Easy Factor Graph, aka EFG, is a general purpose c++ library for handling undirected graphical models, sometimes called also factor graphs. Undirected graphical models are probabilistic models similar to bayesian networks, but offerring some nicer properties. Not familiar with this kind of concepts? Don't worry, have a look at the documentation in the doc folder before diving into the code ;). Random Fields as well as Conditional Random Fields are particular classes of undirected graphical models and can be easily created and trained using this library.
Rememebr to leave a star if you have found this repository useful.
Training can be done using the gradient-base approaches implemented of this external library.
In particular, EFG is able to:
- dynamically build and update undirected factor graph, inserting one by one the factors that compose the model
- dynamically set the group of evidences
- perform belief propagation on both loopy graph and polytree like structure in order to
- get the marginal conditioned distribution of an hidden variable w.r.t. the current evidence set
- get the maximum a posteriori of an hidden variable (or for the entire hidden set in one single call) w.r.t. the current evidence set
- import or export models from and to xml file
- import or export models from and to json string (or file)
- draw samples for the variables composing the model
- train random and conditional random fields with gradient based approaches (gradient descend, conjugate gradient descend, quasi newton method, etc.)
With respect to similar libraries, EFG is also able to:
- enforce the fact that group of tunable factors should have the same weight
- exploits an internal thread pool in order to dramatically reduce the time required to:
- perform belief propagation
- train random and conditional random fields
- draw samples for the variables involved in the model
- python bindings of this library are offered by this repo, which is actually a python package that can be pip installed.
Haven't yet left a star? Do it now! :).
This project is structured as follows:
- the documentation in ./doc explains both how to use EFG as well give some theoretical background about undirected graphical models
- the sources of the EFG library are contained in ./src
- ./samples contains 8 classes of examples, extensively showing how to use EFG
Haven't yet left a star? Do it now! :).
The samples contained in the samples folder and extensively shows how to use EFG. All of the samples consume a library of utilities called Samples-Helpers, which contain common functionalities like printing utilities, that are not part (and don't need to be) of EFG.
Wait a minute ... Easy Factor Graph is great, but I don't know C++ and I am more used to python ... well this package is a pip installable wrapper of EFG created with pybind. Combine the power of EFG and python in your next project!
Haven't yet left a star? Do it now! :).
To consume this library you can rely on CMake. More precisely, You can fetch this package and link to the EFG library:
include(FetchContent)
FetchContent_Declare(
efg
GIT_REPOSITORY https://github.com/andreacasalino/Easy-Factor-Graph
GIT_TAG master
)
FetchContent_MakeAvailable(efg)and then link to the EFG library:
target_link_libraries(${THE NAME OF THE TARGET NEEDING EFG}
EFG-Core
)The possibility to train a model is enabled by deafult. However, such functionality rely on this external library, which might slow down the time required to set up the cmake project. Therefore, if you don't need that you can set the CMake option BUILD_EFG_TRAINER_TOOLS equal to OFF. However, after disabling that option you will still able to get the tunable weights of a model, as well as their gradient, allowing you to use or implement another gradient based trainer.
The external package for performing training uses Eigen as internal linear algebra engine. Eigen is by default fetched from the official gitlab repo by CMake and made available. However, if you already have installed Eigen on your machine you can also decide to use that local version, by setting the CMake option EIGEN_INSTALL_FOLDER equal to the root folder storing the local Eigen you want to use.
By default, the features to export and import models from XML files are enabled. If you don't need them, put the CMake option BUILD_EFG_XML_CONVERTER to OFF.
By default, the features to export and import models from JSON files are enabled. They rely on the famous nlohmann library, which is internally fetched and linked. If you don't need such functionalities, put the CMake option BUILD_EFG_JSON_CONVERTER to OFF.
This library exploits virtual inheritance to define some objects hierarchies. This might trigger this weird warning when compiling in Windows with Visual Studio. You can simply ignore it or tell Visual Studio to ignore warning code 4250, which is something that can be done as explained here.
Haven't yet left a star? Do it now! :).
For convenience, assume all these namespaces are used by default:
using namespace EFG;
using namespace EFG::categoric;
using namespace EFG::factor;
using namespace EFG::model;
using namespace EFG::io;
using namespace EFG::train;
using namespace EFG::strct;EFG allows you to define categoric variables and factors by calling simple functions. This is what you would do to build a couple of variables:
// define a couple of variables, with the same size
VariablePtr A = make_variable(3, "A"); // size is 3
VariablePtr B = make_variable(3, "B"); // size is 3Such variables can be referred by the factors correlating them. In order to build a simple correlating factor this is what you would do:
// build a simply correlating factor connecting the 2 variables
Factor factor_AB(VariablesSoup{B, A}, // the order in the specified
// group matters: B is assumed
// as the first variable, A
// will be the second
Factor::SimplyCorrelatedTag{});And this is what you would do to generate an exponential simple correlating factor:
// build an exponential factor using as base `factor_AB`: values of the
// images are assumed as exp(weight * images_factor_AB)
FactorExponential factor_AB_exponential(
factor_AB,
1.5f // this will be the value assumed for the weight
);You can also define custom factors, specifying the shape function that maps the values in their domain with their images. For example:
// define another variable
VariablePtr C = make_variable(2, "C"); // size is 2
// define a factor connecting C to B
// we start building an empty factor, having all images equal to 0
Factor factor_BC(VariablesSoup{B, C});
// set some individual images of factor_BC
// set for <0,1> -> 2
factor_BC.set(std::vector<std::size_t>{0, 1}, 2.f);
// set for <2,0> -> 1.3f
factor_BC.set(std::vector<std::size_t>{2, 0}, 1.3f);Factor graphs can be built incrementally, passing one by one the factors that compose them. Without loss of generality suppose to start from an empty random field:
// start building an empty random field
RandomField model;then, you can build some factors and enrich the model with them:
// define some variables, which will be later connected
auto A = make_variable(4, "varA");
auto B = make_variable(4, "varB");
auto C = make_variable(4, "varC");
// without loss of generality, add to the model some simply correlating
// factors
model.addConstFactor(std::make_shared<Factor>(
VariablesSoup{A, B},
Factor::SimplyCorrelatedTag{})); // the generated smart
// pointer is shallow
// copied
model.copyConstFactor(
Factor{VariablesSoup{A, C},
Factor::SimplyCorrelatedTag{}}); // the passed factor is
// deep-copied into the
// modelThe previously added factor are kept constant in the model. In order to enrich the model with a tunable factor you can call a different method:
// build some additional tunable exponential factors that will be too added
auto factor_exp_BC = std::make_shared<FactorExponential>(
Factor{VariablesSoup{B, C}, Factor::SimplyCorrelatedTag{}}, 1.f);
model.addTunableFactor(factor_exp_BC);
auto D = make_variable(4, "varD");
auto factor_exp_CD = std::make_shared<FactorExponential>(
Factor{VariablesSoup{C, D}, Factor::SimplyCorrelatedTag{}}, 1.5f);
model.addTunableFactor(factor_exp_CD);You can also add a tunable factor, that must share its weigth with an already inserted factor of the model:
// insert another tunable factor, this time specifying that it needs to
// share the weight with already inserted exponential factor that connects B
// and C
model.addTunableFactor(
std::make_shared<FactorExponential>(
Factor{VariablesSoup{C, D}, Factor::SimplyCorrelatedTag{}},
2.f // actually this value is irrelevant, as the weight of
// factor_exp_BC will be assumed from now on
),
VariablesSet{B, C}
// this additional input is to specify that this exponential factor
// needs to share the weight with the one connecting B and C
);You can also import the entire graph defined in an xml file:
// absorb the structure defined in an xml file
xml::Importer::importFromFile(model, std::string{"file_name.xml"});check the documentation or the samples for the expected format the xml file should be compliant with.
Similarly, you can also import the structure defined in a json
// absorb the structure encoded in a json string
nlohmann::json json_defining_a_structure = ...;
json::Importer::importFromJson(model, json_defining_a_structure);check the documentation or the samples for the expected format the json should be compliant with.
A generated model can be queried in many ways. However, any query that you can do, is conditioned to the latest set of evidences.
Setting the evidences can be easily done by calling:
// set some evidences
model.setEvidence("variable_1", 0); // setting variable_1 = 0
model.setEvidence("variable_2", 2); // setting variable_2 = 2You can get the conditioned marginal distribution of a variable by calling:
// get the marginal conditioned distribution of an hidden variable
std::vector<float> conditioned_marginals =
model.getMarginalDistribution("var_A");Or you might be interested in the maximum a posteriori estimation of the entire evidence set:
// get maxiomum a posteriori estimation of the entire hidden set
std::vector<std::size_t> MAP_hidden_set = model.getHiddenSetMAP();As already mentioned, results are subjected to the latest evidences set (which can be also empty). Of course, you can update the evidences and get the updated marginals:
// set some new evidences
model.removeAllEvidences();
model.setEvidence("evid_1", 1);
// compute new conditioned marginals: the should be different as the
// evidences were changed
conditioned_marginals = model.getMarginalDistribution("var_A");Tunable models are characterized by the exponential factors added to the model itself. Such kind of modelscan be trained. This is done by relying on a training set, which can be for example imported from a file:
// assume we have a training set for the model stored in a file
TrainSet training_set = import_train_set("file_name.txt");Then, a training approach must be chosen. You can rely on one of the ready to use approaches implemented in this (by default) fetched package. Suppose you want to use a quasi Newton method:
// we can train the model using one of the ready to use gradient based
// approaches
::train::QuasiNewton ready_to_use_trainer;
ready_to_use_trainer.setMaxIterations(50);Then, you are ready to train the model:
// some definitions to control the training process
TrainInfo info = TrainInfo{
4, // threads to use
1.f // stochasticity. When set different from 1, the stochastich
// gradient descend approaches are actually used
};
train_model(tunable_model, ready_to_use_trainer, training_set, info);Sometimes, it might be useful to draw samples from the model. This can be done with the Gibbs sampling strategy provided by EFG:
// some definitions to control the samples generation process
GibbsSampler::SamplesGenerationContext info =
GibbsSampler::SamplesGenerationContext{
1000, // samples number
0, // seed used by random engines
500 // number of iterations to discard at the beginning (burn out)
};
// get samples from the model using Gibbs sampler
std::vector<std::vector<std::size_t>> samples =
model.makeSamples(info,
4 // threads to use
);If you have found this library useful, please find the time to leave a star :). Just before you go, be aware that Easy-Factor-Graph-GUI wraps this library as C++ backend to a nice graphical user interactive application:
