Improving Vision-and-Language Navigation with Image-Text Pairs from the Web

Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra

Paper: https://arxiv.org/abs/2004.14973

Model Zoo

A variety of pre-trained VLN-BERT weights can accessed through the following links:

	Pre-training Stages	Job ID	Val Unseen SR	URL
0	no pre-training	174631	30.52%	TBD
1	1	175134	45.17%	TBD
3	1 and 2	221943	49.64%	download
2	1 and 3	220929	50.02%	download
4	1, 2, and 3 (Full Model)	220825	59.26%	download

Usage Instructions

Follow the instructions in INSTALL.md to setup this codebase. The instructions walk you through several steps including preprocessing the Matterport3D panoramas by extracting regions with a pretrained object detector.

Training

To preform stage 3 of pre-training, first download ViLBERT weights from here. Then, run:

python \
-m torch.distributed.launch \
--nproc_per_node=8 \
--nnodes=1 \
--node_rank=0 \
train.py \
--from_pretrained <path/to/vilbert_pytorch_model_9.bin> \
--save_name [pre_train_run_id] \
--num_epochs 50 \
--warmup_proportion 0.08 \
--cooldown_factor 8 \
--masked_language \
--masked_vision \
--no_ranking

To fine-tune VLN-BERT for the path selection task, run:

python \
-m torch.distributed.launch \
--nproc_per_node=8 \
--nnodes=1 \
--node_rank=0 \
train.py \
--from_pretrained <path/to/pytorch_model_50.bin> \
--save_name [fine_tune_run_id]

Evaluation

To evaluate a pre-trained model, run:

python test.py \
--split [val_seen|val_unseen] \
--from_pretrained <path/to/run_[run_id]_pytorch_model.bin> \
--save_name [run_id]

followed by:

python scripts/calculate-metrics.py <path/to/results_[val_seen|val_unseen].json>

Citation

If you find this code useful, please consider citing:

@inproceedings{majumdar2020improving,
  title={Improving Vision-and-Language Navigation with Image-Text Pairs from the Web},
  author={Arjun Majumdar and Ayush Shrivastava and Stefan Lee and Peter Anderson and Devi Parikh and Dhruv Batra},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2020}
}