Generative Models as Distributions of Functions

This repo contains code to reproduce all experiments in Generative Models as Distributions of Functions.



Requirements for training the models can be installed using pip install -r requirements.txt. All experiments were run using python 3.8.10.

Training a model

To train a model on CelebAHQ64, run

python configs/config_celebahq64.json

Example configs to reproduce the results in the paper are provided in the configs folder. Note that you will have to provide a path to the data you wish to train on in the config.

Downloading datasets

The shapenet voxels and point cloud datasets can be downloaded at this link. The CelebAHQ datasets can be downloaded from here.

Loading trained models

All trained models can be downloaded from here. The load-trained-model.ipynb notebook shows an example of using a trained model.

Rendering 3D samples

The requirements in requirements.txt allow for basic plotting of 3D shapes with matplotlib. However, to properly render 3D models, you will need to install mcubes (for marching cubes) and pytorch3d. Pytorch3D is not directly pip installable (depending on your version of torch), so please follow the install instructions provided in their repo.

Once these requirements have been installed, you can render samples using

python -sp samples -mr 32 -rs 64 -ns 8 -mp trained-models/shapenet_voxels/

See for a full list of rendering options.