SSL-VQA

Here is the implementation of our IJCAI 2020 paper Overcoming Language Priors with Self-supervised Learning for Visual Question Answering. This repository contains code modified from here, many thanks!

Requirements

• python 3.6.8

• pytorch 1.0.1

• zarr

• tdqm

• spacy

• h5py

Download and preprocess the data

cd data 
bash download.sh
python preprocess_image.py --data trainval
python create_dictionary.py --dataroot vqacp2/
python preprocess_text.py --dataroot vqacp2/ --version v2
cd ..

Training

• Train our model with multi-label VQA loss

CUDA_VISIBLE_DEVICES=0 python main.py --dataroot data/vqacp2/ 
--img_root data/coco/ --output saved_models_cp2/ --self_loss_weight 3 --ml_loss

• Train our model with corss-entropy VQA loss

CUDA_VISIBLE_DEVICES=0 python main.py --dataroot data/vqacp2/ 
--img_root data/coco/ --output saved_models_cp2/ --self_loss_weight 1.2 --ce_loss

• Train the model with 80% of the original training set

CUDA_VISIBLE_DEVICES=0 python main.py --dataroot data/vqacp2/ 
--img_root data/coco/ --output saved_models_cp2/ --self_loss_weight 3 --ml_loss --ratio 0.8

Evaluation

• A json file of results from the test set can be produced with:

CUDA_VISIBLE_DEVICES=0 python test.py --dataroot data/vqacp2/ --img_root data/coco/ --checkpoint_path saved_models_cp2/best_model.pth --output saved_models_cp2/result/

• Compute detailed accuracy for each answer type:

python comput_score.py --input saved_models_cp2/result/XX.json --dataroot data/vqacp2/

Pretrained model & Well-trained model

If you don't want to train from scratch, you can download the pretrained base model from here(for ml_loss), and fine-tune it with our self-supervised loss as below:

CUDA_VISIBLE_DEVICES=0 python main.py --dataroot data/vqacp2/ 
--img_root data/coco/ --output saved_models_cp2/ --self_loss_weight 3 --ml_loss --checkpoint_path ml_pretrained.pth

A well-trained model (for ml_loss) can be found here. The test results file produced by it can be found here and its performance is as follows:

Overall score: 58.58
Yes/No: 87.47 Num: 40.3 other: 48.45

Reference

If you found this code is useful, please cite the following paper:

@inproceedings{ijcai2020-151,
  title     = {Overcoming Language Priors with Self-supervised Learning for Visual Question Answering},
  author    = {Zhu, Xi and Mao, Zhendong and Liu, Chunxiao and Zhang, Peng and Wang, Bin and Zhang, Yongdong},
  booktitle = {Proceedings of the Twenty-Ninth International Joint Conference on
               Artificial Intelligence, {IJCAI-20}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  editor    = {Christian Bessiere}	
  pages     = {1083--1089},
  year      = {2020},
  month     = {7},
  note      = {Main track}
  doi       = {10.24963/ijcai.2020/151},
  url       = {https://doi.org/10.24963/ijcai.2020/151},
}