« Changelog History


Accord.NET v3.2.0 Release Notes

Release Date: 2016-08-20 // over 7 years ago

  • 🚀 > Build 3.2.0.5706, released on 20.08.2016

    🚀 Accord.NET Framework 3.2.0 release notes

    20.08.2016.

    🚀 Accord.NET 3.2 "auto-generated" release

    🔖 Version updates and fixes:

    • GH-76/GC-24: Add easier creating and handling of factors for categorical variables
    • GH-123: Bug in the Euclidean on Accord.Math.Distance
    • GH-124: Fixing the Envelop filter as missing loop variables were not being incremented
    • GH-135: When the from and to ranges are equal, scaled values should remain unchanged
    • GH-159: Gamma Distribution Fit stalls for some arrays
    • GH-162: ntdll on OS X
    • GH-167: Posterior method has wrong signature in continuous hidden Markov Models
    • GH-171: Quadratic Programming (Goldfarb-Idnani) NoPossibleSolution on possible problems
    • GH-188: ProbabilisticOutputCalibration Class Example Incorrect Object Name
    • GH-206: Chessboard distance is incorrect
    • GH-214: Bug found in ReplaceChannel filter
    • GH-215: Bug Found in DecisionTrees.Learning.ID3Learning.
    • GH-225: Independent Component Analysis not converging
    • GH-232: Bug in Levenshtein distance.
    • GH-234: The subset of observations corresponding to a decision node may contains duplicates
    • GH-235: The getMaxChild method returns the max grandchild
    • GH-236: Possibly-biased comparison between errors
    • GH-237: The subset of observations corresponding to a decision node may contains duplicates
    • GH-240: Re() and Im() function of ComplexMatrix generates a OutOfRangeException
    • General
      • In this release, the Matrix library from Accord.Math has been almost completely
        redesigned to make heavy use of automatic code generation. This results in more
        code reuse, more consistent interfaces and the availability of many methods which
        before were available only for Double to almost all native numerical types in the
        .NET Framework;
      • The framework now contains core classes and interfaces for defining classification
        and regression models and their respective learning algorithms, offering a more
        standard interface when using different parts of the framework;
      • The framework now offers a Accord.Serializer class that should be responsible for
        serializing and deserializing any object from the framework, and will take care of
        🔖 versioning in case of breaking changes between releases;
      • All AForge.NET namespaces have been finally moved to inside Accord.NET, although
        some functionality is still duplicate.
    • Core
      • Adding Interlocked operations (Increment, Add) for double values;
      • To<> universal converter can now convert jagged arrays;
      • Adding a common framework to unify all classification models, and all learning algorithms;
      • Integrating the AForge.NET Range classes in the framework, adding ByteRange;
      • Adding a common serialization mechanism to the framework to manage backwards compatibility;
      • All classes from Accord.MachineLearning.Structures have been moved into Accord.Collections;
      • Updating RedBlackTrees to implement the new base classes for tree structures;
      • Updating KD-Trees to implement the base classes for tree structures (introduces breaking changes).
    • Sample applications
      • Fixing wrong arguments in sample applications.
    • Math
      • Revamped matrix library making heavy use of code generation with T4 templates;
      • Matrix dot products, and elementwise operations are now auto-generated;
      • Renaming InnerProduct to Dot, and marking previous products as obsolete;
      • Vector Range, Scale and Interval are now auto-generated;
      • Standardizing the way Vectors, Matrices and Jagged matrices are created and handled
        in the framework;
      • Adding OneHot and KHot methods overloads for creating vectors using boolean masks;
      • Adding ArgMin and ArgMax methods to Vector, Jagged and Multidimensional matrices;
      • Re-implementing Matrix.Sum and Matrix.Product using T4 templates;
      • Breaking change: Sum() now computes the Sum over the entire matrix (before it needed
        to be done with Sum().Sum(). In order to compute the sum vector over rows, use matrix.Sum(0)
        and for columns, matrix.Sum(1);
      • Chessboard distance has been removed as it is the same as Chebyshev;
      • Moving AForge.NET's old Random classes into the framework, and marking them as deprecated;
      • Adding a log1pexp method for computing (1.0 + Math.Exp(-sum)) without loss of precision;
      • Adding new random generators based on Marsaglia's Ziggurat method;
      • Introducing a new, generic IRandomNumberGenerator interface so existing statistical
        distributions can be used as Random Number Generators;
      • Updating Matrix.IsEqual method to use the auto-generated overloads if possible;
      • Replacing the previous framework-wide generator with a better API;
      • Improving the framework-wide random number generator so generators created in short
        ⚡️ timespans do not get initialized with the same seed: Now, updating a seed will not
        affect existent random generators in other threads. It will affect only newly created
        generators and the one in the current thread;
      • Fixing the DiagonalMatrix property in SingularValueDecomposition and
        JaggedSingularValueDecomposition so the returned diagonal matrices has the necessary
        dimensions to reconstruct the original matrix using the decomposition main formulation;
      • Fixing a bug in Combinatorics.Sequences method where the current vector would be returned
        instead of a copy when inPlace = false;
      • Distance functions can now be auto-generated from classes from the framework;
      • Adding Dice, Jaccard, Kulczynski, Matching, Rogers-Tanimoto, Russel-Rao, Sokal Michener,
        Sokal Sneath, Yule, Bhattacharyya and LogLikelihood distances as proper classes;
      • Updating IsEqual to support absoluete and relative tolarance thresholds;
      • Adding a Histogram method for creating a histogram from an array of integer values;
      • Updating the Interval, Range and Scale method overloads to be automatically generated;
      • Adding loss functions to be used in the unified framework;
      • Moving the Elementwise class to a separate Accord.Math.Core project in order to avoid
        🏗 excessive build times due the number of auto-generated methods in this class;
      • Adding overloads to Eigenvalue decomposition to automatically sorter eigenvectors and
        eigenvalues in descending order of absolute eigenvalue;
      • Adding a dedicated Sort static class with ordering-related methods such as Partition,
        Introsort and NthElement.
      • Expanding decompositions with two additional methods: GetInformationMatrix and Reverse
        GetInformationMatrix can be used to retrieve the standard errors for each coefficient
        when solving a linear system; Reverse reconstructs the original matrix using the definition
        of the decomposition;
      • Deprecating Submatrix in favor of Get (methods with non-inclusive last indices);
      • Adding ArgSort function for retrieving the indices that can be used to sort a vector;
      • Adding LogSumExp to the set of special functions.
    • MachineLearning
      • Adding a base foundation to encompass all classification and regression models in the
        framework as well as their learning algorithms: common interfaces and base classes for
        classifiers, distance-based classifiers and generative classifiers; common interfaces
        and base classes for supervised and unsupervised learning algorithms;
      • Updating Support Vector Machines, Decision Trees, Naive Bayes, Regressions and Analyses
        to use the new classes;
      • Unifying Linear and Kernel SupportVectorMachines, updating their classes to accept the
        Kernel function as a generic parameter: when the kernel function is a ValueType, this
        👮 forces generic classes to be compiled specifically for each kernel type, allowing for
        the inlining of the kernel function calls;
      • Updating the way compact SVMs are represented: instead of having only a weight vector
        👍 and no support vectors, compact machines have a single support vector and a single weight
        of value one, eliminating what before was a special case;
      • Adding classes for OneVsOne and OneVsRest classifiers, separating the functionality that
        👍 was previously inside MulticlassSupportVectorMachine and MultilabelSupportVectorMachine;
      • Fixing multiple issues with ErrorBasedPruning (YaronK);
      • Updating GridSearch to implement ToString methods for easier debugging;
      • Updating Linear machines and learning algorithms to accept sparse kernels;
      • Deprecating the previous sparse vector implementations and moving the current implementation
        to the existing Linear class, since they represent the same operation;
      • Adding a true implementation for LibSVM-style Sparse vectors;
      • Updating SparseReader to read sparse vectors using the new Sparse representation;
      • Refactoring the clustering namespace to increase code reuse between the different algorithms;
      • Updating K-Means, GMM and BagOfWords to expose a ParallelOptions object that can
        🔧 be used to configure and stop the parallelization of those algorithms;
      • Updating K-Means to support sample weights;
      • Correcting multiple random initializations of Gaussian mixture model;
      • Adding a PriorityQueue class based on the MIT-licensed code by Daniel "BlueRaja" Pflughoeft.
      • Adding Vantage-Point and Space-Partitioning trees and Barnes Hutt t-SNE based on the original
        code from Laurens van der Maaten BH t-SNE implementation;
      • Adding a basic implementation for the Apriori algorithm.
    • Imaging
      • Updating static methods in AForge.NET's Image class to become extension methods;
      • Implementing ICloneable in all corner and feature detectors.
    • Neuro
      • Updating ResilientBackpropagation with the improvements from iRProp+.
    • Statistics
      • Adding Non negative Least Squares regression;
      • Adding Procrustes Analysis;
      • Deprecating IAnalysis in favor of the new framelet for classification,
        regression and transformation methods;
      • Merging AForge.NET and Accord.NET Histogram classes;
      • Updating IFittingOptions to implement ICloneable;
      • Adding constructors to Independent distributions accepting a lambda function
        to initialize inner components instead of relying on cloning;
      • Adding a Classes class to provide methods that operate with categorical/label data,
        such as converting boolean, double or integer values to [0;1] or [-1; +1] indicators;
      • Adding Decide methods to unambiguously transform a distance/score value into a boolean;
      • Updating statistic distributions to implement the IRandomNumberGenerator interface, meaning
        any distribution can now be used as random number generator;
      • Adding the Metropolis-Hasting sampler to generate samples from multivariate distributions
        that do not have specialized samplers;
      • Adding named constructors for building regressions directly from coefficient vectors;
      • Updating kernels to rely in Accord.Math.IDistance instead of the previous IDistance from
        the Statistics namespace;
      • Adding Pearson's Universal Kernel, Thin Spline Plate and Hellinger kernels
        contributed by Diego Catalano;
      • Moving standard statistical measures (i.e. mean, standard deviation, variance, ...) to a
        separate Measures class;
      • Updating Mean methods to operate in the same way as Sum: if a dimension is not specified,
        the Mean will be computed across all dimensions of the matrix;
      • Updating Hidden Markov Models to use the new Tagger interfaces and base classes.
    • Genetics
      • Updating the Genetics project to use the new sample generators based on statistical
        distributions;