This repo is cretaed as a part of project course work for IFT6010 - 2021. In this project, we adapt Liu Bang’s Answer-Clue-Style-aware approach to the VQG task. Ourgoal is to develop a scalable system that can automatically generate diverse and high-quality visual question-answers pairs from a set of un-labelled images. The proposed pipeline is a combination of an image captioning module that extracts detailed and meaningful information from an image and an ACS-aware question generation module that produces question-answers pairs from the extracted visual infor-mation.
This project works on on paper " Asking Questions the Human Way: Scalable Question-Answer Generation from Text Corpus" (https://arxiv.org/abs/2002.00748) and the code is adapted from https://github.com/BangLiu/ACS-QG
- Down load the datasets to train - use setup.sh file to download all required files and use Create_Glove_bin.py file to create Glove bin files
- Download VQA datasets from https://visualqa.org/download.html
- Install requirements.txt
- Run_mergevqa.sh. This will merge Caption generated from Image model,Questions from VQA datset,image id from Annotations to give vqa_test.json file.
- Train the GPT2-ACS model with Squad1.1 dataset using Step_1_QG_train_gpt2.sh
- Evaluate model with Step_2_Generate_Evaluate.sh