A package for model fitting and Bayesian evidence calculation.
(In case you are wondering what that is about take a
quick look at [this example.]
Citation index for the BayesicFitting package:
- 29 Oct 2021 version 2.8.0
- New class: BernoulliErrorDistribution and SoftMaxModel, tests, examples and data
- Adaptations To BernoulliED in some other classes
- Multi-dim input and output issues
- Updated some other tests and examples
- Documentation and other small issues.
The BayesicFitting package is a python version of the the fitter classes
in Herschel Common Science System (HCSS). HCSS was the all encompassing
software system for the operations and analysis of the ESA satelite
Herschel. The HCSS version was written
in JAVA mostly by me. I encoded features and classes that were requested
by my Herschel colleagues or that I remembered having used myself during
my lifelong career as data analyst for earlier satellites as IRAS, ISO
and AKARI. So most of the stuff in here was needed and used at a certain
moment in time. Even now the package is developing in directions that
are needed by my work for the James Webb Space Telescope (JWST).
The HCSS system is in the public domain under GPL3. It was used by the 3
instrument groups of the Herschel satellite to write calibration and
analysis software. Since the end of the mission HCSS is not being
I used a customized version of java2python (j2py on github) to translate
the JAVA classes to python. However, the actual code needed serious
pythonization. Every line has been inspected. Every construct has been
The documentation got most profit from the automated conversion. Also
the structure into classes, the inheritance, methods and dependencies
are largely the same as in the original HCSS.
The package is written in python3 although I am not aware of using any
specific python3 features. It uses numpy (>= 1.9) for its array
structure, scipy (>=1.0) for linear algebra and other stuff and astropy
(>=2.0) for units. Matplotlib (>=2.0) is used for plotting.
Download and unpack the BayesicFitting zip file from github. Move into
the BayesicFitting-master directory and run:
python setup.py install
where python is python3. Or install it as :
pip install BayesicFitting
The BayesicFitting package consists of over 100 classes, each class in
its own file. These classes can be divided into 3 broad categories:
models, fitters and nested sampling. About 50 models, 10 fitters and the
remainder is needed to run the nested sampling algorithm. All these
classes are in a directory BayesicFitting/source. A special type of
functions are found in BayesicFitting/source/kernels. They can be used
to construct a model.
a number of scripts can be
found to exercise the classes. They are in the form of jupyter
notebooks. Some are using real data; others have synthetic data
specially constructed to make some point.
All examples can be inspected by clicking on them. They will fold out in
To actually exercise the examples and maybe adapt then, start a jupyter
notebook in your examples directory.
The program will open a list in your webbrowser where you can select a
notebook file (.ipynb), which can be run.
In the [documenation]
directory a number of documents can be found.
A first draft of a manual. It obviously needs more work.
A list of troublesome situations and what to do about it.
A list of the terms used throughout this package, with explanations.
An architectural design document, displaying the relationships between
A few notes on my style of code and documenation.
A list of external references for BayesicFitting.
Almost all classes have a test harness. These are located in
BayesicFitting/test. They can be execised as:
python -m unittest <file>
where python refers to python3 and file refers to one of the files in
As most functionality is tested in a test harness, examples on how to
use the classes can be found there too.
A package like this is never finished. Always more classes and/or
functionalities can be added. I present it now as it is in the hope it
will be usefull and it will generate feedback.
According to Wikipedia -> "Software release life cycle" it is called
"Perpetual Beta". It continues to be in a beta-release because new
classes and features can be added.
More work needs to be done in:
- Introduction of more Classes: NeuralNetModel, OrderProblem, ...
4 Jan 2018 version 0.9.0.
- Initial upload to github.
26 Jan 2018 version
5 Mar 2018 version 1.0.1
- Package on pypi.com.
- Restructured all import statement to comply with PYPI package.
14 Mar 2018 version 1.0.2
- Added Dynamic Models
- Added piping of models
23 Mar 2018 version 1.0.3
- Some issues with ErrorDistributions and map fitting
- 2-d fitting examples added
- All examples revisited
- Links in README.md updated
28 May 2018 version 1.0.4
- New classes: CircularUniformPrior, PseudoVoigtModel
- VoigtModel uses scipy.special.wozf() and has partials now,
- Refactoring Priors to the BaseModel
- Restructuring Dynamic
- Threading optional in NestedSampler.
- New classes: UniformErrorDistribution, FreeShapeModel and kernels/Tophat
- added to testharnesses and examples
27 June 2018 version 1.0.5
- New classes: RadialVelocityModel and MixedErrorDistribution
- testharnesses and examples
- documentation updates
28 June 2018 version 1.0.6
- longdescription set to markdown (Still not OK on pypi.org)
28 July 2018 version 1.0.7
- small compilation error in 1.0.6
11 October 2018 version 1.0.8
- refactoring the setting of attributes in Models
- documentation (manual, design, etc.) updated.
28 December 2018 version 2.0.0
- Introduction of Problem Classes:
Base class for problems to be handled by NestedSampler.
Common class for everything that was possible in version 1.
ClassicProblem is transparant as all interfaces to NestedSampler have remained
the same as they were in version 1.0, even though behind the scenes a
ClassicProblem has been invoked.
Problem that have errors in the xdata and in the ydata.
- ... more to come.
- Introduction of Walker and WalkerList to represent the internal ensemble
- Adaptations in NestedSampler, ErrorDistributions, Engines, Sample, SampleList.
- Better separation of responsibilities of ErrorDistribution and Problem.
Consequently ErrorDistribution has a new initialisation, which is incompatible
with previous versions. In most cases this has no effect on the calling
sequences of NestedSampler.
- Rename GenGaussErrorDistribution into ExponentialErrorDistribution.
- New testharnesses and examples.
- Adaptations of documentation: manual and design.
16 Jan 2019 version 2.1.0
Problems with more dimensional outputs
A 2 dim output model to calculate the orbit of a double star
To calculate the radius and true anomaly according to Kepplers 2nd law.
- RadialVelocityModel: adapted to Kepplers2ndLaw. A slight change in the
order of the parameters.
- NestedSampler: some improvements in output layout.
- New tests, examples and updates for documentation.
7 Feb 2019 version 2.2.0
- ChordEngine. Implementation of the POLYCHORD engine, developed
by Handley etal. (2015) MNRAS
- OrthogonalBasis. Helper class fot ChordEngine.
- Tests and examples
19 Feb 2019 version 2.2.1
- AmoebaFitter still mentioned GenGaussErrorDistribution; replaced
- Some documentation issues repaired.
20 Jun 2019 version 2.3.0
- Add LogisticModel and SampleMovie
- Periodic residuals in Problem
- Small issues repaired
- Rerun all examples
- Pictures moved to documentation/images
- Some documentation issues repaired.
14 Nov 2019 version 2.4.0 to 2.4.2
- New Classes:
A DecisionTree Model (DTM) is mostly defined on multiple input dimensions (axes).
It splits the data in 2 parts, according low and high values on a certain input axis.
The splitting can continue along other axes.
Interface to define modifiable behaviour of some Models.
Engine to modify Models that implement Modifiable
- Introduce Table from astrolib as (multidimensional) xdata
- Some restructering necessitated by the classes above.
- Testcases and examples for the classes above
3 Feb 2020 version 2.4.3
- Clean up and unification of the python doc strings.
- Reran all examples and test harnasses in python 3.7.
- Add random seed to several examples to make them more stable.
17 Mar 2020 version 2.4.4
- Moved BayesicFitting/BayesicFitting/documentation to BayesicFitting/docs
- Added a references.md file which collects (external) references.
- Updated the docs files.
- Handling of weight in accordance with the definition in the Glossary.
- Add keyword tail= to formatter to display last items of an array.
4 Jun 2020 version 2.5.0
- Add new models: BasicSplinesModel and SplinesDynamicModel
- Option for constraints on the likelihood
- Option for slow engines (working every slow-th iteration in NestedSampler)
- Restructure growPrior setting
- Print formatting in NestedSampler
- Adapt to SplinesDynamicModel
- Homogenized and improved plotoptions in test harnesses
- Three more examples added
5 Jun 2020 versions 2.5.1
- Comment out NeuralNetModel (not yet available) and some typos.
6 Jun 2020 versions 2.5.2
- Two more bugs smashed (in StartEngine and Prior)
29 Jun 2020 versions 2.5.3
- Averaging of circular variables
- Update of static class attributes
- Attribute and printing issues.
23 Oct 2020 version 2.6.0
- New class: PhantomSampler; adaptations in Engines, Explorer, WalkerList
- Restructured NestedSampler to accommodate PhantomSampler
- Test harnass for PhantomSampler
- Option: fix parameters in BasicSplinesModel
- Confusing str method in compound models improved
6 Nov 2020 versions 2.6.1
- Avoid infinities in unbound Priors
- mcycles in initialization of MonteCarlo
- convert xdata, ydata, weights using numpy.asarray
11 Dec 2020 versions 2.6.2
- Add limits and circular to Priors
- Finetune Engines
18 Feb 2021 version 2.7.0
- New class: EvidenceProblem & ModelDistribution; adaptations in NestedSampler and tests.
- Change in constrain method definition
- decay in ExpModel
- some seldom errors, clean-up & new test harnasses.
19 April 2021 version 2.7.1
- remove CrossEngine completely
20 April 2021 version 2.7.2
- put some tests on hold
- few minor issues/errors
29 Oct 2021 version 2.8.0