darkmachines / pybambi

Resurrecting BAMBI for the pythonic deep learning era

pyBAMBI

pyBAMBI:	Resurrecting BAMBI for the pythonic deep learning era
Author:	Dark Machines collaboration
Organiser:	Will Handley
Version:	0.2.0
GitHub:	https://github.com/DarkMachines/pyBAMBI
Documentation:	https://pybambi.readthedocs.io/
Website:	https://darkmachines.org
Paper:	https://arxiv.org/abs/1110.2997

Quick links

• Contributing
• Code of conduct

Notes

Currently, there is a test script in test.py, which shows how to call pyBAMBI. The main piece of work to be done is to implement neural network modelling in pybambi/dumper.py from the view onto the live points every nlive iterations.

Installation instructions

• MultiNest installation
• PolyChord installation

You can clone the repository with

git clone [email protected]:DarkMachines/pyBAMBI.git

Although if you wish to contribute, you can either create a new branch and push (if you are a member of the DarkMachines sampling community) or fork the repository to your own account.

You can run the tests with either of the commands:

python -m pytest tests
python setup.py test

pybambi's Issues

Should we make this repository public?

Pros:

• Can use (for free) modern collaborative tools such as
  • Continuous integration
  • Pull Requests
• External contributors easier to integrate
• We can give ownership of this repository to the Dark Machines collaboration, rather than @williamjameshandley

Cons:

• Risks people using / publishing with it before its ready

Continuous integration for documentation

Ideally using sphinx-build

Create basic neural network code

Leading on from #3, PR #9 introduced a simple 'nearest neighbor' neural network to act as a template for better implementations. Any volunteers with experience implementing neural networks would be much appreciated.

Give kerasnet choice of neural network architecture

@mjw63 has done a good job starting off implementing neural networks in keras with #12.

Currently pybambi/neuralnetworks/kerasnet.py has a hard-coded, three layer 200 neuron-per-layer neural network. Ideally there would be more user-choice about this.

Comments/discussion welcome as to what is the most pythonic/re-usable way to provide this functionality. Perhaps allowing __init__ to take model as an argument, where model is a keras-class?

Add pep8 checking to .travis.yml

For example:

https://www.dominicrodger.com/2012/04/29/build-breaking/

Better solutions welcome

Implement BAMBI framework in python

This will require carefully reading section 3.3.0 and 3.4 of the BAMBI paper (skip 3.3.1 and 3.3.2, as they are skynet-specific). Requires a placeholder function to be delivered from #3

Define a default neural network architecture

We need to try and define a default neural network architecture, loss function, training scheme and activation function, that works reasonably well on a variety of likelihood functions. This can be tested in parallel with #4 using an external script which @mjw63 can supply. It relies on samples from #15 being available.

Also involved: @melli1992

100% test coverage

The newest functionality is not fully covered by the test suite. Any help in increasing coverage would be much appreciated

Bayesian neural nets and uncertainty measures

@bstienen has previous experience with Bayesian neural nets + related uncertainty measures, and can test these in an external code once #15 is finished.

Improve documentation

After milestone 0.2.0 the documentation is in need of checking and expanding.

Test suite should include MPI and OSX

To do this, we need to to update the .travis.yml file, making sure that mpi libraries are installed (both on system, and in mpi4py).

This should be done as a build matrix, i.e. for compiling with and without MPI using environment variables

Define interesting likelihood functions for testing

This can be done in parallel to #4 , in order to try to find a good neural network architecture @melli1992

Add CircleCI continuous integration

Write framework for neural network code

This should be placed in pybambi/neural_networks.py, and be called in dumper. A placeholder 'identity with noise' neural network would do for testing the bambi setup, but we should ideally have a keras-based framework for choosing a neural network