This repository contains the re-implementation along with improved results of the paper: Reinforced Video Captioning with Entailment Rewards (EMNLP 2017)

This code is tested on python 3.6 and pytorch 0.3.

Install all the required packages from requirements.txt file.

Download the ResNet-152 frame-level features + ResNeXt-101 motion features for MSR-VTT videos.

Download the captions and vocabulary data for MSR-VTT, and place the downloaded data in 'data' folder.

Clone the code from here to setup the evaluation metrics, and place it the parent directory on this repository. Note that this is also required during training since tensorboard logs the validation scores.

To train Baseline-XE model:

python main.py --model_name "model_name"

To train CIDEr-RL model:

python main.py --model_name "model_name" --load_path "path_to_baseline_model_folder" --lr 0.00001 --reward_type CIDEr --max_epoch 40 --loss_function xe+rl

To train CIDEnt-RL model:

python main.py --model_name "model_name" --load_path "path_to_baseline_model_folder" --load_entailment_path "path_to_entailment_model_end_with_*pth" --lr 0.00001 --reward_type CIDEnt --max_epoch 40 --loss_function xe+rl

For testing:

python main.py --mode test --load_path "path_to_model_folder" --beam_size 5 

Download the pretrained models for Baseline, CIDEr-RL, and CIDEnt-RL from here.

For running the pretrained models:

python main.py --mode test --load_path "path_to_model_ending_with_*.pth" --beam_size 5 

On running the above given pretrained models you should achieve the following results:

Note that first our CIDEr-RL model achieves stat. signif. improvements over Baseline-XE (on CIDEr, BLEU, and ROUGE) and then our CIDEnt-RL model achieves stat. signif. improvements over CIDEr-RL (on CIDEr, BLEU, and METEOR).

Note that this released baseline model has several improvements (and significantly better results) w.r.t. our EMNLP17 paper results, due to the following enhancements:

• Better visual features: (1) ResNet-152 frame-level features (2) ResNeXt-101 motion features
• Used SCST approach instead of MIXER for reinforcement leanring
• Used better entailment classifier

If you find this code helpful, please consider citing the following papers:

@inproceedings{pasunuru2017reinforced,
    title={Reinforced Video Captioning with Entailment Rewards},
    author={Pasunuru, Ramakanth and Bansal, Mohit},
    booktitle={EMNLP},
    year={2017}
}

@inproceedings{pasunuru2017multi,
    title={Multi-Task Video Captioning with Video and Entailment Generation},
    author={Pasunuru, Ramakanth and Bansal, Mohit},
    booktitle={ACL},
    year={2017}
}