Vector Quantized Variational Autoencoder

This is an implementation of Vector Quantized Variational Autoencoder for the final project of ECE 176: Introduction to Deep Learning & Applications at UC San Diego.

Environment Setup

First, ensure you have a clean Python distribution that has version >=3.11. One way of doing it is through miniconda:

conda create --name vqvae python=3.12
conda activate vqvae

Then, you can install poetry:

pip install poetry

Finally, install the dependencies and the package itself:

poetry install

To download the CelebA dataset, you might need the library gdown. You can install it with:

poetry install --with download

when installing the package.

Training

To train the model, you can run (assuming you're in the root of the repository):

python scripts/train/cifar10.py
python scripts/train/celeba.py

The hyperparameters can be changed directly in the script files. The results will be saved in the logs directory.