For doing simple vs complex classification we used simpcomp.py file. You can load your dataset accordinlgy. TallyQA already
has a field issimple that says whether the question is simple or complex so you don't need to run that file. For example,
{'answer': 4,
'data_source': 'imported_genome',
'image': 'VG_100K_2/2410408.jpg',
'image_id': 92410408,
'issimple': False, ## this entry here
'question': 'How many headlights does the black bus have?',
'question_id': 30095774}
The main paper uses the RN_BG_OG_embd model with the reported hyper-parameters in this repo.
There are other types of Relational Models avaialabe in the repo which we tried for our project.
To Run the code as:
python main.py --model RN_OG_embd
Don't forget to edit config.py to point to appropriate locations for files and features.
Please, feel free to ask if you have any other questions.