Release ABCpy 0.6.3

In this release we have done the following additions and fixes:

• Added a DrawFromPrior inference scheme, with similar API as the other inference schemes and is used to sample from prior with no need to specify observation. That also has another method which can be used to generate simulation-parameter pairs to be used for learning the automatic summary statistics utilities.
• Added a method to the Journal which allows to resample the posterior samples (bootstrap and subsample). That generates a new journal which is returned by the method.
• Added a new GenerateFromJournal class which allows to generate simulations from a given model using parameter values stored in a journal. Together with the previous method this allows to perform predictive check.
• Some reformatting to the Statistics; specifically, I've added the capability to standardize the different statistics by dividing them by their standard deviation on a set of reference simulations.
• Added the Statistics learning with exponential family based on Score Matching.
• Some refactoring of the Journal class
• Added option to not store the simulated dataset in the Journal for SMCABC. That breaks the possibility of restarting inference from the journal but greatly reduces journal size.
• Fixed the MCMC routine (used with BSL and others) when a starting point for the chain is passed.
• Added an example showing how to save to disk and re-load the learned neural network statistics.
• Some minor fixes in docs and tests.

Release ABCpy 0.6.2

In this new release of ABCpy version 0.6.2, we have specifically added the following things:

• LogNormal and Exponential continuous models.
• semiparametric Synthetic Likelihood in the approx_lhd module. We have also reorganized the Approx_likelihood abstract class and added there some methods which are used by all subclasses of that.
• Several divergences between empirical distributions to be used as ABC distances: sliced Wasserstein, gamma divergence, KL divergence, MMD, Energy Distance, Squared Hellinger distance.
• New version of SMCABC following Bernton et al. 2019 (the Wasserstein ABC paper); now SMCABC allows choosing between the original version and this one, which works fine with divergences between empirical distributions. Further, SMCABC now implements both the original MH kernel and two versions of r-hit kernels by Anthony Lee.
• Slightly updated SMCABC and PMCABC inference routines so that restarting the inference from a partial journal file ensures reproducibility.
• W have added the possibility to specify simulation budget rather than n_samples in Rejection ABC. The changes do not impact retrocompatbility. When setting simulation budget, it is now also possible to set a distance quantile instead of explicitly choosing epsilon.
• A new MCMC inference scheme based on Metropolis Hastings algorithm to be used with the Approx_likelihood's, alternative to PMC. The resulting journal class also allows to produce traceplot of the MCMC.
• A way to save the journal file to disk directly from the inference scheme, with a new argument path_to_save_journal. When that is specified, the sequential inference routines save the journal at each sequential step, so that partial results are saved in case the inference does not conclude correctly.
• Improved the default perturbation kernel for discrete variables RandomWalkKernel.
• Some minor fixes and optimizations.
• Small improvement to README highlighting the features of the library
• Improvement to docs
• Travis tests and explicit support for Python 3.9

Release ABCpy 0.6.0

In this release we include the following new features and improvements of ABCpy:

• Improved documentation;
• Better explanation of the examples with comments;
• Addition of a Jupyter notebook;
• Addition of convergence diagnostic tools;
• Addition of Wasserstein distance for ABC;
• Small bug fixes related to Pypi installation.

Release ABCpy 0.5.7

In this release we include the following new features and improvements of ABCpy:

* We now provide tools to learn summary statistics from dataset using neural networks.
  These features presently uses extensions of semi-automatic summary statistics learning 
  and statistics learning based on distance learning.

* We have provided some additional features for the output analysis to create figures of 
  smoothed posterior distribution.

* We have fixed some bugs.

* Some improvements of the documentation.

Release ABCpy 0.5.6

In this release we include the following new features and improvements of ABCpy:

• We now allow the nodes of the Bayesian network to be multivariate random variables, previously only uni-variate random variables were allowed.

• We have provided some additional features for the output analysis, to access the posterior mean at different steps, access to distances computed etc.

• We have fixed some bugs related to transition kernels.

• We have corrected the discrete random walk kernel.

• Some improvements of the documentation.

Release ABCpy 0.5.5

New release enabling MPI-parallelized simulator model using nested-MPI backend.

ABCpy 0.5.4

This release (0.5.4) contains

• bugfixes in PMC algorithm
• better distribution of parallel jobs in PMC
• addition of discrete uniform as a new discrete probabilistic model
• fixing of some deprecation warnings
• updates in ReadMe

ABCpy 0.5.3

This release comes with several small improvements and fixes.

• Fix issue with singular matrix in perturbation kernel
• Add verbose output for inference schemes using Python logging module
• Add exception handling for map phase in MPI backend

ABCpy 0.5.2

This release fixes some bugs.

ABCpy 0.5.1

This release contains bugfixes only.