Latent Diffusion Models for Generation of Neural Network Parameters

Research project examining the power of latent diffusion models for neural network parameter generation from hyper-representations.

Description

In this repository we are implementing a novel approach to "learning" neural network parameters. For this we adapt the approach of latent diffusion models to model the learning process of neural networks. We do this by utilizing the LDM from Rombach et al. 2022 and use it together with the ModelZooDataset which we transform with the help of Kschuerholt et al.'s Hyperrepresentations.

Development

To start development start with

pip install -e ".[develop]"

Note

This project has been set up using PyScaffold 4.0.2. For details and usage information on PyScaffold see https://pyscaffold.org/.