Gacs-Korner Common Information Variational Autoencoder

Official Pytorch Implementation for Gacs-Korner Common Information Variational Autoencoder (NeurIPS 2023) .

Quick Start

The main training script is main.py.

For example, to train on multiple views of dsprites:

python main.py cvae_ddsprites_randSample_klqq=0.1_klu=10_epoch=70_batch=128_z=8_zu=3_seed=0 -s 0 -d ddsprites -m Doubleburgess -md Doubleburgess -l CVAE --lr 0.001 -b 128 -e 70 -z 8 -zu 3 --gamma-klu 10 --gamma-klqq 0.1

These commands (and the others used in the paper) are generated in scripts/generate_paper_commands.py.

Evaluating Information and Disentanglement

The main evaluation script is main_eval.py. For example, run:

python main_eval.py --name cvae_ddsprites_randSample_klqq=0.1_klu=10_epoch=70_batch=128_z=8_zu=3_seed=0 --dataset ddsprites --nu 3 --nz 8 --num-factors 5

Note that --num-factors 5 corresponds to the number of ground-truth latent factors.

Generating visualizations

The main visualization script is main_viz.py. To generate traversals, run:

python main_viz.py cvae_ddsprites_randSample_klqq=0.1_klu=10_epoch=70_batch=128_z=8_zu=3_seed=0 traversals -r 8

Plotting scripts (which show how to load logged data) begin with plot_*. For example, run:

python plot_disentanglement_hinton.py --result-dir results --name cvae_ddsprites_randSample_klqq=0.1_klu=10_epoch=70_batch=128_z=8_zu=3_seed=0

Visualizations, plots, and metrics will be in results/{name}

Requirements

• python 3.6+
• torch
• torchvision
• scipy
• seaborn (for plotting)

This repository builds off: https://github.com/YannDubs/disentangling-vae

---

If you find this useful for your work, please consider citing

@inproceedings{
kleinman2023gacskorner,
title={Gacs-Korner Common Information Variational Autoencoder},
author={Michael Kleinman and Alessandro Achille and Stefano Soatto and Jonathan Kao},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=e4XidX6AHd}
}