KDMCSE: Knowledge Distillation Multimodal Sentence Embeddings with Adaptive Angular margin Contrastive Learning

This repository contains code and pre-trained models for our NAACL-2024 paper KDMCSE: Knowledge Distillation Multimodal Sentence Embeddings with Adaptive Angular margin Contrastive Learning.

Quickstart

Setup

• Pytorch
• Other packages:

pip install -r requirements.txt

Data Preparation

Please organize the data directory as following:

REPO ROOT
|
|--data    
|  |--wiki1m_for_simcse.txt  
|  |--flickr30k_ViT_L14.json
|  |--train_coco_ViT_L14.json

Wiki1M

wget https://huggingface.co/datasets/princeton-nlp/datasets-for-simcse/resolve/main/wiki1m_for_simcse.txt

Flickr30k & MS-COCO
Download the preprocessed data:

• flickr30k_ViT_L14.json
• train_coco_ViT_L14.json

Or preprocess the data by yourself (using raw images from Flickr and COCO with CLIP).

Train & Evaluation

1. Prepare the senteval datasets for evaluation:

   cd SentEval/data/downstream/
   bash download_dataset.sh
   

2. Run scripts:

   # For example:
   sh scripts/run_wiki_flickr.sh
   
   sh scripts/run_wiki_coco.sh

Acknowledgements

• The extremely clear and well organized codebase: MCSE
• SentEval toolkit