Changelog
[1.2.2]
Added
- tests for Julia 1.9 in CI matrix
- HDF5.jl v0.17 compatibility
[1.2.1]
Added
- tests for
fitprior2
in CI pipeline.
Fixed
- error when constructing
MCMCData
with input distance matrix that has non-zero diagonal entries. - edge case for `fitprior2' when
Kmin = Kmax = 1
orKmin = Kmax = N
generates a warning and fallback to default repulsion/cohesion parameters.
[1.2.0]
Added
- added function
fitprior2
, alternative method to fit prior hyperparameters.
Changed
- pretty print for
MCMCResult
also shows whether the model includes repulsion. - moved
basic_example.jl
from theexamples
folder todocs
. - removed
examples
(added to a separate branch).
[1.1.0]
Added
- added function signature for
sampleK
to accept individual parameters.
[1.0.1]
Documentation now updated to use Literate.jl to generate the example.
[1.0.0]
Added
- examples from the main paper are now included in the examples folder. Data from the original paper is included in the data folder in the standard HDF5 format.
- the examples now have more plots.
- functionality to sample the distances from the prior predictive distribution (see
sampledist
). - functionality to sample $K$ from its marginal prior predictive (see
sampleK
). generatemixture
accepts a random number generator or a seed for the default RNG for reproducibility of results.
Removed
- dependency on RCall and the various calls to the salso algorithm were removed. It is left to the user to make these calls if necessary.
[0.2.2]
Fixed
- corrected time not printing in the correct format.
- moved
prettytime
andprettynumber
toutils.jl
.
[0.2.1]
Added
- added tests for pretty printing.
- added tests for custom loss function in
getpointestimate
.
Fixed
- verbose output in sampler missing newline.
- custom loss function when computing a point estimate.
Removed
_infodist
not in use, removed.
[0.2.0]
Added
- added log-posterior to result.
- added log-likelihood and log-posterior plots to basic example.
- point-estimation via maximum likelihood and maximum a posteriori.
infodist
function to compute the information distance between clusterings.- convenience constructor for
MCMCData
. - add verbose option for
runsampler
andfitprior
. - pretty printing for
MCMCData
,MCMCOptionsList
, andPriorHyperparamsList
. - added bibliographic references to documentation.
- added changelog.
Breaking
- MCMCOptions constructor changed account for change in point-estimate calculation.
fitprior
now only returns the hyperparameter list.- removed
summarise
forMCMCResult
objects (use pretty printing instead).
Fixed
- corrected computation of log-likelihood in result.
- fixed edge case error in
fitprior
when the found value ofK
is either 1 orN
. This case arises only for edge values ofKmin
andKmax
, since the elbow method will otherwise not choose 1 orN
forK
.
Changed
- added bounds checking for
Kmin
andKmax
infitprior
. - added input message for better debugging in
fitprior
. - added progress bar for
fitprior
ifuseR = false
. - added input validation for
generatemixture
. - added input validation for
binderloss
andevaluateclustering
. - added input validation for the constructors of
MCMCData
andMCMCOptionsList
. - separated computation of log-likelihood and log-prior.
- removed redundant loss function calculations when computing point-estimate.
- Binder loss calculation (
binderloss
) is now approximate (using randindex from Clustering.jl) but faster. runsampler
no longer automatically calculates a point estimate.summarise
has a separate signature for MCMC output and for point estimate summary, and now provides more measures for point estimates.- minor optimisations using
@inbounds
,@simd
, and@turbo
. - using StaticArrays.jl and separate function for restricted Gibbs scan.
- speedup via non-generic implementation of matrix and vector sums with LoopVectorization.jl.