This is a Pytorch implementation for Mucko: Multi-Layer Cross-Modal Knowledge Reasoning for Fact-based Visual Question Answering (IJCAI 2020).
NOTE: The offical publication has not been published!
- Install Python 3.7.
- Install PyTorch 1.2.
- Install other dependency packages.
- Clone this repository and enter the root directory of it.
git clone https://github.com/astro-zihao/mucko.git
For training the model
CUDA_VISIBLE_DEVICES=0 python train.py --config-yml exp_fvqa/exp.yml --cpu-workers 8 --gpus 0 --save-dirpath fvqa/exp_data/checkpoints
- config-yml: Path to a config file listing reader, model and solver parameters.
- cpu-workers: Number of CPU workers for dataloader.
- save-dirpath: Path of directory to create checkpoint directory and save checkpoints.
- load-pthpath: To continue training, path to .pth file of saved checkpoint.
- validate: Whether to validate on val split after every epoch.
@inproceedings{zhu2020mucko,
title={Mucko: Multi-Layer Cross-Modal Knowledge Reasoning for Fact-based Visual Question Answering,
author={Zhu, Zihao and Yu, Jing and Sun, Yajing and Hu, Yue and Wang, Yujing and Wu, Qi},
booktitle={International Joint Conference on Artificial Intelligence (IJCAI)},
year={2020}
}
Part of this code uses components from DualVD. We thank authors for releasing their code.