git HEAD
--------
* Improved documentation of include/dai/bipgraph.h
* Moved example to examples/ and added examples/example_bipgraph.cpp
* Cleanup of matlab interface
* Small improvement of utils/fginfo
* Small cleanup of BP code
libDAI-0.2.2 (2008-09-30)
-------------------------
New features:
* Approximate inference methods now report the number of iterations needed.
* Added damping to various algorithms to improve convergence properties.
* Added more features to utils/createfg for creating factor graphs.
* Added ExactInf class for brute force exact inference.
* [Giuseppe Pasino] Added "logdomain" property to BP, a boolean that controls
whether calculations are done in the log-domain or in the linear domain;
doing calculations in the log-domain may help if the numerical range
of a double is too small.
* [Claudio Lima] Added Max-Product functionality to BP.
* Improved documentation.
Improved architecture:
* Added Exceptions framework.
* Pervasive change of BipartiteGraph implementation (based on an idea by
Giuseppe Passino). BipartiteGraph no longer stores the node properties
(former _V1 and _V2), nor does it store a dense adjacency matrix anymore,
nor an edge list. Instead, it stores the graph structure as lists of
neighboring nodes. This yields a significant memory/speed improvement for
large factor graphs, and is more elegant as well. Iterating over neighbors is
made easy by using boost::foreach.
* Added conditional compilation of inference methods.
* VarSet is now implemented using a std::vector` instead of a
std::set``, which yields a significant speed improvement. Furthermore,
the implementation has been generalized, resulting in the small_set class
which can be used to represent sets of small cardinality; VarSet is the
specialization with T = Var.
* Improved ClusterGraph implementation, yielding significant speedups
for the JunctionTree algorithm on large factorgraphs.
Code cleanup:
* Moved everything into namespace "dai".
* Renamed DEBUG to DAI_DEBUG to avoid conflicts.
* Replaced ENUM2,ENUM3,ENUM4,ENUM5,ENUM6 by single DAI_ENUM macro.
* Removed utils/remove_short_loops and matlab/remove_short_loops.
* Replaced sub_nb class in mr.h by boost::dynamic_bitset.
* Improved index.h:
- Renamed Index -> IndexFor
- Added some .reserve()'s to IndexFor methods which yields a
25% speedup of testregression
- Replaced multind by Permute
- Added MultiFor
- Added State
* New funcstionality of factor.h.
* Moved Properties and MaxDiff frameworks from InfAlg to each individual
inference algorithm, because the Properties framework was not as
convenient as I hoped, and not every inference algorithm needs a maxdiff
variable. Also, replaced some FactorGraph functionality in InfAlg by a
function that returns the FactorGraph. The result is cleaner (less
entangled) code.
* Removed x2x.
* Replaced Complex with real numbers (negative potentials are just too rare
to warrant the additional "complexity" :)).
Miscellaneous improvements:
* Now compiles also with MS Visual C++ (thanks to Jiuxiang Hu) and with
GCC under cygwin.
* Contributions by Peter Gober:
- Renamed variable _N in mr.* for compatibility with g++ under cygwin.
* Misc contributions by Giuseppe Passino:
- removed "using namespace std;" from header files - bad practice;
- moved header files in include/dai and sources in src;
- changed #ifndefs to GNU style;
- added extra warning checks (-W -Wextra) and fixed resulting warnings;
- dai::TProb:
o removed copy constructor and assignment operators (redundant);
o implementation of some methods via STL algorithms;
o added methods takeExp, takeLog, takeLog0 for transformation in-place;
o explicit constructor (prevents implicit conversion from size_t to TProb);
o added operator+,+=,-,-=, with argument T (for calculations in log-scale);
* Misc contributions by Christian Wojek:
- New FactorGraph constructor that constructs from given ranges of factors
and variables;
- Optimization of FactorGraph constructors using tr1::unordered_map.
* FactorGraph constructors no longer check for short loops (huge speed
increase for large factor graphs), nor for negative entries. Also, the
normtype is now Prob::NORMPROB by default.
* Improved MaxSpanningTreePrims algorithm (uses boost::graph).
Interface changes:
* VarSet::
- VarSet::stateSpace() -> nrStates(const VarSet &)
- VarSet( const std::set`` ) -> VarSet( begin, end, sizeHint=0 )
- VarSet( const std::vector`` ) -> VarSet( begin, end, sizeHint=0 )
- removed bool operator||
- operator&&(const VarSet&) -> intersects(const VarSet&)
- operator&&(const Var&) -> contains(const Var&)
* FactorGraph::
- delta(const Var &) -> delta(size_t)
- Delta(const Var &) -> Delta(size_t)
- makeCavity(const Var &) -> makeCavity(size_t)
- vars() -> vars
- factors() -> factors
- removed MakeFactorCavity(size_t)
- removed ExactMarginal(const VarSet &)
- removed ExactlogZ()
- removed updatedFactor(size_t)
- removed _normtype and NormType()
- removed hasShortLoops(...) and removeShortLoops(...)
- WriteToDotFile(const char *filename) -> printDot( std::ostream& os )
- undoProb(size_t) -> restoreFactor(size_t)
- saveProb(size_t) -> backupFactor(size_t)
- undoProbs(const VarSet &) -> restoreFactors(const VarSet &)
- saveProbs(const VarSet &) -> backupFactors(const VarSet &)
- ReadFromFile(const char*) returns void (throws on error)
- WriteToFile(const char*) returns void (throws on error)
- removed hasNegatives()
* RegionGraph::
- nr_ORs() -> nrORs()
- nr_IRs() -> nrIRs()
- ORs() -> ORs
- IRs() -> IRs
* *::Regenerate() -> *::construct()
* Renamed Index -> IndexFor
* Diffs:
- max() -> maxDiff()
- max_size() -> maxSize()
* Prob::max() -> Prob::maxVal()
* Factor::
- max() -> maxVal()
- part_sum() -> partSum()
* toc() in util.h now returns seconds as a double
* VarSet::operator&&
* Properties -> PropertySet
libDAI-0.2.1 (2008-05-26)
-------------------------
Bugfix release.
* added missing cstdio header in util.h
* fixed Properties in MR_CLAMPING_* and MR_EXACT_*
* added description of the factor graph fileformat
* improved Makefile
libDAI-0.2.0 (2006-11-30)
-------------------------
First public release.
0.1.5 (2006-11-30)
------------------
Regressions
- tests/testlcbp and tests/testlcbp are broken.
- EXACT method does not work anymore.
- The Properties framework gives a speed penalty because of the lookup
costs involved; inner loops should be optimized.
General framework
- DAIAlg is now a template class; typedefs for DAIAlg and for
DAIAlg are provided. In this way, we do not have to write "wrapper"
functions to forward functionality from either FactorGraph or RegionGraph
to DAIAlg. Functionality like clamping can be implemented in FactorGraph
and in RegionGraph and no explicit interface is needed in descendants.
- New abstract base class InfAlg added, representing an inference algorithm,
from which DAIAlg inherits. This solves the incompatibility problems of
DAIAlg for different T (e.g. DAIAlg was incompatible with
DAIAlg). More work is required to reduce code duplication
(make FactorGraph part of InfAlg).
- Added generic interface (nrVars(), Vars(), nrFactors(), factor(size_t),
beliefs(), belief(VarSet &), ...) to InfAlg and descendants.
- Added a saveProbs/undoProbs interface to InfAlg and descendants that enables
one to save a few factors, modify them (e.g. clamp them), and then restore them
to their old values. Undo should also init the corresponding messages / beliefs.
This can be used if a given factor graph repeatedly needs to be clamped in
different ways and an approximation method is run for each clamping; using the
saveProbs/undoProbs can give a significant speed increase.
- Switched to a general Properties framework that handles the parameters of
all inference methods in a uniform manner. The Properties class is a map of
several properties in boost::any objects, indexed by their names (strings).
It can read from a stream and write to a stream. It is recursive, in the sense
that a Properties object can hold a variable of type Properties as well.
- Added a generic way of constructing inference algorithms from a factor graph,
name and properties object. Added the newInfAlg function which constructs
the requested object. This is used by LCBP, the Matlab interface and the
command line (test) interface.
- Added a generic enum framework for enum parameters. Although implemented as a
hack, it is quite useful in that it drastically simplifies and reduces the
amount of code for handling enum parameters.
- Provided generic functions for calculating marginals in different ways that
work for all approximate inference methods.
Bugfixes
- Fixed GBP free energy.
- Fixed bug in junctiontree (it didn't init the _vars variable).
- Corrected two bugs in operator&& and operator|| in VarSet (they returned
the logical NOT of what they should return).
- Fixed bug in RegionGraph::RecomputeOR(s).
- Fixed bug in utils/create_dreg_fg:
graph structure was not random for given parameters (forgot to call srand()).
- TreeEP bug workaround: use the complete junction tree instead of a subtree.
- Fixed bug in JTree::HUGIN() and JTree:ShaferShenoy() in case of junction tree
that consists of one outer region only.
- Fixed INIT bug in LCBP2::UpdatePancake().
- Fixed MaxDiffs flow (except for MR).
New functionality
- HAK supports several default cluster choices:
minimal (only factors)
delta (Markov blankets)
loop (all loops consisting of loops consisting of or less variables)
Only the maximal clusters are used as outer clusters.
- Implemented TreeEP. It generalizes the heuristic method described in the
Minka & Qi paper for obtaining a tree with the most relevant interactions to
higher order interactions. Almost all optimizations described in the Minka & Qi
paper are used, except that evidence is passed over the whole tree instead of
relevant subsets (the latter is almost implemented but buggy). Also added
alternative (worst-case) algorithm that uses a maximum spanning tree on the
weighted graph where the weight between neighbours i and j is given by
N(psi,i,j), where psi is the product of all the factors involving both i and j
(which is an upper bound on the effective interaction between i and j).
- Implemented MR (MontanariRizzo) based on Bastian's code, but extended it
to be able to handle connectivities larger than 3 (in principle, up to 32).
It supports different initialization methods (the original RESPPROP,
the CLAMPING method and EXACT which uses JTree) and different update methods
(FULL and LINEAR).
- Implemented LCBP2, an analogon of LCBP which represents pancakes as little
networks and uses some approximate inference method on them for calculating
marginals.
- Now there are several LCBP variants (LCBP, LCBPI, LCBPJ, LCBPK, LCBPL);
LCBPJ works only for pairwise, LCBPK is LCBP improved for higher order
interactions and LCBPL is LCBPI improved for higher-order interactions.
- Wrote one single program utils/createfg for creating various types of
random factor graphs.
- Wrote utility to visualize factor graphs using graphviz.
(it uses the BOOST Program Options library)
- Added fginfo utility that displays some info about a .fg file.
- Implemented Factor::strength function that calculates the potential strength
N(psi,i,j) between variables i and j as described in cs.IT:0504030
- Wrote a general MatLab interface matlab/dai (similar to tests/test);
this unified the matlab functions dai, dai_bp, dai_mf, dai_jt, dai_tep, dai_cvm.
- Added MATLAB routine that returns contraction matrix A for BP convergence analysis.
- Implemented a MATLAB interface ai_potstrength for Factor::strength
- Added Martijn's x2x
Improvements of existing code
- Reimplemented RegionGraph and descendants: a RegionGraph ISA FactorGraph
and also a BipartiteGraph. It now also keeps a map that
associates outer region indices to factor indices (no powers yet, this
is deemed superfluous) and provides functions to recompute (some of) the
outer regions from the factors.
- InfAlg descendants run() methods now stop immediately and return NAN in case
they detect NANs. Only BP does not do NAN checking for performance reasons.
- LCBP now works with factors containing zeroes (by defining x/0 = 0).
- HAK, GBP and DoubleLoop now conform to the standards for verbose reporting,
timing and convergence criteria.
- Implemented logZ() for JTree. It does the calculation during message-passing.
- Marginal2ndO now optionally divides by the single node beliefs (to the power n-2);
hopefully this will give more accurate approximations.
- Marginal and Marginal2ndO (optionally) use the new saveProbs/undoProbs functionality
for a faster way of calculating marginals, which does not require a call to init()
nor cloning the whole object for each clamping. This leads to a significant speedup.
- LCBP (and LCBP2) now have complete flexibility in the specification of the
inner method, i.e. the method used to generate the initial cavity approximations.
One can pass two strings, a name and a properties string, and LCBP simply invokes
newInfAlg to construct the corresponding inference algorithm and uses the generic
marginal functions to approximate cavity marginals.
- Replaced the global "method" variable by local properties and removed ai.h
- Added some methods to Factor (operators *, *=, /, /= with doubles as
second argument, operators -, +=, -= with other Factors as second
arguments, randomize(), RemoveFirstOrderInteractions) and similar
operations to Prob
- Moving towards boost::program_options for handling command line arguments
(tests/test is done).
- Renamed some FactorGraph methods:
nr_vars -> nrVars
nr_factors -> nrFactors
varind -> findVar
factorind -> findFactor
makeFacCavity -> makeFactorCavity
- LCBP_SEQMAXRES has been removed because it did strange things.
- Implemented RandomDRegularGraph
- Implemented JTree::calcMarginal for marginals that are not confined
within one cluster (using cut-set conditioning).
- Added isConnected() method to FactorGraph (some methods do not work with
disconnected factor graphs).
- Pair beliefs are now calculated in a symmetrical way by calcPairBeliefs
- Removed single node interaction "correction" code from clamping methods
- Removed calcCavityDist and calcCavityDist2ndO
- No longer depends on GSL.
- Increased portability by combining platform dependant utility functions
in util.{h,cpp}.
- Wrote *.m files providing help
Testing framework
- Made a new and significantly improved testing framework that provides most
functionality from the command line.
- The basis is provided by tests/test, which uses the newInfAlg functionality
and enables the user to easily compare from the command line different
inference methods on a given factorgraph. All parameters can be specified.
Output consists of CPU time, average and maximum single variable marginal
errors, relative logZ error and MaxDiff().
- tests/aliases.conf contains aliases for standard combinations of methods
and their options (which can be used in tests/test).
- tests/large contains several bash/python scripts that create random factor
graphs, compare several approximate inference algorithms (using tests/test) and
allow for easy visualization of the results using PyX.
- Added several .fg files for test purposes to /tests (e.g. two ALARM versions
alarm.fg and alarm_bnt.fg; testfast.fg, a 4x4 periodic Ising grid for
regression testing).
- Added a regression test to the Makefile which is included in the standard
target. It compares all inference methods on tests/testfast.fg with the
results stored in tests/testfast.out
Miscellaneous
- Expanded all tabs to spaces (":set tabstop 4\n:set expandtab\n:retab" in vim)
- Experimental MATLAB code added for StarEP approximation on cavity
- Renamed project to libDAI and changed directory name accordingly.
- Renamed JunctionTree to JTree.
- Fixed licensing (now it's officially GPL).
- Improved README
revision 252
------------
Functionality
- Added RegionGraph, GBP, CVM and HAK (double-loop).
- Added JunctionTree (with two update algorithms, HUGIN and Shafer-Shenoy), which is a
RegionGraph.
- NormType is now chosen automatically (in case of negative factors, Prob::NORMLINF is used,
otherwise the default Prob::NORMPROB is used). Also, in case of negative factors, the
RegionGraph constructors assign each Factor to a unique outer region instead of dividing
it over all subsuming outer regions. See README for when negative factors are known to work
and when not.
- FactorGraph::FactorGraph(const vector) only gives a warning in case of short loops,
it does not automatically merge factors anymore.
- Removed BP_SEQMAXRESNOCLEAR (all cavity initialization methods now are implicitly NOCLEAR)
- Added MATLAB interface functions ai_readfg, ai_removeshortloops and ai_bp
- Added LCBP-III type that should be equivalent to LCBP-II, but can handle zeroes
in potentials. Note that it is significantly slower than LCBP-II (and has to be reimplemented
such that it does not store the complete pancakes, but represents them as little factor graphs).
Implementation / code
- Made a seperate type WeightedGraph, which until now only implements Prim's
maximal spanning tree algorithm and is only used by the junction tree code. It might
be good to make it a class somewhere in the future and extend it's interface.
- Made a seperate class ClusterGraph, which is only used by the junction tree
code. It's main purpose is a graph-theoretical variable elimination algorithm.
- Implemented the heuristic "minimum-new-edges-in-clique-graph" for variable elimination.
- Massive code cleanup, moving towards "generic" programming style, using
multiple inheritance and polymorphism.
o BP, LCBP, MF, HAK and JunctionTree now inherit from a common DAIAlg class
o Made generic functions Marginal, Marginal2ndO, calcCavityDist, calcCavityDist2ndO, clamp
that can be used by FactorGraph-based DAIAlgs.
o Created TProb class, which stores a probability vector (without the accompanying indexing
and VarSet) and provides functionality for it (which is used by TFactor).
o Rewrote the VarSet class. It now caches its statespace(). It now privately inherits from set``.
I had to overload the insert methods of set`` so that they calculate the new statespace.
o Rewrote the TFactor class. The TFactor class now HAS a TProb and HAS a VarSet.
- Rewrote BP to use the new TProb interface. Performance of BP improved up to a factor 6 by:
o using Prob's instead of Factor's;
o splitting the multiplication of the messages into two parts (thanks to Vicenc!);
o optimizing the calculation of the beliefs (only the message calculations were optimized till now).
o replacing FactorGraph::_nb1 and _nb2 (which were set) by vector
- LCBP now seperately stores cavitydists and pancakes. Added InitPancakes() method
that takes the cavitydists and multiplies them with the relevant factors. This
resulted in an API change in AI which now accepts and returns initial cavitydists
instead of initial pancakes.
Minor changes
- Started writing DoxyGen documentation
- Renamed lcptab2fg.m matlab/ai_writefg.m
- Moved all matlab stuff to matlab/
- More detailed reporting (also reports clocks used).
- Marginal and Marginal2ndO now use *differences* in logZ to avoid NaNs.
- Improved testing suite.
- Removed logreal support.
- FactorGraph now also supports input streams and ignores comment lines in .fg files.
- Added tests/create_full_fg.cpp and tests/create_ising_fg.cpp which create
full and periodic 2D Ising networks according to some command line parameters.
- Now logZ really returns logZ instead of -logZ.
- Added FactorGraph::WriteToDotFile
0.1.4 (2006-04-13)
------------------
- Added file IO routines to read and write factorgraphs.
- Added L-infinity normalization for use with (partially) negative factors.
- Renamed BetheF, MFF to logZ, which now use complex numbers to be able to
handle (partially) negative factors.
- Added test suite.
- All probabilities are now represented using double instead of LogReal.
- Combined Alg and Alg3 into LCBP. Several update schemes possible.
- Combined several variants of BP into doBP. Several update schemes possible.
Now uses precalculated indices for optimization.
- Renamed Node -> Var and Potential -> Factor.
- Extensive code cleanup. More use of OO techniques. Changed API.
- MaxIter now means maximum number of passes (corresponding to number of
_parallel_ updates).
- MaxDiff now means the maximum L-infinity distance between the updated and
original single variable beliefs, for all AI methods.
- Implemented RBP which converges faster than BP for difficult problems.
- Now uses method parameter which is a bitmask combining outmethod and inmethod
(see ai.h).
0.1.3 (2006-03-22)
--------------------
- All AI methods now return maxdiff
- ai.cpp:
o Now returns maxdiffinner and maxdiffouter
o New BP2ndO innermethod (estimate only 2nd order cavity interactions)
o New InitCav outermethod (only create initial cavity distributions)
- bp.cpp:
o New CavityDist2ndO which estimates 2nd order cavity interactions
- Makefile:
o Bugfix: removed dependencies on algwim.*
0.1.2 (2006-02-28)
--------------------
- Cleaned up alg.cpp (removed Alg2 and its corresponding data structures).
- Added the possibility to provide initial cavity distributions as an input
argument to ai (not much error checking is done, so be careful).
- Potentials2mx now correctly sets the dimensions of the P field (i.e. for
the output arguments Q, Q0 of ai).
- Removed algwim.* since it does not work.
0.1.1 (2006-02-28)
--------------------
- The constructors of (Log)FactorGraph and LogFactorGraph from a
vector<(Log)Potential> now merge potentials to prevent short loops (of length
4) in the factor graph. These are used in ai to construct the factor graphs
from the psi argument. If compiled with DEBUG defined, the method calc_nb()
of BipGraph checks for the existence of short loops.
- Changed calling syntax of ai (now the actual syntax *does* correspond to its
description in the help).
- ai does not hook cout anymore (which caused weird segfaults).
- Fixed a bug in an assert statement in the matlab interface code in ai.cpp.
- Removed network.* since it is not useful.
0.1.0 (2006-02-28)
--------------------
First version worthy a version number.
`