ywu40 / hilloc

This repository accompanies the paper HiLLoC: Lossless Image Compression with Hierarchical Latent Variable Models by James Townsend, Thomas Bird, Julius Kunze and David Barber, appearing at ICLR '20.

The Craystack software for lossless compression can be found at https://github.com/j-towns/craystack, code to reproduce the experiments in the paper is in the experiments directory.

We recommend using the colab notebook, which demonstrates compression using HiLLoC and a ResNet VAE model. The notebook is a minimal implementation written in JAX, taking the trained weights resulting from the tensorflow implementation in the experiments directory. You can download trained weights for the 4 layer and 24 layer model, as well as a sample cifar image. Feel free to compress your own images!