System.Numerics.Tensors

About

Provides methods for performing mathematical operations over tensors represented as spans. These methods are accelerated to use SIMD (Single instruction, multiple data) operations supported by the CPU where available.

Key Features

• Numerical operations on tensors represented as ReadOnlySpan<float>
• Element-wise arithmetic: Add, Subtract, Multiply, Divide, Exp, Log, Cosh, Tanh, etc.
• Tensor arithmetic: CosineSimilarity, Distance, Dot, Normalize, Softmax, Sigmoid, etc.

How to Use

using System.Numerics.Tensors;

var movies = new[] {
    new { Title="The Lion King", Embedding= new [] { 0.10022575f, -0.23998135f } },
    new { Title="Inception", Embedding= new [] { 0.10327095f, 0.2563685f } },
    new { Title="Toy Story", Embedding= new [] { 0.095857024f, -0.201278f } },
    new { Title="Pulp Function", Embedding= new [] { 0.106827796f, 0.21676421f } },
    new { Title="Shrek", Embedding= new [] { 0.09568083f, -0.21177962f } }
};
var queryEmbedding = new[] { 0.12217915f, -0.034832448f };

var top3MoviesTensorPrimitives =
    movies
        .Select(movie =>
            (
                movie.Title,
                Similarity: TensorPrimitives.CosineSimilarity(queryEmbedding, movie.Embedding)
            ))
        .OrderByDescending(movies => movies.Similarity)
        .Take(3);

foreach (var movie in top3MoviesTensorPrimitives)
{
    Console.WriteLine(movie);
}

Main Types

The main types provided by this library are:

• System.Numerics.Tensors.TensorPrimitives

Additional Documentation

• API documentation

Feedback & Contributing

System.Numerics.Tensors is released as open source under the MIT license. Bug reports and contributions are welcome at the GitHub repository.