🧠 Visual Question Answering

This project implements a Visual Question Answering (VQA) pipeline using two model architectures:

• 🔹 LSTM + CNN
• 🔸 Attention

📁 Dataset

The dataset used for this project is the COCO-QA dataset.

🔧 Setup

1. Create a virtual environment

python -m venv vqa_env
source vqa_env/bin/activate

2. Install the required packages

pip install -r requirements.txt

3. Download the dataset

python data/cocoqa_preprocess.py

🧼 Preprocessing

1. Create Question Features

python data/preprocessing.py

2. Create Image Features

Run notebook: data/processing.ipynb on Kaggle/Colab to use GPU for faster processing.

🚀 Training

1. LSTM + Multimodal Fusion

python train_lstm.py --batch_size 16 --max_epochs 1000

2. Attention

python train_attention.py --batch_size 16 --max_epochs 1000

🧪 Evaluation

You can use notebook vqa_main.ipynb for end-to-end training and evaluation.

📄 References

• [1] VQA: Visual Question Answering (Agrawal et al, 2016): https://arxiv.org/pdf/1505.00468v6.pdf
• [2] Hierarchical Question-Image Co-Attention for Visual Question Answering (Lu et al, 2017): https://arxiv.org/pdf/1606.00061.pdf