This project explores various generative models for image classifications as well as extensions of their methodology. The focus is on the VAE and DDPM architectures.

Main Notebook

As a main notebook is required for this project, we have provided one: Main Notebook. As many of the experiments do not make sense to run in interactive Python, we plotted data and called make commands from the notebook to show how our work was done.

Requirements

Download requirements:

make requirements

Note

You should setup a venv or conda environment for the project

Downlaod data

To download the data and process it, run the following command:

make get-data

To get the pokemon fusion data, run:

make get-fusion-data

Checkpoints and samples

All checkpoints and samples can be found on Google Drive

Jobscripts

We have provided shell scripts to submit jobs to the DTU HPC cluster in /jobscripts.

Other useful make commands

We have also made make commands for training models, sampling and calculating FID-scores. These should be self-contained and examples are also shown in the main notebook.