Skip to content

Latest commit

 

History

History
109 lines (102 loc) · 5.78 KB

CHANGELOG.md

File metadata and controls

109 lines (102 loc) · 5.78 KB
Raw

Changelog

Changes in version 3

  • Changes to the client code API:
    • relaxed requirements on return type of configuration();
    • require member function random_configuration();
    • require static member function config_policy_from_parameters(), a factory function for the configuration policy (see below);
    • retired member function configuration_size();
    • retired member types phase_classifier and phase_sweep_policy_type.
  • Split config_policy (formerly exclusive to the gauge client code) into an inheritance hierarchy of reusable classes:
    • abstract tksvm::config::policy interface class;
    • tksvm::config::monomial_policy handles index rearrangement, block structure as necessary when using the tensorial kernel at given rank;
    • tksvm::config::clustered_policy handles calculation of feature mapping for a given ClusterPolicy, where the latter defines how spin components within a chosen spin cluster geometry are retrieved from the raw configuration;
    • a policy that is specific to the client code; can be a type alias to one of the former or a derived class.
  • The *-learn programs have each been split into two separate programs: *-sample and *-learn. The former collects samples and stores these in the *.clone.h5 file, the latter performs the actual SVM optimization.
  • Parallelization has been revamped:
    • previous OpenMP-based parallelization has been replaced by one based on MPI;
    • support for Parallel Tempering (PT);
    • simulation class required to inherit from base class pt_adapter<phase_point> in lieu of alps::mcbase if PT is to be used, otherwise from embarrassing_adapter<phase_point> for embarassingly parallel sampling;
    • uses of the measurements object (formerly a protected member inherited from alps::mcbase) must be replaced by calls to the measurements() function;
    • only the programs *-sample and *-test are parallelized.
  • Sweep policies:
    • added nonuniform_grid, log_scan, and sweep_grid policies;
    • removed gaussian_temperatures, uniform_temperatures (previously deprecated in version 2).
  • Optional "LAZY" mode: rather than eagerly mapping raw spin configurations sampled from the simulation to the tensorial features and storing those in *.clone.h5, this may be deferred to the invocation of *-learn. This allows for a systematic study of multiple different TK ranks and cluster choices from a single run of *-sample but comes at the cost of having to store the raw spin configurations. This behavior can be enabled by setting the build option CONFIG_MAPPING to LAZY. The old behavior, "EAGER", remains the default.
  • The classifier is no longer chosen at compile time using a build option, but at run time using the classifier.policy parameter.
  • The fixed_from_sweep classifier replaces both fixed_from_cycle and fixed_from_grid and is applicable to arbitrary sweep policies.
  • New client code frustmag: highly generic code for classical MC simulations of frustrated spin systems.
  • Features added to *-learn program:
    • flag --infinite-temperature allows for inclusion of fictious spin configuration samples corresponding to completely uncorrelated "infinite temperature" spins (obtained via the random_configuration() function) as a control group;
    • flag --statistics-only applies the classifier, prints label statistics, but foregoes the SVM optimization;
    • flag --skip-sampling removed, as *-learn no longer samples;
    • flag --merge allows to combine the samples of multiple *.clone.h5 files into one big optimization problem.
  • Features added to *-test program:
    • The set of points at which the SVM decisions function(s) and observables are to be measured is no longer constrained to a simple line in parameter space but can be specified in the form of a sweep policy using parameters test.policy and the like.
    • program can be launched with an *.ini parameter file, in which case no SVM decision functions are measured, only the observables registered and measured by the simulation class;
    • calculation and output of variances on decision functions and observables (provided also their square is measured);
    • *.test.txt output files are self-documenting: a comment annotating their column layout is automatically generated.
  • Changes to the *-coeffs program:
    • --exact and --diff are now conditional on the presence of the --result=<result-name> flag, specifying the name of the (hardcoded) reference result that is to be compared to.
  • Features added to the graph analysis (*-segregate-phases):
    • binary decision on whether or not to include an edge in the graph is superseded by an option --weight to specify a weighting function which maps biases to continuous edge weights in [0, 1]; default value box reproduces previous behavior;
    • multiple graphs may be combined into one (by multiplication of edge weights)
    • support for masks (--threshold, --invert-mask, --mask, --masked-value)

Changes in version 2

  • Support for multiclassification
    • in upstream SVM wrapper library
    • all decision functions are being measured in *-test
  • Gauge client code uses both J1 and J2 as parameters which locate the model in the phase diagram
    • introduced sweep policy concept to traverse phase diagram in varied ways
    • introduced classifier concept to map points in the phase diagram to labels which are used as input for the SVM
  • Spectral graph partitioning analysis using separate program: *-segregate-phases