Please check the Release Notes file for news on what's been updated in each new release.
TorchSharp is now in the .NET Foundation!
If you are using TorchSharp from NuGet, you should be using a version >= 0.98.3 of TorchSharp, and >= 1.12.0 of the libtorch-xxx redistributable packages. We recommend using one of the 'bundled' packages: TorchSharp-cpu, TorchSharp-cuda-windows, or TorchSharp-cuda-linux. They will pull in the right LibTorch backends.
TorchSharp examples has their own home!
Head over to the TorchSharp Examples Repo for convenient access to existing and upcoming examples.
IMPORTANT NOTES:
When targeting .NET FX on Windows, the project configuration must be set to 'x64' rather than 'Any CPU' for anything that depends on TorchSharp.
As we build up to a v1.0 release, we will continue to make breaking changes, but only when we consider it necessary for usability. Similarity to the PyTorch experience is a primary design tenet, and we will continue on that path.
TorchSharp is a .NET library that provides access to the library that powers PyTorch. It is part of the .NET Foundation.
The focus is to bind the API surfaced by LibTorch with a particular focus on tensors. The design intent is to stay as close as possible to the Pytorch experience, while still taking advantage of the benefits of the .NET static type system where it makes sense. For example: method overloading is relied on when Pytorch defines multiple valid types for a particular parameter.
The technology is a "wrapper library": no more, no less. DiffSharp uses this repository extensively and has been a major factor in iterating support.
Things that you can try:
using TorchSharp;
using static TorchSharp.torch.nn;
var lin1 = Linear(1000, 100);
var lin2 = Linear(100, 10);
var seq = Sequential(("lin1", lin1), ("relu1", ReLU()), ("drop1", Dropout(0.1)), ("lin2", lin2));
using var x = torch.randn(64, 1000);
using var y = torch.randn(64, 10);
var optimizer = torch.optim.Adam(seq.parameters());
for (int i = 0; i < 10; i++) {
using var eval = seq.forward(x);
using var output = functional.mse_loss(eval, y, Reduction.Sum);
optimizer.zero_grad();
output.backward();
optimizer.step();
}While the intent has been to stay close to the Pytorch experience, there are some peculiarities to take note of:
-
We have disregarded .NET naming conventions in favor of Python where it impacts the experience. We know this will feel wrong to some, but after a lot of deliberation, we decided to follow the lead of the SciSharp community and embrace naming similarity with Python over .NET tradition. We believe this will make it easier to take Python-based examples and snippets and apply them in .NET.
-
In order to make a constructor call look more the Pytorch code, each class has a factory method with the same name. Because we cannot have a method and a class with the same name in a scope, we moved the class declarations to a nested scope 'Modules.'
For example:
Module conv1 = Conv1d(...);
creates an instance of
Modules.Conv1d, which has 'torch.Module' as its base class. -
C# uses ':' when passing a named parameter, while F# and Python uses '=', and Pytorch functions have enough parameters to encourage passing them by name. This means that you cannot simply copy a lot of code into C#.
-
There are a number of APIs where Pytorch encodes what are effectively enum types as strings. We have chosen to use proper .NET enumeration types in most cases.
-
The type
torch.deviceistorch.Devicein TorchSharp. We felt that using all-lowercase for a class type was one step too far. The device object constructors, which is what you use most of the time, are still calleddevice()
TorchSharp is distributed via the NuGet gallery: https://www.nuget.org/packages/TorchSharp/
We recommend using one of the 'bundled' packages, which will pull in both TorchSharp and the right backends:
- TorchSharp-cpu (CPU, Linux/Windows/OSX)
- TorchSharp-cuda-windows (CPU/CUDA 12.1, Windows)
- TorchSharp-cuda-linux (CPU/CUDA 12.1, Linux)
Otherwise, you also need one of the LibTorch backend packages: https://www.nuget.org/packages?q=libtorch, specifically one of
-
libtorch-cpu-linux-x64(CPU, Linux) -
libtorch-cpu-win-x64(CPU, Windows) -
libtorch-cpu-osx-x64(CPU, OSX) -
libtorch-cpu(CPU, references all three, larger download but simpler) -
libtorch-cuda-12.1-linux-x64(CPU/CUDA 12.1, Linux)NOTE: Due to the presence of very large native binaries, using the
libtorch-cuda-12.1-linux-x64package requires .NET 6, e.g. .NET SDK version6.0.100-preview.5.21302.13or greater. -
libtorch-cuda-12.1-win-x64(CPU/CUDA 12.1, Windows)
Alternatively you can access the LibTorch native binaries via direct reference to existing local native binaries of LibTorch installed through other means (for example, by installing PyTorch using a Python package manager). You will have to add an explicit load of the relevant native library, for example:
using System.Runtime.InteropServices;
NativeLibrary.Load("/home/gunes/anaconda3/lib/python3.8/site-packages/torch/lib/libtorch.so")NOTE: Some have reported that in order to use TorchSharp on Windows, the C++ redistributable needs to be installed. This will be the case where VS is installed, but it maybe necessary to install this version of the C++ redist on machines where TorchSharp is deployed:
Microsoft Visual C++ 2015-2022 ( 14.36.32532 )
This project has adopted the code of conduct defined by the Contributor Covenant to clarify expected behavior in our community. For more information see the .NET Foundation Code of Conduct.
See DEVGUIDE.md and CONTRIBUTING.md.
DiffSharp also uses this repository extensively and has been a major factor in iterating support.
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent