adversarial_AIRBERT/README.md

Finetuning Airbert on Downstream VLN Tasks

This repository stores the codebase for finetuning Airbert on downstream VLN tasks including R2R and REVERIE. The code is based on Recurrent-VLN-BERT. We acknowledge Yicong Hong for releasing the Recurrent-VLN-BERT code.

Prerequisites

1. Follow instructions in Recurrent-VLN-BERT to setup the environment and download data.

2. Download the trained models.

REVERIE

Inference

To replicate the performance reported in our paper, load the trained models and run validation:

bash scripts/valid_reverie_agent.sh 0

Training

To train the model, simply run:

bash scripts/train_reverie_agent.sh 0

R2R

Inference

To replicate the performance reported in our paper, load the trained models and run validation:

bash scripts/valid_r2r_agent.sh 0

Training

To train the model, simply run:

bash scripts/train_r2r_agent.sh 0

Citation

Please cite our paper if you find this repository useful:

@misc{guhur2021airbert,
  title ={{Airbert: In-domain Pretraining for Vision-and-Language Navigation}},
  author={Pierre-Louis Guhur and Makarand Tapaswi and Shizhe Chen and Ivan Laptev and Cordelia Schmid},
  year={2021},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
}