https://github.com/chenlin9/Gaussianization_Flows

https://github.com/chenlin9/Gaussianization_Flows

Last synced: 3 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/chenlin9/Gaussianization_Flows
• Owner: chenlin9
• Created: 2020-03-31T07:48:51.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-05-06T08:43:28.000Z (over 4 years ago)
• Last Synced: 2024-07-12T17:54:10.926Z (4 months ago)
• Language: Python
• Size: 33.2 KB
• Stars: 21
• Watchers: 2
• Forks: 8
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-normalizing-flows - [Code

README

        
# Gaussianization Flows

This repo contains the implementation for [Gaussianization Flows](https://arxiv.org/abs/2003.01941).

-------------------------------------------------------------------------------------

Iterative Gaussianization is a fixed-point iteration procedure that can transform any continuous random vector into a Gaussian one. 

Based on iterative Gaussianization, we propose a new type of normalizing flow model that enables both efficient 

computation of likelihoods and efficient inversion for sample generation. We demonstrate that these models, 

named Gaussianization flows, are universal approximators for continuous probability distributions under some regularity 

conditions. Because of this guaranteed expressivity, they can capture multimodal target distributions without compromising 

the efficiency of sample generation. Experimentally, we show that Gaussianization flows achieve better or comparable 

performance on several tabular datasets compared to other efficiently invertible flow models 

such as Real NVP, Glow and FFJORD. In particular, Gaussianization flows are easier to initialize, 

demonstrate better robustness with respect to different transformations of the training data, 

and generalize better on small training sets.

## Dependencies

* PyTorch

* seaborn

## Running Experiments

### RBIG Experiments

To run RBIG experiments, simply run

``python rbig.py``

### Tabular Dataset Experiments

To download tabular datasets, follow the instructions [here](https://github.com/gpapamak/maf).

To run the experiments, run

```

python tabular_experiment.py --multidim_kernel --usehouseholder

```

and specify the dataset and settings by using the flags

```angular2

--total_datapoints  --process_size --dataset --layer --epoch --lr --batch_size

```