This code accompanies the proximity variational inference paper: https://arxiv.org/abs/1705.08931

If you use this code, please cite us:

@article{altosaar2017proximity,
	author={Altosaar, J and Ranganath, R and Blei, DM},
	eprint={arXiv:1311.1704},
	title={Proximity Variational Inference},
	url={https://arxiv.org/abs/1705.08931},
	year={2017}
}

The promise: Variational inference (left) is sensitive to initialization. Proximity variational inference (right) can help correct this.

Get the binarized MNIST dataset from Hugo & Larochelle (2011), write it to /tmp/binarized_mnist.hdf5.

python get_binary_mnist.py

I recommend anaconda: brew cask install anaconda on a mac, bash installer otherwise. To use the same environment:

conda env create -f environment.yml  # may need to edit to choose between CPU or GPU version of tensorflow
source activate proximity_vi

The code assumes you have set the following environment variables. This enables easy switching between local and remote workstations.

> export DAT=/tmp
> export LOG=/tmp

This benchmarks proximity variational inference against deterministic annealing and vanilla variational inference, with good initialization and bad initialization (Tables 1 and 2 in the paper).

Each experiment takes about half a day on a Tesla P100 GPU:

./sigmoid_belief_network_grid.sh

# List final estimates of the ELBO and marginal likelihood
tail -n 1 $LOG/proximity_vi/*/*/*.log

# View training statistics on tensorboard
tensorboard --logdir $LOG/proximity_vi

This tests the orthogonal proximity statistic to make optimization easier in a variational autoencoder. (Table 3 in the paper)

Each run takes a few minutes on a Tesla P100 GPU:

./deep_latent_gaussian_model_grid.sh

# List final estimates of the ELBO and marginal likelihood
tail -n 1 $LOG/proximity_vi/*/*/*.log

# View training statistics on tensorboard
tensorboard --logdir $LOG/proximity_vi

Please email me with any questions: altosaar@princeton.edu.