Projects

nll

Numerical learning libraryNLL is a thread-safe multi-platform open source project entirely written in C++. Its goal it to propose generic and efficient algorithms for machine learning and more specifically computer vision. It is intended to be very easy to integrate and it is mainly composed of ... [More] header files with no dependency on any library but the STL. ArchitectureNLL implements generic algorithms using template metaprogramming and a minimalist interface. Several layers are used: core: the very basic structures and operations algorithm_impl: generic algorithms with very limited dependencies and interface algorithm: algorithms taking advantage of the NLL framework preprocessing: it is built around the concept of typelist and propose a unified pipeline to process and optimize features. The processing of the data must comply with a specific set of transformations to eventually feed a classifier's input. DetailsHere is an overview of some algorithms implemented in NLL: classifiers (k-nearest neighbour, multi-layered neural networks, support vector machines, boosting, gaussian mixture model) feature selection (best-first, wrapper using genetic algorithm, relief-f, pearson) feature transformation (PCA) optimizers (grid search, harmony search, genetic algorithms) math library (matrix, vector, linear algebra, distributions) image library (resampling, morphology, transformations, convolutions, region growing, labelling) clustering (k-means, LSDBC) kd-trees, gabor filters... and much more soon! ExampleHere is a typical use of the framework: /** In this test a neural network will be optimized using a harmony search algorithm. */ void test() { typedef nll::benchmark::BenchmarkDatabases::Database::Sample::Input Input; typedef nll::algorithm::Classifier Classifier; // find the cancer1.dt benchmark const nll::benchmark::BenchmarkDatabases::Benchmark* benchmark = nll::benchmark::BenchmarkDatabases::instance().find( "cancer1.dt" ); ensure( benchmark, "can't find benchmark" ); Classifier::Database dat = benchmark->database; // use a multi layered perceptron as a classifier typedef algorithm::ClassifierMlp ClassifierImpl; ClassifierImpl classifier; // optimize the parameters of the classifier on the original dataset // we will use a harmony search algorithm. // For each point, the classifier is evaluated: a 10-fold cross validation is // run on the learning database Classifier::OptimizerClientClassifier classifierOptimizer = classifier.createOptimizer( dat ); // configure the optimizer options nll::algorithm::StopConditionIteration stop( 10 ); nll::algorithm::MetricEuclidian metric; nll::algorithm::OptimizerHarmonySearchMemory parametersOptimizer( 5, 0.8, 0.1, 1, &stop, 0.01, &metric ); // run the optimizer on the default constrained classifier parameters // if the default values don't fit, other constraint parameters should be given std::vector params = parametersOptimizer.optimize( classifierOptimizer, ClassifierImpl::buildParameters() ); // learn the LEARNING and VALIDATION database with the optimized parameters, and test the classifier // on the TESTING database classifier.learnTrainingDatabase( dat, nll::core::make_buffer1D( params ) ); Classifier::Result rr = classifier.test( dat ); TESTER_ASSERT( rr.testingError < 0.025 ); } TutorialsTutorials are included in the tutorial sub-project. [Less]

mdsst

A method for discovering Metadata in semi-structured text. This is a continuation of a previous method which is used to extracta data, metadata and attributes in semi-structured text with HMMs.