Skip to content

Latest commit

 

History

History
90 lines (67 loc) · 5.61 KB

README.md

File metadata and controls

90 lines (67 loc) · 5.61 KB
Raw

Prior Hessian

Library for fast computation of log-likelihoods and derivatives of multivariate priors defined as composites of univariate multivariate distributions especially designed for support of Archimedean copulas.

Documentation

The PriorHessian Doxygen documentation can be build with the OPT_DOC CMake option and is also available on online:

Installation

The PriorHessian library uses CMake and is designed to be installed either as a native package, or as a relocatable package in an arbitrary install prefix. The default build script will install to the _install directory underneath the repository root.

  • Default release build script
    • builds to ${BUILD_PATH} if set or default of: ${CMAKE_SOURCE_DIR}\_build\Release
    • installs to ${INSTALL_PATH} if set or default of: ${CMAKE_SOURCE_DIR}\_install
$ ./build.sh <cmake-extra-opts>
  • Default Debug build script
    • builds to ${BUILD_PATH} if set or default of: ${CMAKE_SOURCE_DIR}\_build\Debug
    • installs to ${INSTALL_PATH} if set or default of: ${CMAKE_SOURCE_DIR}\_install
$ ./build.debug.sh <cmake-extra-opts>
  • Manual CMake Build
$ cmake -H. -B<build-path> -DCMAKE_INSTALL_PREFEX=<install-path> -DCMAKE_BUILD_TYPE=Release <other-cmake-args>
$ cd <build-path> && make -j<num-procs> install

Dependencies

Note the OPT_BLAS_INT64 CMake option controls whether Armadillo uses BLAS and LAPACK libraries that use 64-bit interger indexing. Matlab uses 64-bit by default, and to link PriorHessian to Matlab MEX libraries, this option must be on. Many linux systems only provide 32-bit integer versions of BLAS and Lapack, and the option can be disabled if Matlab support is not a concern and 64-bit support is difficult to manage on

CMake Options

Standard options

  • BUILD_SHARED_LIBS - Build shared libraries [Default: On]
  • BUILD_STATIC_LIBS - Build static libraries [Default: On]
  • BUILD_TESTING - Build tests [Default: On if CMAKE_BUILD_TYPE=Debug]
  • OPT_INSTALL_TESTING - Install tests. [Default: Off]
  • OPT_DOC - Build and install documentation (enables make doc and make pdf) [Default: Off]
  • OPT_EXPORT_BUILD_TREE - Enable CMake export and find_package(BacktraceException) support from the build-tree.

Dependency options

  • OPT_BLAS_INT64 - Enable 64-bit integer BLAS library support [Default: Off]

External Projects

These packages are specialized CMake projects. If they are not currently installed on the development machines we use the AddExternalDependency.cmake which will automatically download, configure, build and install to the CMAKE_INSTALL_PREFIX, enabling their use through the normal CMake find_package() system.

Motivation

For many likelihood-based methods, they can be extended to Bayesian methods like MAP Estimation and MCMC Posterior sampling, by incorporating a prior. This prior must provide fast methods for computing log-likelihood and it's derivatives over the parameter space. The prior log-likelihood, as well as it's gradient and hessian are then added to the equivalent quantities from the likelihood to create a Bayesian objective for MAP Estimation.

Static Polymorphism

The PriorHessian library is designed using static polymorphism (templates), and as such avoids virtual functions for small-grained tasks, and instead uses templates, which allow many small functions to be inlined. This aggressive inlining by the compiler produces log-likelihood, gradient, and hessian functions that are nearly as fast as hand-coded functions. But our flexible CompositeDist class is able to be easily created with any mix of UnivariateDist and MultivariateDist elements.

In Mappel we use this ability to create heterogeneous priors for each Model's parameters, (e.g., [x,y,I,bg,sigma]).

Functionally, the PriorHessian library stores sequences of distributions as std::tuples. Using this approach as opposed to the runtime polymorphism of using std::vector<std::unique_ptr<Base>> gains several advantages. Most importantly, without the need for virtual functions, the tuple-based approach has the ability to inline the many small computational functions that must be combined for every call to compute the log-likelihood or other computationally important quantities.