Relation-aware Graph Attention Network for Visual Question Answering

This repository is the implementation of Relation-aware Graph Attention Network for Visual Question Answering.

This repository is based on and inspired by @hengyuan-hu's work and @Jin-Hwa Kim's work. We sincerely thank for their sharing of the codes.

Prerequisites

You may need a machine with 4 GPUs with 16GB memory each, and PyTorch v1.0.1 for Python 3.

1. Install PyTorch with CUDA10.0 and Python 3.7.
2. Install h5py.
3. Install block.bootstrap.pytorch.

If you are using miniconda, you can install all the prerequisites with tools/environment.yml.

Data

Our implementation uses the pretrained features from bottom-up-attention, the adaptive 10-100 features per image. In addition to this, the GloVe vectors and Visual Genome question answer pairs. For your convenience, the below script helps you to download preprocessed data.

source tools/download.sh

In addition to data, this script also download several pretrained models. In the end, the data folder and pretrained_models folder should be organized as shown below:

├── data
│   ├── Answers
│   │   ├── v2_mscoco_train2014_annotations.json
│   │   └── v2_mscoco_val2014_annotations.json
│   ├── Bottom-up-features-adaptive
│   │   ├── train.hdf5
│   │   ├── val.hdf5
│   │   └── test2015.hdf5
│   ├── Bottom-up-features-fixed
│   │   ├── train36.hdf5
│   │   ├── val36.hdf5
│   │   └── test2015_36.hdf5
│   ├── cache
│   │   ├── cp_v2_test_target.pkl
│   │   ├── cp_v2_train_target.pkl
│   │   ├── train_target.pkl
│   │   ├── val_target.pkl
│   │   ├── trainval_ans2label.pkl
│   │   └── trainval_label2ans.pkl
│   ├── cp_v2_annotations
│   │   ├── vqacp_v2_test_annotations.json
│   │   └── vqacp_v2_train_annotations.json
│   ├── cp_v2_questions
│   │   ├── vqacp_v2_test_questions.json
│   │   └── vqacp_v2_train_questions.json
│   ├── glove
│   │   ├── dictionary.pkl
│   │   ├── glove6b_init_300d.npy
│   │   └──- glove6b.300d.txt
│   ├── imgids
│   │   ├── test2015_36_imgid2idx.pkl
│   │   ├── test2015_ids.pkl
│   │   ├── test2015_imgid2idx.pkl
│   │   ├── train36_imgid2idx.pkl
│   │   ├── train_ids.pkl
│   │   ├── train_imgid2idx.pkl
│   │   ├── val36_imgid2idx.pkl
│   │   ├── val_ids.pkl
│   │   └── val_imgid2idx.pkl
│   ├── Questions
│   │   ├── v2_OpenEnded_mscoco_test-dev2015_questions.json
│   │   ├── v2_OpenEnded_mscoco_test2015_questions.json
│   │   ├── v2_OpenEnded_mscoco_train2014_questions.json
│   │   └── v2_OpenEnded_mscoco_val2014_questions.json
│   ├── visualGenome
│   │   ├── image_data.json
│   │   └── question_answers.json

├── pretrained_models (each model folder contains hps.json and model.pth)
│   ├── regat_implicit
│   │   ├── ban_1_implicit_vqa_196
│   │   ├── ban_4_implicit_vqa_cp_4422
│   │   ├── butd_implicit_vqa_6371
│   │   └── mutan_implicit_vqa_2632
│   ├── regat_semantic
│   │   ├── ban_1_semantic_vqa_7971
│   │   ├── ban_4_semantic_vqa_cp_9960
│   │   ├── butd_semantic_vqa_244
│   │   └── mutan_semantic_vqa_2711
│   ├── regat_spatial
│   │   ├── ban_1_spatial_vqa_1687
│   │   ├── ban_4_spatial_vqa_cp_4488
│   │   ├── butd_spatial_vqa_5942
│   │   └── mutan_spatial_vqa_3842

Training

python3 main.py --config config/butd_vqa.json

Evaluating

# take ban_1_implicit_vqa_196 as an example
# to evaluate cp_v2 performance, need to use --dataset cp_v2 --split test
python3 eval.py --output_folder pretrained_models/regat_implicit/ban_1_implicit_vqa_196

Citation

If you use this code as part of any published research, we'd really appreciate it if you could cite the following paper:

@article{li2019relation,
  title={Relation-aware Graph Attention Network for Visual Question Answering},
  author={Li, Linjie and Gan, Zhe and Cheng, Yu and Liu, Jingjing},
  journal={ICCV},
  year={2019}
}

License

MIT License