SuS-X: Training-Free Name-Only Transfer of Vision-Language Models

Official code for the ICCV'23 paper "SuS-X: Training-Free Name-Only Transfer of Vision-Language Models". Authors: Vishaal Udandarao, Ankush Gupta and Samuel Albanie.

Introduction

Contrastive Language-Image Pre-training (CLIP) has emerged as a simple yet effective way to train large-scale vision-language models. CLIP demonstrates impressive zero-shot classification and retrieval on diverse downstream tasks. However, to leverage its full potential, fine-tuning still appears to be necessary. Fine-tuning the entire CLIP model can be resource-intensive and unstable. Moreover, recent methods that aim to circumvent this need for fine-tuning still require access to images from the target distribution. We pursue a different approach and explore the regime of training-free "name-only transfer" in which the only knowledge we possess about the downstream task comprises the names of downstream target categories. We propose a novel method, SuS-X, consisting of two key building blocks: "SuS" and "TIP-X", that requires neither intensive fine-tuning nor costly labelled data. SuS-X achieves state-of-the-art zero-shot classification results on 19 benchmark datasets. We further show the utility of TIP-X in the training-free few-shot setting, where we again achieve state-of-the-art results over strong training-free baselines.

Getting started

All our code was tested on Python 3.6.8 with Pytorch 1.9.0+cu111. Ideally, our scripts require access to a single GPU (uses .cuda() for inference). Inference can also be done on CPUs with minimal changes to the scripts.

Setting up environments

We recommend setting up a python virtual environment and installing all the requirements. Please follow these steps to set up the project folder correctly:

git clone https://github.com/vishaal27/SuS-X.git
cd SuS-X

python3 -m venv ./env
source env/bin/activate

pip install -r requirements.txt

Setting up datasets

We provide detailed instructions on how to set up our datasets in data/DATA.md.

Directory structure

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

SuS-X/
|–– data
|–––– ucf101
|–––– ... 18 other datasets
|–– features
|–– gpt3_prompts
|–––– CuPL_prompts_ucf101.json
|–––– ... 18 other dataset json files
|–– README.md
|–– clip.py
|–– ... all other provided python scripts

Running the baselines

Zero-shot CLIP

You can run Zero-shot CLIP inference using:

python run_zs_baseline.py --dataset <dataset> --backbone <CLIP_visual_backbone>

The backbone parameter can be one of [RN50, RN101, ViT-B/32, ViT-B/16].

CALIP

You can run our re-implementation of the CALIP baseline using:

python run_calip_baseline.py --dataset <dataset> --backbone <CLIP_visual_backbone>

CuPL

You can run the CuPL and CuPL+e baselines using:

python run_cupl_baseline.py --dataset <dataset> --backbone <CLIP_visual_backbone>

This script will also save the CuPL and CuPL+e text classifier weights into features/.

SuS Construction

We provide scripts for both SuS-SD generation and SuS-LC retrieval.

Photo prompting strategy

The prompts used for the Photo prompting strategy can be found in utils/prompts_helper.py.

CuPL prompting strategy

To generate customised CuPL prompts using GPT-3, we require access to an OpenAI token. Please create an account on OpenAI and find your key under the keys tab. Please ensure that the key is in the format sk-xxxxxxxxx. You can then run the following command to generate CuPL prompts for any dataset:

python generate_gpt3_prompts.py --dataset <dataset> --openai_key <openai_key>

For ensuring reproducibility, we provide all 19 dataset CuPL prompt files generated by us (and used for SuS generation and CuPL inference) in gpt3-prompts.

SuS-SD generation

For generating images using the Stable-Diffusion v1-4 checkpoint, we need a huggingface token. Please create an account on huggingface and find your token under the access tokens tab. Please ensure that the token is in the format hf_xxxxxxxxx. You can then generate the support set images using the command:

python generate_sd_sus.py --dataset <dataset> --num_images <number_of_images_per_class> --prompt_shorthand <prompting_strategy> --huggingface_key <huggingface_token>

<prompting_strategy> is photo for the Photo strategy and cupl for the CuPL strategy (refer Sec. 3.1 of the paper for more details).The generated support set is saved in data/sus-sd/<dataset>/<prompting_strategy>.

SuS-LC retrieval

There are two steps for correctly creating the SuS-LC support sets:

1. Downloading the URLs of the top-ranked images from LAION-5B. You can download the URLs for the images in the support set using:

python retrieve_urls_lc.py --dataset <dataset> --num_images <number_of_image_urls_per_class> --prompt_shorthand <prompting_strategy>

This will download all the URLs for the images to be downloaded in data/sus-lc/download_urls/<dataset>/<prompting_strategy>.

2. Downloading the top-ranked images using the downloaded URLs. You can download the support set images using:

python retrieve_images_lc_sus.py --dataset <dataset> --num_images <number_of_images_per_class> --prompt_shorthand <prompting_strategy>

The generated support set is saved in data/sus-lc/<dataset>/<prompting_strategy>.

Constructing the features

Test, validation and few-shot features

You can create the test and validation image features using:

python encode_datasets.py --dataset <dataset>

This script will save the test, validation and few-shot features in features/.

SuS features

You can create the curated SuS features using:

# for SuS-LC
python encode_sus_lc.py --dataset <dataset> --prompt_shorthand <prompting_strategy>
# for SuS-SD
python encode_sus_sd.py --dataset <dataset> --prompt_shorthand <prompting_strategy>

These scripts will save the respective SuS image features in features/.

Text classifier weights

You can create the different text classifier weights using:

python generate_text_classifier_weights.py --dataset <dataset>

This script will again save all the text classifier weights in features/.

For ensuring reproducibility, we release the features used for all our baselines and our best performing SuS-X-LC-P model here. We further provide detailed descriptions of the naming of the feature files in features/FEATURES.md.

TIP-X Inference

Once you have correctly saved all the feature files, you can run TIP-X using:

python tipx.py --dataset <dataset> --backbone <CLIP_visual_backbone> --prompt_shorthand <prompting_strategy> --sus_type <SuS_type>

The sus_type parameter is lc for SuS-LC and sd for SuS-SD.

Citation

If you found this work useful, please consider citing it as:

@inproceedings{udandarao2022sus-x,
  title={SuS-X: Training-Free Name-Only Transfer of Vision-Language Models},
  author={Udandarao, Vishaal and Gupta, Ankush and Albanie, Samuel},
  booktitle={ICCV},
  year={2023}
}

Acknowledgements

We build on several previous well-maintained repositories like CLIP, CoOp, CLIP-Adapter, TIP-Adapter and CuPL. We thank the authors for providing such amazing code, and enabling further research towards better vision-language model adaptation. We also thank the authors of the amazing Stable-Diffusion and LAION-5B projects, both of which are pivotal components of our method.

Contact

Please feel free to open an issue or email us at vu214@cam.ac.uk.