Discouraging Posterior Collapse in Hierarchical Variational Autoencoders Using Context

This repository contains the code for the experiments from the paper:

Discouraging Posterior Collapse in Hierarchical Variational Autoencoders Using Context

Anna Kuzina, Jakub M. Tomczak

---

Abstract

Hierarchical Variational Autoencoders (VAEs) are among the most popular likelihood-based generative models. There is a consensus that the top-down hierarchical VAEs allow effective learning of deep latent structures and avoid problems like posterior collapse. Here, we show that this is not necessarily the case, and the problem of collapsing posteriors remains. To discourage this issue, we propose a deep hierarchical VAE with a context on top. Specifically, we use a Discrete Cosine Transform to obtain the last latent variable. In a series of experiments, we observe that the proposed modification allows us to achieve better utilization of the latent space and does not harm the model's generative abilities.

Experiments

Environment Setup

We list the required packages in conda_requirements.yaml:

conda env create -f conda_requirements.yaml

Weights and Biases Setup

We use w&b to track experiments, save and load trained models, thus it is required to run the code. Variables entity, project and api_key in the configs/wandb/defaults.yaml file should be specified before running the experiments.

Run experiments

Configuration for all the experiments are provided in configs/experiments folder.

For example, to run the experiment with the MNIST dataset and the DCT-based hierarchical VAE, run the following command:

python -u run_experiment.py experiment=mnist_dct_ladder

We use 4-GPUs to train model on the CIFAR-10 dataset (make sure that train.ddp: True):

mpiexec -n 4 python -u run_experiment.py experiment=cifar_dct_ladder

Pre-trained model

Dataset	Model	Test NLL/BPD	Link
MNIST	DCT-VAE	76.62	link
OMNIGLOT	DCT-VAE	86.11	link
CIFAR-10	DCT-VAE	3.26	link

Cite

If you found this work useful in your research, please consider citing:

@article{kuzina2023discouraging,
  title={Discouraging Posterior Collapse in Hierarchical Variational Autoencoders Using Context},
  author={Kuzina, Anna and Tomczak, Jakub M},
  journal={},
  year={2023}
}

Acknowledgements

This research was (partially) funded by the Hybrid Intelligence Center, a 10-year programme funded by the Dutch Ministry of Education, Culture and Science through the Netherlands Organisation for Scientific Research, https://hybrid-intelligence-centre.nl.

This work was carried out on the Dutch national infrastructure with the support of SURF Cooperative.