Implementation of Visual Question Answering

ECE285_Jarvis_ProjectA

This code is an implementation based on Vahid Kazemi and Ali Elqursh's paper Show, Ask, Attend, and Answer: A Strong Baseline For Visual Question Answering using PyTorch by team Jarvis whose members are listed below. This is part of the final project for the UCSD course Machine Learning for Image Processing in Spring 2019

Pre-trained model (50epoch.pth) is available here

Team members:

• Chandhini Grandhi, cgrandhi@eng.ucsd.edu, A53272378
• Gayathri Cuddappa Sudhinder, gcuddappa@eng.ucsd.edu, A53266365
• Shashank Solomon, s2solomo@eng.ucsd.edu, A53281006
• Srinithya Nagiri, snagiri@eng.ucsd.edu, A53277642

Required Packages (python3):

1. torch
2. h5py
3. tqdm
4. pillow
5. json

For installing these packages, you can use either pip to install packages. For example,

pip install torch

To run a demo:

To demo the model, run the demo.ipynb jupyter notebook

To train the model:

• Change the parameters in config.py
• run python train.py <path-to-save-model>

File Structure:

.
├── efforts                              # Other models we tried
|   ├── Bottom-up-attention-imp          # Implementation of https://arxiv.org/abs/1708.02711 paper
|   ├── san-vqa-tensorflow               # Implementation of https://arxiv.org/pdf/1511.02274.pdf paper 
|   ├── Show_Attend_and_Tell_beam        # Implementation of https://arxiv.org/pdf/1502.03044.pdf paper 
|   └── show_attend_and_tell             # Implementation of https://arxiv.org/pdf/1502.03044.pdf paper 
|
├── resnet                              
|   ├── convert.py                       # Used to build resnet model
|   ├── convert2.py                      # Used to build resnet model
|   ├── resnet.py                        # Used to build resnet model
|   └── synset.py                        # Used to build resnet model
|
├── Visual-Question-Answering_Report.pdf # Project report 
├── config.py                            # Contains parameters for training the model
├── data.py                              # Data loaders for the dataset
├── demo.ipynb                           # Notebook to run the demo
├── dogs.jpg                             # Image used in demo
├── model.py                             # File to build the neural net in
├── preprocess-images.py                 # Extract features from images
├── preprocess-vocab.py                  # Do word embedding
├── tennis.jpg                           # Image used in demo
├── test_img.jpg                         # Image used in demo
├── train.py                             # To train the model
├── utils.py                             # Contains some utility functions
├── vocab.json                           # Vocabulary file created during training
└── README.md