Skip to content
Pytorch implementation of Block Neural Autoregressive Flow (http://arxiv.org/abs/1904.04676)
Branch: master
Clone or download
Latest commit 1a629a2 May 30, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
data Initial commit third part Apr 5, 2019
optim Initial commit second part Apr 5, 2019
.gitignore Added folders in .gitignore Apr 7, 2019
LICENSE Initial commit Apr 2, 2019
README.md Update README.md May 30, 2019
bnaf.py Small fixes Apr 7, 2019
density_estimation.py Small fixes Apr 7, 2019
download_datasets.sh Small fixes Apr 7, 2019
toy2d.py Small fixes Apr 7, 2019

README.md

BNAF

Pytorch implementation of Block Neural Autoregressive Flow based on our paper:

De Cao Nicola, Titov Ivan and Aziz Wilker, Block Neural Autoregressive Flow (2019)

Requirements

  • python>=3.6 (it will probably work on older versions but I have not tested on them)
  • pytorch>=1.0.0

Optional for visualization and plotting: numpy, matplotlib and tensorboardX

Structure

  • bnaf.py: Implementation of Block Neural Normalzing Flow.
  • toy2d.py: Experiments of 2d toy task (density estimation and energy matching).
  • density_estimation.py: Experiments on density estimation on real datasets.
  • optim: A custom extension of torch.optim.Adam and torch.optim.Adamax with Polyak averaging. A custom extension of torch.optim.lr_scheduler.ReduceLROnPlateau with callbacks.
  • data: Data classes to handle the real datasets.

Usage

Below, example commands are given for running experiments.

Download datasets

Run the following command to download the datasets:

./download_datasets.sh

Run 2D toy density estimation

This example runs density estimation on the 8 Gaussians dataset using 1 flow of BNAF with 2 layers and 100 hidden units (50 * 2 since the data dimensionality is 2).

python toy2d.py --dataset 8gaussians \    # which dataset to use
                --experiment density2d \  # which experiment to run
                --flows 1 \               # BNAF flows to concatenate
                --layers 2 \              # layers for each flow of BNAF
                --hidden_dim 50 \         # hidden units per dimension for each hidden layer
                --save                    # save the model after training
                --savefig                 # save the density plot on disk

Imgur

Run 2D toy energy matching

This example runs energy matching on the t4 function using 1 flow of BNAF with 2 layers and 100 hidden units (50 * 2 since the data dimensionality is 2).

python toy2d.py --dataset t4 \            # which dataset to use
                --experiment energy2d \   # which experiment to run
                --flows 1 \               # BNAF flows to concatenate
                --layers 2 \              # layers for each flow of BNAF
                --hidden_dim 50 \         # hidden units per dimension for each hidden layer
                --save                    # save the model after training
                --savefig                 # save the density plot on disk

Imgur

Run real dataset density estimation

This example runs density estimation on the MINIBOONE dataset using 5 flows of BNAF with 0 layers.

python density_estimation.py --dataset miniboone \  # which dataset to use
                             --flows 5 \            # BNAF flows to concatenate
                             --layers 0 \           # layers for each flow of BNAF
                             --hidden_dim 10 \      # hidden units per dimension for each hidden layer
                             --save                 # save the model after training

Citation

De Cao Nicola, Titov Ivan, Aziz Wilker,
Block Neural Autoregressive Flow,
35th Conference on Uncertainty in Artificial Intelligence (UAI19) (2019).

BibTeX format:

@article{bnaf19,
  title={Block Neural Autoregressive Flow},
  author={De Cao, Nicola and
          Titov, Ivan and
          Aziz, Wilker},
  journal={35th Conference on Uncertainty in Artificial Intelligence (UAI19)},
  year={2019}
}

Feedback

For questions and comments, feel free to contact Nicola De Cao.

License

MIT

You can’t perform that action at this time.