About The Project

This project aims to uncover the latent patterns in the space generated by two data-generating functions (Checkerboard and Gaussian_mixtures) by leveraging two different methods: Denoising Score Matching (DSM) and Variational Autoencoder (VAE). The learned model is then utilized to generate new data points that remain within the same latent space, preserving the underlying structure and distribution of the original data.

(back to top)

Built With

• Python
• Jupyter Notebook

(back to top)

Prerequisites and Installation

Models are based on JAX and these key packages need to be installed in advance.

• jax

  pip install -U jax

• flax

  pip install flax

• chex

  pip install chex

• optax

  pip install git+https://github.com/google-deepmind/optax.git

• haiku

  pip install git+https://github.com/deepmind/dm-haiku

(back to top)

Project Structure

generative-model-project/
├── args.py                           # Model configs
├── data_generation.py                # Custom data generation
├── model.py                          # Implementation of the DSM & VAE
├── generative_model_notebook.ipynb   # Main file: training, evaluation, visualization
└── README.md                          

(back to top)

Output Examples

Training loss over time

Generated Samples

(back to top)