER] [DATASET]

Zero-Shot CLIP Baseline

Please go to the ./CoOp/script directory and run bash zeroshot.sh [DATASET].

Citation

If you find this work useful, please consider citing our paper. We provide a BibTeX entry of our paper below:

@article{ma2022understanding,
  author={Ma, Chengcheng and Liu, Yang and Deng, Jiankang and Xie, Lingxi and Dong, Weiming and Xu, Changsheng},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={Understanding and Mitigating Overfitting in Prompt Tuning for Vision-Language Models}, 
  year={2023},
  volume={33},
  number={9},
  pages={4616-4629},
  doi={10.1109/TCSVT.2023.3245584}
}

Acknowledgements

  • This repo is based on CoOp and Dassl.pytorch.
  • Thanks to DLDR repo (see reference [33] in our paper). It inspires me a lot.

Contact

Feel free to contact me via machengcheng2016@gmail.com if you have any problems about our paper or codes.