avensun / torchsharp

.NET bindings for the Pytorch engine

TorchSharp

TorchSharp is a .NET library that provides access to the library that powers PyTorch. It is a work in progress, but already provides a .NET API that can be used to perform (1) various operations on ATen Tensors; (2) scoring of TorchScript models; (3) Training of simple neural networks.

Our current focus is to bind the entire API surfaced by libtorch.

Things that you can try:

using AtenSharp;

var x = new FloatTensor (100);   // 1D-tensor with 100 elements
FloatTensor result = new FloatTensor (100);

FloatTensor.Add (x, 23, result);

Console.WriteLine (x [12]);

Discussions

We have a chat room on Gitter

Building

Windows

Requirements:

• Visual Studio
• git
• cmake (tested with 3.14)

Commands:

• Building: build.cmd
• Building from Visual Studio: first build using the command line
• See all configurations: build.cmd -?
• Run tests from command line: build.cmd -runtests
• Build packages: build.cmd -buildpackages

Linux/Mac

Requirements:

• requirements to run .NET Core 2.0
• git
• cmake (tested with 3.14)
• clang 3.9

Example to fulfill the requirements in Ubuntu 16:

sudo apt-get update
sudo apt-get install git clang cmake libunwind8 curl
sudo apt-get install libssl1.0.0
sudo apt-get install libomp-dev

Commands:

• Building: ./build.sh
• Building from Visual Studio: first build using the command line
• See all configurations: ./build.sh -?
• Run tests from command line: ./build.sh -runtests
• Build packages: ./build.sh -buildpackages

Updating package version for new release

To change the package version update this file. Everything is currently considered in preview.

Use the following two MSBuild arguments in order to control the -preview and the build numbers in the name of the nuget packages produced (use one of the two generally):

Name	Value	Example Version Output
StabilizePackageVersion	true	1.0.0
IncludeBuildNumberInPackageVersion	false	1.0.0-preview

Sample command: ./build.cmd -release -buildpackages -- /p:StabilizePackageVersion=true

GPU support

For GPU support it is required to install CUDA 9.0 and make it available to the dynamic linker.

Examples

Porting of the more famous network architectures to TorchSharp is in progress. For the moment we only support MNIST and AlexNet