Pypmc

Improve the CI support and improve building and testing wheels across more platforms and architectures.

Changes to the build system
* Require cython to build the code to avoid brokens builds with newer version of python/cython
* Use ``cibuildwheel`` to build and test wheels.
* Tidy up test cases to work with ``cibuildwheel``
* Build wheels for platforms ``linux`` (archs ``x86_64`` and ``aarch64``) and ``macOS`` (archs ``x86_64`` and ``arm64``)

Improve the CI support and support building and deploying wheels

Fixed issues
62

changes to the build system
* Require cython to build the code to avoid brokens builds with newer version of python/cython
* move to xenial and test on python 3.7
* add doc and deploy stages
* test mpi with just one process due to travis limitations
* build wheels for linux and deploy to Pypi
* declare pypmc stable on pypi

Changes to the functional core of pypmc
None

No change to the actual source code was done but it is now easier to install `pypmc` from pypi. If `numpy` is not available, it will be properly installed before at attempt is made to build `pypmc` itself.

The basic ground work for binary wheels for multiple platforms via `cibuild` is done, too. It requires a final tweak and travis to actually build and deploy.

===========

This is a maintenance release.

Infrastructure
--------------

`pypmc` now runs on travis and the docs are automatically built and deployed to http://fredros.github.io/pypmc/html/. Every release is tagged on zenodo.

Fixes
-----

* explicit integer/float handling for compatibility with numpy v1.13
* update color map for compatibility with matplotlib v2.0

[](https://zenodo.org/badge/latestdoi/15123/fredRos/pypmc)

New features
- `ImportanceSampler` optionally stores the values of the target density
- `plot_mixture` can color components based on their weight rather than their index
- the new `PMC` class can run multiple EM updates until convergence

Fixes
- fix installation on mac without cython
- compatible to newer version of numpy
- fix too large a prune value in `GaussianInference`

Breaking changes
- remove `DeterministicIS`. No class was needed, all the functionality is now contained in `combine_weights`
- `MarkovChain` and `ImportanceSampler` don't have a combined `history` anymore, weights and samples are accessible as their own `History` objects

The code has matured and represents the state at the end of Stephan Jahn's master's thesis.
[](http://dx.doi.org/10.5281/zenodo.20045)