Skip to content

Latest commit

 

History

History
188 lines (134 loc) · 5.33 KB

VERSION.md

File metadata and controls

188 lines (134 loc) · 5.33 KB

LibBi VERSION.md

v1.4.5

Fixes:

  • Fixed OpenMP check
  • Fixed border case for binomial distribution

v1.4.4

Fixes:

  • Fixed a bug in the calculation of binomial densities

v1.4.3

New features:

  • Metropolis-Hastings sampling with Gaussian proposals now work if the standard deviation is zero; this makes it possible to repeatedly propose the same theta, for example to adjust the number of particles, without the need to change the model and recompile.
  • build-dir, version and with-lldb command line options.
  • Parameter or initial condition assignments are now overwritten by a given init file.

Fixes:

  • The input-interval has been changed to run from [t, t+1) so that timings are preserved if output from a previous run is used as input.
  • Fixed compile error in Metropolis resampler.
  • Fixed some CUDA problems.

v1.4.2

Performance:

  • Improved beta and negative binomial random generators.

v1.4.1

Fixes:

  • Fixed compatibility with automake 1.16.
  • Fixes in truncated Gaussian and negative binomial probability distributions.

v1.4.0

New features:

  • New distributions: Poisson, binomial, negative binomial, beta-binomial, exponential.
  • Added --cuda-arch option for specifying the CUDA architecture.
  • Added --enable-cudafastmath option for using fast CUDA maths.

Changes:

  • State and noise variables now can have inputs.

Performance:

  • Improved truncated Gaussian sampler.
  • Models are now only simulated if the parameter prior probability density is >0.
  • Parsing and translating models is now faster.

Fixes:

  • No more error is thrown if all particle weights are 0.
  • optimise now works again.
  • No more warning is thrown if the model name does not match the file name.

v1.3.0

New features:

  • Added SMC$^2$ with anytime moves, enabled with --tmoves command-line option.
  • Added --mpi-hostfile option.

v1.2.0

New features:

  • Streamlined installation instructions: Perl module dependencies are now installed automatically, make test checks for these dependencies, and MAGMA is optional.
  • CPU builds now work on Windows under Cygwin.
  • CUDA builds now work on Mac OS X.
  • AVX support added (use --enable-avx).
  • SMC^2 now works with a Kalman filter over the state, not just a particle filter, and has been renamed marginal SIR to reflect this.
  • Output at observation times can now be disabled (use --without-output-at-obs).
  • Non-const scalar expressions are now allowed as indexes on the left side of actions.

Changes:

  • Distribution now contains a compiled version of the manual.
  • Major refactoring of methods code to reduce replication.

Performance:

  • Improved performance under --enable-sse.

Fixes:

  • Fixed the first observation in a time series being zero when --start-time did not match the earliest time in --obs-file.
  • Fixed CUDA kernel execution configurations for large numbers of variables.
  • Fixed matrix range being ignored on the right side of actions.

v1.1.0

New features:

  • Added bridge particle filter (from Del Moral & Murray, 2014).
  • Added built-in variables t_now, t_last_input and t_next_obs.
  • Added transpose and cholesky actions.
  • Added log argument to pdf action.
  • Added matrix-matrix multiply.
  • Added range syntax, e.g. x[0:4].
  • Added checks for *.sh, data/*.nc and oct/*.m files to libbi package --validate checks.

Changes:

  • Removed ill-defined uninformative action.
  • Action syntax made stricter: an action that returns a scalar cannot be applied to a vector on the left.

Performance:

  • Minor performance improvements in GPU random number generation and resampling.
  • Minor I/O performance improvements.

Fixes:

  • Restored NetCDF 4.1 support.
  • Fixed initialisation of parameters from init file when --with-transform-initial-to-param used.
  • Fixed C_ and U2_ variables in Kalman filter output files.
  • Fixed reporting of log-likelihood in PMCMC output when particle filter degenerates.
  • Fixed build error when model name does not begin with an uppercase letter.
  • Fixed runtime error when empty --output-file given.
  • Fixed race condition in locking the build directory under some circumstances.
  • Fixed unnecessary recompiles triggered by new hash implementation in newer versions of Perl.

v1.0.2

  • Removed dependency on NetCDF C++ interface, the C interface is now used directly.
  • Added 'extended' boundary condition for dimensions.
  • Added --enable-openmp/--disable-openmp command-line options.
  • Added --enable-gpu-cache/--disable-gpu-cache command-line options for better control of GPU memory usage.
  • Added --adapter-ess-rel command-line option to avoid adaptation of proposal in SMC$^2$ when ESS too low.
  • Several bug and compatibility fixes.

v1.0.1

  • Added additional material to manual, including new section with guidance on tuning the proposal distribution and number of particles when using PMMH.
  • Fixed sampling of joint distribution (--target joint now implies --with-param-to-state, just as --target prior and --target prediction do).
  • Fixed reordering of actions and blocks when the same variable appears on the left more than once.
  • Fixed bug in GPU implementation of multinomial resampler.
  • Added --dry-parse option to remove parsing overhead when binaries have already been compiled.

v1.0.0

  • First public release.