/DSR-RL

Pytorch implementation of DSR-RL for Video Summarization Task

Primary LanguagePython

Self-Attention Recurrent Summarization Network with Reinforcement Learning for Video Summarization Task

IEEE International Conference on Multimedia and Expo (ICME) 2021

This is an implementation of DSR-RL (Deep Self-attention Recurrent summarization network with Reinforcement Learning).

Datasets Preparation

Download the pre-processed datasets into datasets/ folder, including TVSum, SumMe, OVP, and YouTube datasets.

mkdir -p datasets/ && cd datasets/
wget https://www.dropbox.com/s/8lvlf3knqwfeth9/datasets.zip
unzip datasets.zip

If the Dropbox link is unavailable to you, try downloading from below links.

Now the datasets structure should look like

DSR-RL
└── datasets/
    ├── eccv16_dataset_ovp_google_pool5.h5
    ├── eccv16_dataset_summe_google_pool5.h5
    ├── eccv16_dataset_tvsum_google_pool5.h5
    ├── eccv16_dataset_youtube_google_pool5.h5

Userscore Preparation

For user score files can be download from Summarizer project.

and put file summarizer_dataset_summe_google_pool5 and summarizer_dataset_tvsum_google_pool5.h5 in datasets/ folder

DSR-RL
└── datasets/
    ├── eccv16_dataset_ovp_google_pool5.h5
    ├── eccv16_dataset_summe_google_pool5.h5
    ├── eccv16_dataset_tvsum_google_pool5.h5
    ├── eccv16_dataset_youtube_google_pool5.h5
    ├── summarizer_dataset_summe_google_pool5.h5
    ├── summarizer_dataset_tvsum_google_pool5.h5

Training Model in Canonical Mode

SumMe

python main.py -d datasets/eccv16_dataset_summe_google_pool5.h5 -s splits/summe_splits.json -m summe --gpu 0  --save-dir log/summe-split --split-id 0 --verbose --rnn-cell gru --max-epoch 50 --userscore datasets/summarizer_dataset_summe_google_pool5.h5 --save-results

TVSum

python main.py -d datasets/eccv16_dataset_tvsum_google_pool5.h5 -s splits/tvsum_splits.json -m tvsum --gpu 0 --save-dir log/tvsum-split --split-id 0 --verbose --rnn-cell gru --max-epoch 50 --userscore datasets/summarizer_dataset_tvsum_google_pool5.h5 --save-results

Training Model in Augmented Mode

SumMe

python main.py -s datasets/splits/vasnet_summe_aug_splits.json -m summe --gpu 0 --save-dir log/summe-aug-split --split-id 0 --verbose --rnn-cell gru --max-epoch 50 --userscore datasets/summarizer_dataset_summe_google_pool5.h5

TVSum

python main.py -s datasets/splits/vasnet_tvsum_aug_splits.json -m tvsum --gpu 0 --save-dir log/tvsum-aug-split --split-id 0 --verbose --rnn-cell gru --max-epoch 50 --userscore datasets/summarizer_dataset_tvsum_google_pool5.h5

Training Model in Transfer Mode

SumMe

python main.py -s datasets/splits/vasnet_summe_tran_splits.json -m summe --gpu 0 --save-dir log/summe-tran-split --split-id 0 --verbose --rnn-cell gru --max-epoch 50 --userscore datasets/summarizer_dataset_summe_google_pool5.h5

TVSum

python main.py -s datasets/splits/vasnet_tvsum_tran_splits.json -m tvsum --gpu 0 --save-dir log/tvsum-tran-split --split-id 0 --verbose --rnn-cell gru --max-epoch 50 --userscore datasets/summarizer_dataset_tvsum_google_pool5.h5

Citation

If this code is useful for you, please consider citing the paper as following detail:

@inproceedings{phaphuangwittayakul2021self,
  title={Self-Attention Recurrent Summarization Network with Reinforcement Learning for Video Summarization Task},
  author={Phaphuangwittayakul, Aniwat and Guo, Yi and Ying, Fangli and Xu, Wentian and Zheng, Zheng},
  booktitle={2021 IEEE International Conference on Multimedia and Expo (ICME)},
  pages={1--6},
  year={2021},
  organization={IEEE}
}

Acknowledgement

This work is heavily based on VASNet and DR-DSN. Thanks them for their great work!

DSR-RL is freely available for non-commercial use. Don't hesitate to drop e-mail if you have any problem.