Implementation of Not All Features Matter:Enhancing Few-shot CLIP with Adaptive Prior Refinement, a few-shot framework designed for CLIP and other vision-language models.

Adaptive Prior rEfinement (APE) is a new method for few-shot CLIP, which achieves superior accuracy with high computational efficiency. Via a prior refinement module, we analyze the inter-class disparity in the downstream data and decouple the domain-specific knowledge from the CLIP-extracted cache model. On top of that, we introduce two model variants, a training-free APE and a training-required APE-T. We explore the trilateral affinities between the test image, prior cache model, and textual representations, and only enable a lightweight category-residual module to be trained. For the average accuracy over 11 benchmarks, both APE and APE-T attain SOTA.

Create a conda environment and install dependencies:

conda create -n APE python=3.7
conda activate APE

pip install -r requirements.txt

# Install the according versions of torch and torchvision
conda install pytorch torchvision cudatoolkit

Follow DATASET.md to install ImageNet and other 10 datasets referring to CoOp.

In this code, we seperate the feature extraction and model inference. The features should be extracted first, then conduct model inference.

We recommend users directly download the features from this google drive.

After setting up the datasets and downloading the features, the project root folder should look like this:

APE/
|–– caches
|–––– caltech101
|–––– dtd
|–––– eurosat
|–––– ... 8 other datasets' features
|–– clip
|–– configs
|–– data
|–––– caltech101
|–––– ... 10 other datasets
|–– datasets
|–––– caltech101.yaml
|–––– ... 10 other yamls
|–– gpt3_prompts
|–––– CuPL_prompts_caltech101.json
|–––– ... 10 other dataset json files
|–– extract_features.py
|–– main.py
|–– utils.py
|–– README.md

Or you can extract the features by youself by running

CUDA_VISIBLE_DEVICES=0 python extract_features.py

to extract the features of all 11 datasets. Including the few-shot training set representation, validation and test set representation, and textual representation.

By running

CUDA_VISIBLE_DEVICES=0 python main.py --config configs/[dataset_name].yaml --shot [shot_number]

users can test the proposed APE and APE-T, where you can change the dataset_name and shot_number. dataset_name should be one of [caltech101, dtd, eurosat, fgvc, food101, imagenet, oxford_flowers, oxford_pets, stanford_cars, sun397, ucf101], and shot_number is chosen from 1/2/4/8/16.

We build on several previous well-maintained repositories like TIP-Adapter, CLIP, and CoOp, SuS-X, and CuPL. We thank the authors for providing such amazing code, and enabling further research towards better vision-language model adaptation.

Not All Features Matter:Enhancing Few-shot CLIP with Adaptive Prior Refinement

@misc{zhu2023features,
      title={Not All Features Matter: Enhancing Few-shot CLIP with Adaptive Prior Refinement}, 
      author={Xiangyang Zhu and Renrui Zhang and Bowei He and Aojun Zhou and Dong Wang and Bin Zhao and Peng Gao},
      year={2023},
      eprint={2304.01195},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}