/score-based-generative-models

Experiments to improve the performance of score-based generative models. Serves as our course project for CS726: Adv ML Course Project. Authors: Pranava Singhal, Aditya Sriram, Keshav Patel Keval, Jujhaar Singh.

Primary LanguageJupyter Notebook

Experiments on Score-based Generative Models to Improve Performance

The following ideas have been implemented and compared by us

  1. Using Hamiltonian MC sampling instead of standard methods to improve performance and accuracy when sampling from the model

  2. Generating continuous embeddings using VAEs for discrete variables in the latent space to train the diffusion models with

  3. Exploiting the sparsity of data using signal processing techniques like wavelet transformations