cross-modal-ablation icon indicating copy to clipboard operation
cross-modal-ablation copied to clipboard

[EMNLP 2021] Code and data for our paper "Vision-and-Language or Vision-for-Language? On Cross-Modal Influence in Multimodal Transformers"

Cross-Modal Ablation

This is the implementation of the approaches described in the paper:

Stella Frank*, Emanuele Bugliarello* and Desmond Elliott. Vision-and-Language or Vision-for-Language? On Cross-Modal Influence in Multimodal Transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), Nov 2021.

We provide the code for reproducing our results.

The cross-modal ablation task has also been integrated into VOLTA, upon which our repository was origally built.

Repository Setup

You can clone this repository issuing:
git clone [email protected]:e-bug/cross-modal-ablation

1. Create a fresh conda environment, and install all dependencies.

conda create -n xm-influence python=3.6
conda activate xm-influence
pip install -r requirements.txt

2. Install apex. If you use a cluster, you may want to first run commands like the following:

module load cuda/10.1.105
module load gcc/8.3.0-cuda

3. Setup the refer submodule for Referring Expression Comprehension:

cd tools/refer; make

Data

For textual data, please clone the Flickr30K Entities repository:
[email protected]:BryanPlummer/flickr30k_entities.git

For visual features, we use the VOLTA release for Flickr30K.

Our datasets directory looks as follows:

data/
 |-- flickr30k/
 |    |-- resnet101_faster_rcnn_genome_imgfeats/
 |
 |-- flickr30k_entities/
 |    |-- Annotations/
 |    |-- Sentences/
 |    |-- val.txt

Once you have defined the path to your datasets directory, make sure to update the cross-modal influence configuration file (e.g. volta/config_tasks/xm-influence_test_tasks.yaml).

Our Dataset class for cross-modal ablation on Flickr30K Entites is implemented in volta/volta/datasets/flickr30ke_ablation_dataset.py.

The LabelMatch subset can be derived following the notebook notebooks/Data-MakeLabelMatch.ipynb.

Models

Most of the models we evaluated were released in VOLTA (Bugliarello et al., 2021).

If you are interested in using some of the variations we studied in our paper, reach out to us or open an issue on GitHub.

Training and Evaluation

We introduce the following scripts in this repository:

  • volta/train_bert_concap.py: Pretrain a text-only model on the textual modality of Conceptual Captions. We use this script to train BERT-CC in our paper.
  • volta/train_concap_vis.py: Pretrain only on the visual modality of Conceptual Captions.
  • volta/ablate_textonly_lang.py: Evaluate the performance of text-only models in predicting the masked phrases.
  • volta/ablate_vis4lang.py: Evaluate the performance of V&L models in predicting masked phrases as visual inputs are ablated.
  • volta/ablate_lang4vis.py: Evaluate the performance of V&L models in predicting masked objects as textual inputs are ablated.

We provide all the scripts we used in our study under experiments/.

We share our results aggregated in TSV files under notebooks/.

License

This work is licensed under the MIT license. See LICENSE for details. Third-party software and data sets are subject to their respective licenses.
If you find our code/data/models or ideas useful in your research, please consider citing the paper:

@inproceedings{frank-etal-2021-vision,
    title = "Vision-and-Language or Vision-for-Language? {O}n Cross-Modal Influence in Multimodal Transformers",
    author = "Frank, Stella and Bugliarello, Emanuele and
      Elliott, Desmond",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021)",
    month = "nov",
    year = "2021",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/2109.04448",
}