MMBT is the accompanying code repository for the paper titled, "Supervised Multimodal Bitransformers for Classifying Images and Text" by Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Ethan Perez and Davide Testuggine.
The goal of the repository is to provide an implementation of the MMBT model and replicate the experiments in the paper.
Paper Link: https://arxiv.org/abs/1909.02950
- PyTorch version >= 1.0.0
- Python version >= 3.6
pip install torch torchvision sklearn pytorch-pretrained-bert numpy tqdm matplotlib
train.py provides the common training pipeline for all datasets.
- task: mmimdb, food101, vsnli
- model: bow, img, concatbow, bert, concatbert, mmbt
The following paths need to be set to start training.
- data_path: Assumes a subfolder for each dataset.
- savedir: Location to save model checkpoints.
- glove_path: Path to glove embeds file. Needed for bow, concatbow models.
Example command:
python mmbt/train.py --batch_sz 4 --gradient_accumulation_steps 40 \
--savedir /path/to/savedir/ --name mmbt_model_run \
--data_path /path/to/datasets/ \
--task food101 --task_type classification \
--model mmbt --num_image_embeds 3 --freeze_txt 5 --freeze_img 3 \
--patience 5 --dropout 0.1 --lr 5e-05 --warmup 0.1 --max_epochs 100 --seed 1
MMBT is also available in HuggingFace Transformers. See https://github.com/huggingface/transformers/tree/master/examples/research_projects/mm-imdb for an example that shows how easy it is to run MMBT in that framework.
MMBT is licensed under Creative Commons-Non Commercial 4.0. See the LICENSE file for details.
Please cite it as follows
@article{kiela2019supervised,
title={Supervised Multimodal Bitransformers for Classifying Images and Text},
author={Kiela, Douwe and Bhooshan, Suvrat and Firooz, Hamed and Testuggine, Davide},
journal={arXiv preprint arXiv:1909.02950},
year={2019}
}