LibSVMsharp

LibSVMsharp is a simple and easy-to-use C# wrapper for Support Vector Machines. This library uses LibSVM version 3.37 with x64 and arm64 support.

For more information visit the official libsvm webpage.

What's Changed in This Fork

The items below are the focus of recent updates over the upstream 3.23 baseline.

• Native libsvm upgraded 3.23 → 3.37. Core/svm_model.cs now includes the prob_density_marks field so the managed struct matches the 3.37 memory layout. Without it sv_indices / label / nSV were misaligned by 8 bytes and one-class / probability models read garbage. Core/libsvm.cs VERSION bumped to 3.37.
• Multi-target netstandard2.1;net8.0. The library is now consumable from .NET Core 3.1 / .NET 5/6/7/8 (and Mono/Xamarin via netstandard2.1), not just net8. The smart NativeLibrary resolver is compiled in only on the net8.0 target (#if NET6_0_OR_GREATER); the netstandard2.1 target falls back to the runtime's default libsvm resolution.
• build-native.sh rewritten. Builds both linux-x64 and linux-arm64 (arm64 cross-compile on x64 needs gcc-aarch64-linux-gnu).
• One-class SVM validated. Added LibSVMSharp.Tests/TestSVMOneClass.cs — 4 tests covering train / predict (+1 inlier / −1 outlier), decision values, save-load roundtrip, and an ABI regression guard on SVIndices. Full suite: 35/35 green.
• Native binaries no longer in source control. runtimes/ is gitignored; run build-native.sh locally to populate it.

How to Install

To install LibSVMsharp, download the Nuget package or run the following command in the Package Manager Console:

PM> Install-Package LibSVMsharp

Platform Support

LibSVMsharp multi-targets netstandard2.1 and net8.0, so it can be referenced from .NET Core 3.1 / .NET 5/6/7/8 (and Mono/Xamarin via netstandard2.1). It runs on Windows, Linux, and macOS on both x64 and arm64. The managed wrapper is platform-agnostic; it only needs the native libsvm shared library to be present at run time.

OS	Native file	Architectures	Source
Windows	libsvm.dll	x64 (arm64: build)	Included in the repo / NuGet package
Linux	libsvm.so	x64, arm64	Build from source (see below)
macOS	libsvm.dylib	x64, arm64 (Apple S)	Build from source

On the net8.0 target the library registers a custom NativeLibrary resolver that derives the current runtime identifier (e.g. linux-arm64, win-x64, osx-arm64) from the OS and the process architecture, then looks for the native file under runtimes/<rid>/native/ (and next to the assembly as a fallback). No LD_LIBRARY_PATH tweaking is required, and x64/arm64 libraries can coexist side by side — the matching one is loaded automatically. On the netstandard2.1 target (NativeLibrary is .NET 5+) the runtime's default resolution is used, so place libsvm.so / libsvm.dll next to the host application or on the system library path.

Building on Linux

The native libsvm shared library is not shipped for Linux. Build it with the provided helper script (requires wget or curl, tar, make, and a C/C++ compiler such as gcc/g++); it downloads libsvm-3.37.tar.gz from the official mirror:

./build-native.sh            # build for the host architecture
./build-native.sh x64        # build x86-64 explicitly
./build-native.sh arm64      # build aarch64 (native on arm64, or cross-compiled on x64)
dotnet build LibSVMsharp.sln

Run it once per architecture to populate both runtimes/linux-x64/native/ and runtimes/linux-arm64/native/. For arm64 cross-compilation on an x86-64 host, install the aarch64 toolchain first (sudo apt install gcc-aarch64-linux-gnu g++-aarch64-linux-gnu).

Then run an example:

cd LibSVMsharp.Examples.Classification
dotnet run

License

LibSVMsharp is released under the MIT License and libsvm is released under the modified BSD Lisence which is compatible with many free software licenses such as GPL.

Example Codes

Simple Classification

SVMProblem problem = SVMProblemHelper.Load(@"dataset_path.txt");
SVMProblem testProblem = SVMProblemHelper.Load(@"test_dataset_path.txt");

SVMParameter parameter = new SVMParameter();
parameter.Type = SVMType.C_SVC;
parameter.Kernel = SVMKernelType.RBF;
parameter.C = 1;
parameter.Gamma = 1;

SVMModel model = SVM.Train(problem, parameter);

double[] target = new double[testProblem.Length];
for (int i = 0; i < testProblem.Length; i++)
  target[i] = SVM.Predict(model, testProblem.X[i]);

double accuracy = SVMHelper.EvaluateClassificationProblem(testProblem, target);

Simple Classification with Extension Methods

SVMProblem problem = SVMProblemHelper.Load(@"dataset_path.txt");
SVMProblem testProblem = SVMProblemHelper.Load(@"test_dataset_path.txt");

SVMParameter parameter = new SVMParameter();

SVMModel model = problem.Train(parameter);

double[] target = testProblem.Predict(model);
double accuracy = testProblem.EvaluateClassificationProblem(target);

Simple Regression

SVMProblem problem = SVMProblemHelper.Load(@"dataset_path.txt");
SVMProblem testProblem = SVMProblemHelper.Load(@"test_dataset_path.txt");

SVMParameter parameter = new SVMParameter();

SVMModel model = problem.Train(parameter);

double[] target = testProblem.Predict(model);
double correlationCoeff;
double meanSquaredErr = testProblem.EvaluateRegressionProblem(target, out correlationCoeff);

One-Class SVM (Anomaly Detection)

SVMProblem problem = SVMProblemHelper.Load(@"normal_data.txt");

SVMParameter parameter = new SVMParameter();
parameter.Type = SVMType.ONE_CLASS;   // distribution estimation
parameter.Kernel = SVMKernelType.RBF;
parameter.Nu = 0.1;                   // upper bound on training outliers

SVMModel model = SVM.Train(problem, parameter);

// Predict returns +1 for inliers (normal) and -1 for outliers (anomalies).
double label = SVM.Predict(model, new[]
{
    new SVMNode(1, 0.0),
    new SVMNode(2, 0.0)
});