CSL

This repository contains the reference code for the a novel video-text retrieval method "Complementarity-aware Space Learning", codes and trained models will be released soon!

Environment setup

• Python 3.6
• Pytorch 1.7.1 (strongly recommand)
• Numpy 1.19

Data preparation

1. Please follow dual_enc to download the required features for MSRVTT. Then put the folder “msrvtt10k” under ${DataPath }.

2. Please follow HGR to download the required annotations (int2word.json, word2int.json, word_embeds.glove32b.th). Then put them under the file ${DataPath }.

# clone the repository
git clone 
cd CSL
export PYTHONPATH=$(pwd):${PYTHONPATH}

Inference procedure

To reproduce the results of our paper, do the following two steps:

1. modify the /path/to/data in ./inference.py into $DataPath

2. please run the code below:

python inference.py

Training procedure

To train a our CSL model, do the following two steps:

1. modify the /path/to/data in ./train_msrvtt.py into $DataPath
2. please run the code below:

cd ./
python main.py

Note that it takes 1 v100 GPUs and around 20 hours to train this model.