Skip to content

Release ABCpy 0.6.2
Compare
Choose a tag to compare
@statrita2004 statrita2004 released this 09 Apr 16:41
· 62 commits to master since this release
v0.6.2
be58367
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

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
Assets 2