General

• Fixing target framework versions: projects that were targetting 4.5.2 and 4.6.1 have been updated to target 4.5 and 4.6, respectively;
• Adding a Accord.DataSets namespace to contain classes that can download and pre-process well-known machine learning datasets directly from the web;
• Adding a Accord.Text namespace to contain classes related to text processing;
• Creating a separate Accord.Video.FFMPEG NuGet package to target x64 (Win64);
• Upgrading C++ projects to VS2017 and removing dependencies on VS2013 and VS2015;

Core

• Updating the framework random generator to allow fixing the seed of independent threads: new random generators created from existing threads will be initialized with the global random seed, except if their thread-specific random seed has been manually overwritten;
• Adding support for using GZip compression when serializing through Accord.Serializer;
• Adding initial support for rational numbers;

IO

• Adding support for reading and writing NumPy's .npy and .npz formats;
• Adding a SparseWriter to save Sparse in LibSVM sparse file format;

Video

• Updating FFMPEG library to version 3.2.2;
• Updating Accord.Video.FFMPEG to support both x86 and x64 platforms;
• Updating FrameRate properties of VideoFileWriter and VideoFileReader to be represented in Rationals instead of Int32;

DataSets

• Adding downloaders for the RCV1v2, MNIST and News20 datasets;

Text

• Adding C# versions of Snowball's stemmers;

Math

• Fixing ToString() method of Sparse vectors such that they are compatible with LibSVM;
• Adding Angular distance metric (proper distance metric based on the Cosine similarity);
• Updating distance metrics to be structs rather than classes, allowing for optimizations when used in conjunction with generic classes and methods;
• Adding generic versions of the IOptimizationMethod interfaces, and adding a specific IFunctionOptimizationMethod interface to specify function optimization methods that do not necessarily require a gradient function;

Statistics

• Fixing serialization of PCA, KPCA, LDA, KDA and other classes containing a CancellationToken;
• Adding support for learning multinomial logistic regression with any optimization algorithm;
• Correcting/improving Wilcoxon distribution: Adding a separate method for the ComplementaryDistributionFunction for more precision; adding the approximation distribution using an actual NormalDistribution, and new parameters at its constructor to control whether exact or approximate distribution should be used; adding parameters to determine whether to apply different types of continuity corrections; adding an option to determine whether to correct for ties when computing the rankings;
• Correcting/improving Mann-Whitney U's distribution: Adding most corrections as mentioned above, plus checks to ensure we are computing the distribution for the case when the first sample is smaller than the second; adding a correction to the variance of the approximation distribution for the case of ties; adding parallelization support when computing the table of exact values;
• Correcting/improving Wilcoxon and Mann-Whitney tests: adding a correction for ties in the variance of the Normal approximation; adding checks to make sure we don't generate p-values higher than 1; adding more comparison tests against results generated by R;
• Wavelet kernel is now a struct and has support for Int32 inputs;

MachineLearning

• Improving the performance for LinearDualCoordinateDescent and OneVsRest SVMs;
• Adding a ToMulticlass method to Multilabel classifiers based on scores;
• Updating K-Nearest Neighbor to use the new classification/learning interfaces;
• Adding parallelism to the non-data-structure-based version of K-Nearest Neighbors;

Accord.Imaging

• Adding support for LocalBinaryPattern and Haralick in Bag-of-Visual-Words.