We used a small part of the dataset downloading and preprocessing code released by the official body that released the VQA dataset. link: https://github.com/GT-Vision-Lab/VQA_LSTM_CNN/blob/master/data/vqa_preprocessing.py This file is used as the vqa_preprocess.py file in our codebase
Every part of the codebase was developed by the team except for the already noted preprocessing files mentioned above.
Given the current configuration, we can train the models using the following steps
- baseline model: run
python main_baseline.py - single word model: run
python main_one_word.py - multi word model: run
python main_enc_dec.py
For demo of our models:
python demo.py
All files take command line arguments that have some default value set, the parameters can be tweaked as the user likes
Machine and software requirements
- GPU access (compulsory for demo)
- Pytorch v0.4
- torchtext (by pytorch)
- Python 3.7+
- SPaCy
- CUDA toolkit 9.0
- CuDnn