DiffSharp / DiffSharp

Documentation

This is the development branch of DiffSharp 1.0.

NOTE: This branch is undergoing development. It has incomplete code, functionality, and design that are likely to change without notice.

Getting Started

DiffSharp is normally used from an F# Jupyter notebook. You can simply open examples directly in the browser, e.g.

• index.ipynb

• getting-started-libtorch.ipynb

To use locally in Visual Studio Code:

• Install .NET Interactive Notebooks for VS Code

• After opening an .ipynb execute Ctrl-Shift-P for the command palette and chose Reopen Editoer With... then .NET Interactive Notebooks

• To restart the kernel use restart from the command palette.

To use locally in Jupyter, first install Jupyter and then:

dotnet tool install -g --add-source "https://dotnet.myget.org/F/dotnet-try/api/v3/index.json" microsoft.dotnet-interactive
dotnet interactive jupyter install

When using .NET Interactive it is best to completely turn off automatic HTML displays of outputs:

Formatter.SetPreferredMimeTypeFor(typeof<obj>, "text/plain")
Formatter.Register(fun x writer -> fprintfn writer "%120A" x )

You can also use DiffSharp from a script or an application. Here are some example scripts with appropriate package references:

• docs/index.fsx

• docs/getting-started-libtorch.fsx

Available packages and backends

Now reference an appropriate nuget package from https://nuget.org:

• DiffSharp-lite - This is the reference backend.

• DiffSharp-cpu - This includes the Torch backend using CPU only.

• DiffSharp-cuda-linux, DiffSharp-cuda-windows - These include the Torch CPU/GPU backend for Linux and Windows respectively. Large download.

For all but DiffSharp-lite add the following to your code:

dsharp.config(backend=Backend.Torch)

Using a pre-installed or self-built LibTorch 1.8.0

The Torch CPU and CUDA packages above are large. If you already have libtorch 1.8.0 available on your machine you can

1. reference DiffSharp-lite

2. set LD_LIBRARY_PATH to include a directory containing the relevant torch.so, torch_cpu.so and torch_cuda.so, or execute NativeLibrary.Load on torch.so.

3. use dsharp.config(backend=Backend.Torch)

Developing DiffSharp Libraries

To develop libraries built on DiffSharp, do the following:

1. reference DiffSharp.Core and DiffSharp.Data in your library code.

2. reference DiffSharp.Backends.Reference in your correctness testing code.

3. reference DiffSharp.Backends.Torch and libtorch-cpu in your CPU testing code.

4. reference DiffSharp.Backends.Torch and libtorch-cuda-linux or libtorch-cuda-windows in your (optional) GPU testing code.