# VSA4VQA

![pipeline.png](pipeline.png)

Official code for [VSA4VQA: Scaling a Vector Symbolic Architecture to Visual Question Answering on Natural Images](https://perceptualui.org/publications/penzkofer24_cogsci/) published at CogSci'24

## Installation
```shell 
# create environment
conda create -n ssp_env python=3.9 pip
conda activate ssp_env
conda install pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia -y

sudo apt install libmysqlclient-dev

# install requirements
pip install -r requirements.txt

# install CLIP 
pip install git+https://github.com/openai/CLIP.git

# setup jupyter notebook kernel
python -m ipykernel install --user --name=ssp_env

```

## Get GQA Programs
using code by [https://github.com/wenhuchen/Meta-Module-Network](https://github.com/wenhuchen/Meta-Module-Network)<br>
- Download github repo MMN
- Add `gqa-questions` folder with GQA json files
- Run Preprocessing  
`python preprocess.py create_balanced_programs`
- Save generated programs to data folder: 
```
testdev_balanced_inputs.json
trainval_balanced_inputs.json
testdev_balanced_programs.json  
trainval_balanced_programs.json
```

> GQA dictionaries: `gqa_all_attributes.json` and `gqa_all_vocab_classes` are also adapted from [https://github.com/wenhuchen/Meta-Module-Network](https://github.com/wenhuchen/Meta-Module-Network)

## Generate Query Masks 
![query_masks.png](query_masks.png)
> All 37 generated query masks are available in [relations.zip](relations.zip) as numpy files 

```python
# loading a query mask with numpy
import numpy as np 
rel = 'to_the_right_of'
path = f'relations/{rel}.npy'
mask = np.load(path)
mask = mask > 0.05 	# binary mask
```
If you want to run the generation process, run: 
```shell 
python generate_query_masks.py 
``` 
- generates full_relations_df.pkl if not already present
- generates query masks for all relations with more than 1000 samples


## Pipeline 
Execute Pipeline for all samples in GQA: train_balanced (with `TEST=False`) or validation_balanced (with `TEST=True`)
```shell 
python run_programs.py 
``` 
For visualizing samples with all pipeline steps see [VSA4VQA_examples.ipynb](VSA4VQA_examples.ipynb) <br>

## Citation 
Please consider citing this paper if you use VSA4VQA or parts of this publication in your research:
```
@inproceedings{penzkofer24_cogsci,
  author = {Penzkofer, Anna and Shi, Lei and Bulling, Andreas},
  title = {VSA4VQA: Scaling A Vector Symbolic Architecture To Visual Question Answering on Natural Images},
  booktitle = {Proc. 46th Annual Meeting of the Cognitive Science Society (CogSci)},
  year = {2024},
  pages = {}
}
```