This repository hosts a pytorch project that trains and evaluates a Visual Question Answering (VQA) deep learning model. For details and references, refer to the PDF report found at the top directory level of the repository.
This project requires Python 3.6.7. To install the python dependencies, run the following command (we recommend using a virtual environment that supports Python 3.6.7):
pip install -r requirements.txtThe dataset annotations can be found in the annotations directory. To download the dataset, run the setup script from the top directory level of the repository as follows:
chmod +x setup.sh
./setup.shThis script will also create all the directories that are needed to train and test your models.
It is recommended that you train your models on a machine with access to GPU resources. To train your models, run the following python script:
python main.pyTo inspect the parameters you can configure for training, execute python main.py --help. The debug flag should only be set when you want to debug your pipeline.
For more details on model training, refer to the project report.
It is recommended that you evaluate your models on a machine with access to GPU resources. To evaluate your models, run the following python scripts:
python main.py --model {0}
python eval.py{0} should be replaced by the name of the file in the models directory that stores the weights of the pytorch model you want to evaluate. Models are automatically saved during training.
For more details on model evaluation, refer to the project report.
Feedback and comments are highly appreciated. Feel free to open an issue on the repository page to express your thoughts and suggest changes.