vitamin-b 0.2.0

A user-friendly machine learning Bayesian inference library

     [](https://travis-ci.org/hagabbar/vitamin_b)

# [VItamin_B: A Machine Learning Library for Fast Gravitational Wave Posterior Generation](https://arxiv.org/abs/1909.06296) :star: Star us on GitHub — it helps!

Welcome to VItamin_B, a python toolkit for producing fast gravitational wave posterior samples.

This [repository](https://github.com/hagabbar/vitamin_b) is the official implementation of [Bayesian Parameter Estimation using Conditional Variational Autoencoders for Gravitational Wave Astronomy](https://arxiv.org/abs/1909.06296).

Hunter Gabbard, Chris Messenger, Ik Siong Heng, Francesco Tonlini, Roderick Murray-Smith

Official Documentation can be found at [https://hagabbar.github.io/vitamin_b](https://hagabbar.github.io/vitamin_b).

Check out our Blog (to be made), [Paper](https://arxiv.org/abs/1909.06296) and [Interactive Demo](https://colab.research.google.com/github/hagabbar/OzGrav_demo/blob/master/OzGrav_VItamin_demo.ipynb).

Note: This repository is a work in progress. No official release of code just yet.

## Requirements

VItamin requires python3.6. You may use python3.6 by initializing a virtual environment.

` virtualenv -p python3.6 myenv source myenv/bin/activate pip install --upgrade pip `

Make sure to install basemap prior to installing all other packages.

For installing basemap: - Install geos-3.3.3 from source - Once geos is installed, install basemap using pip install git+https://github.com/matplotlib/basemap.git

For other required packages: ` pip install -r requirements.txt `

Install VItamin using pip: ` pip install vitamin-b `

## Training

To train an example model from the paper, try out the [demo](https://colab.research.google.com/github/hagabbar/OzGrav_demo/blob/master/OzGrav_VItamin_demo.ipynb).

Full model definitions are given in models directory. Data is generated from gen_benchmark_pe.py.

## Results

We train using a network derived from first principals: 

We track the performance of the model during training via loss curves: 

Finally, we produce posteriors after training and other diagnostic tests comparing our approach with 4 other independent methods:

Posterior example: 

KL-Divergence between posteriors: 

PP Tests: