Hasktorch 0.2 Pre-Release

Hasktorch is a library for tensors and neural networks in Haskell. It is an independent open source community project which leverages the core C++ libraries shared by PyTorch.

This project is in early development and should only be used by contributing developers. Expect substantial changes to the library API as it evolves. Contributions and PRs are welcome.

Currently we are prepping development and migration for a major 2nd release (please excuse sparsity of docs in the meantime). If you're interested in details or contributing please get in touch (see contributing).

Project Structure

Basic functionality:

deps/ - submodules and downloads for build dependencies - libtorch, mklml, pytorch
examples/ - high level example models (xor mlp, typed cnn)
experimental/ - experimental projects or tips (jupyterlab)
hasktorch/ - higher level user-facing library, calls into ffi/, used by examples/

Internals (for contributing developers):

codegen/ - code generation, parses Declarations.yaml spec from pytorch and produces ffi/ contents
inline-c/ - submodule to inline-cpp fork used for C++ FFI
libtorch-ffi/- low level FFI bindings to libtorch
spec/ - specification files used for codegen/

Getting Started

On OSX or Ubuntu-like OSes'

The following steps run a toy linear regression example, assuming the hasktorch repository has just been cloned.

Starting at the top-level directory of the project, go to the deps/ (dependencies) directory and run the get-deps.sh shell script to retrieve project dependencies with the following commands:

pushd deps           # Change to deps directory and save the current directory.
./get-deps.sh        # Run the shell script to retrieve the dependency. 
popd                 # Go back to the root directory of the project.

If you are using CUDA-9, replace ./get-deps.sh with ./get-deps.sh -a cu92. Likewise for CUDA-10, replace ./get-deps.sh with ./get-deps.sh -a cu101.

These downloads include various pytorch shared libraries. Note get-deps.sh only has to be run once when the repo is initially cloned.

Next, set shell environment to reference the shared library locations:

source setenv

Note source setenv should be run from the top-level directory of the repo.

via nix-shell

Always the artifacts of hasktorch's master branch are uploaded to cachix. If you setup cachix before using nix-shell, nix-shell will be faster.

nix-env -i cachix
cachix use hasktorch
nix-shell ./hasktorch/shell.nix

Will get you into a development environment for hasktorch using the CPU backend. On NixOS you may also pass in a cudaVersion argument of 9 or 10 to provision a CUDA environment:

nix-shell ./hasktorch/shell.nix --arg cudaVersion 9 # or 10

If you are running cabal or stack to develop hasktorch, there is a shell hook to tell you which extra-lib-dirs and extra-include-dirs fields to include in your stack.yaml or cabal.project.local. This hook will also explain how to create a cabal.project.freeze file.

Building examples

Finally, try building and running the linear regression example:

stack run regression

For additional examples, see the examples/ directory.

Set up development environement in VS Code.

If you want to develop the project in VS Code and get Haskell Tooling support, you will need to install HIE(Haskell IDE Enginer). Since this project uses the resolver version lts-14.7, so you will need to install and use the corresponding version of HIE which is hie-8.6.5.

And then install the Haskell Language Server plugin. If you encounter the hie executable missing, please make sure it is installed, see github.com/haskell/haskell-ide-engine when starting VSCode, first make sure that when you run:

which hie

It should give you an output. And the path of the hie executable in the settings.json by adding:

"languageServerHaskell.hieExecutablePath": "~/.local/bin/hie-8.6.5",

Using as a library in a project via `nix`

See the example project in examples/library-example for a default.nix that can be dropped alongside a .cabal file.

Contributing

We welcome new contributors.

Contact Austin Huang or Sam Stites for access to the hasktorch slack channel. You can send an email to hasktorch@gmail.com or on twitter as @austinvhuang and @SamStites.

Developer Information

See the wiki for developer information.

SurajK7/hasktorch