This repository is the implementation of MEVF for the visual question answering task in medical domain. Our model achieved 43.9 for open-ended and 75.1 for close-end on VQA-RAD dataset. For the detail, please refer to link.
This repository is based on and inspired by @Jin-Hwa Kim's work. We sincerely thank for their sharing of the codes.
Please install dependence package by run following command:
pip install -r requirements.txt
All data should be downloaded via link. The downloaded file should be extracted to data_RAD/ directory.
Train MEVF model with Stacked Attention Network
$ python3 main.py --model SAN --use_RAD --RAD_dir data_RAD --maml --autoencoder --output saved_models/SAN_MEVF
Train MEVF model with Bilinear Attention Network
$ python3 main.py --model BAN --use_RAD --RAD_dir data_RAD --maml --autoencoder --output saved_models/BAN_MEVF
The training scores will be printed every epoch.
| SAN+proposal | BAN+proposal | |
|---|---|---|
| Open-ended | 40.7 | 43.9 |
| Close-ended | 74.1 | 75.1 |
In this repo, we include the pre-trained weight of MAML and CDAE which are used for initializing the feature extraction modules.
The MAML model data_RAD/pretrained_maml.weights is trained by using official source code link.
The CDAE model data_RAD/pretrained_ae.pth is trained by code provided in train_cdae.py. For reproducing the pretrained model, please check the instruction provided in that file.
We also provide the pretrained models reported as the best single model in the paper.
For SAN_MEVF pretrained model. Please download the link and move to saved_models/SAN_MEVF/. The trained SAN_MEVF model can be tested in VQA-RAD test set via:
$ python3 test.py --model SAN --use_RAD --RAD_dir data_RAD --maml --autoencoder --input saved_models/SAN_MEVF --epoch 19 --output results/SAN_MEVF
For BAN_MEVF pretrained model. Please download the link and move to saved_models/BAN_MEVF/. The trained BAN_MEVF model can be tested in VQA-RAD test set via:
$ python3 test.py --model BAN --use_RAD --RAD_dir data_RAD --maml --autoencoder --input saved_models/BAN_MEVF --epoch 19 --output results/BAN_MEVF
The result json file can be found in the directory results/.
Please cite these papers in your publications if it helps your research
@inproceedings{aioz_mevf_miccai19,
author={Binh D. Nguyen, Thanh-Toan Do, Binh X. Nguyen, Tuong Do, Erman Tjiputra, Quang D. Tran},
title={Overcoming Data Limitation in Medical Visual Question Answering},
booktitle = {MICCAI},
year={2019}
}
MIT License
AIOZ AI Homepage: https://ai.aioz.io
https://drive.google.com/file/d/1dqJjthrbdnIs41ZdC_ZGVQnoZbuGMNCR/view?usp=sharing