In mlverse/lltm: Long Long Term Memory Neural Network Cells

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

lltm

The goal of lltm is to be a minimal implementation of an extension for torch that interfaces with the underlying C++ interface, called LibTorch.

In this pakage we provide an implementation of a new recurrent unit that is similar to a LSTM but it lacks a forget gate and uses an Exponential Linear Unit (ELU) as its internal activation function. Because this unit never forgets, we’ll call it LLTM, or Long-Long-Term-Memory unit.

The example implemented here is a port of the official PyTorch tutorial on custom C++ and CUDA extensions.

High-Level overview

Writing C++ extensions for torch requires us to coordinate the communication between multiple agents in the torch ecossytem. The following diagram is a high-level overview on how they communicate in this package.

On the torch package side the agents that appear are:

• LibTorch: The PyTorch's C++ interface. This is the library implementing all the heavy computations and the data structures like tensors.
• Lantern: Is a C wrapper for LibTorch and is a part of the torch for R project. We had to develop Lantern because on Windows LibTorch can only be
  compiled with the MSVC compiler while R is compiled with MinGW. Because of the different compilers, only C interfaces (not C++) are compatible.
• torchpkg.so: This is how we are referring to the C++ library, implemented with Rcpp that allows the R API to make calls to Lantern functions. Another important feature it provides is custom Rcpp types that allows users to easily manage memory life time of objects returned by Lantern.

In the extension side the actors are:

• csrc: What we are calling csrc here is the equivalent to Lantern in the torch project. It's a C interface for calling functions from LibTorch that implement the desidered extension functionality. The library produced here must also be compiled with MSVC on Windows thus the C interface is required.
• lltm.so: This is the C++ library implemented using Rcpp that allows the R API to call the csrc functionality. Here, in general, we want to use the torchpkg.so features to manage memory instead of re-implementing that functionality.

Project structure

• csrc: The directory containing library that will call efficient LibTorch code. See the section csrc for details.
• src: Rcpp code that interfaces the csrc library and exports functionality to the R API.
• R/package.R: Definitions for correctly downloading pre-built binaries, and dynamically loading the csrc library as well as the C++ library.

csrc: Implementing the operators and their C wrappers.

• CMakeLists.txt: The first important file that you should get familiar with in this directory is the CMakeLists.txt file. This is the CMake configuration file defining how the project must be compiled and its dependencies. You can refer to comments in the file for almost line by line explanation of definitions.

• csrc/src/lltm.cpp: In this file we define the LibTorch implementation of the operations we want to export. We can use as many functions as we want in the implementation and we mark the functions we want to make available in the R package with // [[torch::export]], similar to what we do when exporting functions with Rcpp. For example we define the lltm_forward implementation with: (For details on the lltm_forward implementation refer to the official guide.)

The // [[torch::export]] marks will allow torchexport that is called during when building with cmake to autogenerate C wrappers necessary to handle errors and to correctly pass data between this library and the R package.

```cpp // [[torch::export]] std::vector lltm_forward( torch::Tensor input, torch::Tensor weights, torch::Tensor bias, torch::Tensor old_h, torch::Tensor old_cell) { auto X = torch::cat({old_h, input}, /dim=/1);

  auto gate_weights = torch::addmm(bias, X, weights.transpose(0, 1));
  auto gates = gate_weights.chunk(3, /*dim=*/1);

  auto input_gate = torch::sigmoid(gates[0]);
  auto output_gate = torch::sigmoid(gates[1]);
  auto candidate_cell = torch::elu(gates[2], /*alpha=*/1.0);

  auto new_cell = old_cell + candidate_cell * input_gate;
  auto new_h = torch::tanh(new_cell) * output_gate;

  return {new_h,
          new_cell,
          input_gate,
          output_gate,
          candidate_cell,
          X,
          gate_weights};

} ```

• csrc/src/exports.cpp: This file is autogenerated by torchexport and should not be modified manually. It wrapps the function that uses LibTorch's API into C API functions that can be called in the R/Rcpp side.

• csrc/include/lltm/exports.h This file includes declarations used by functions defined in exports.cpp. It should always be included in lltm.h. Note that this file is also autogenerated.

• csrc/src/lltm.def: This file is automaticaly generated by a custom CMake command. It lists the functions from lltm.cpp that we want to export. This is only required for Windows, but it's a good practice to keep it up to date. See more information on Module definition files in this link

For example, the current definition is:

LIBRARY LLTM EXPORTS _lltm_forward _lltm_backward

• csrc/include/lltm/lltm.h: In a minimal setup this file only needs to include the lltm/exports.h headers that is auto-generated by torchexport. You might want add other function declarations here, if for some reason you had to bypass the code autogeneration.

The library implemented in csrc can be compiled with CMake. We use the following commands to compile and install it locally:

cd csrc && mkdir build
cmake .. && cmake --build . --target install --config Release

src: Wrapping the library with Rcpp

Now that we implemented the operators that we wanted to call from R, we can now implement the Rcpp wrappers that will allow us to call those operators from R.

• src/exports.cpp: This file is autogenerated and defines Rcpp wrappers for the functions that have been marked with [[torch::export]] in your library. The wrappers defined in this file take R objects and convert them to the correct C type that we need to pass to the C library. Remember that the C library return void* pointers and we need to make sure to free this objects when they are no longer in use, otherwise we will leak memory. The torch.h headers provides Rcpp extension types that act like smart pointers and make sure that the objects created in the C library are correctly freed when they are no longer in use. The types implemented in torch.h also implement convertion from and to SEXPs so we don't need to implement them on our own.

You can find all the available types in the torch namespace available when you include <torch.h>.

• src/lltm.cpp: In a minimal setup this file only needs to include the header files from the torch package as well as from your library and specify a few variables that make sure the implementations are included. It also must define a host_exception_handler that is used to correctly raise exceptions from your C library to the R runtime - in general you don't need to modify the one that's already defined in this template.

cpp #include <Rcpp.h> #define LLTM_HEADERS_ONLY // should only be defined in a single file #include <lltm/lltm.h> #define TORCH_IMPL // should only be defined in a single file #define IMPORT_TORCH // should only be defined in a single file #include <torch.h>

• src/Makevars.win: On Windows, the normal compilation workflow wouldn't work as Windows wouldn't be able to find the implementations of _lltm_forward (as it only sees the headers), so we convert the .def file created in csrc to a .lib file and use this as an argument to the linker. That's what Makevars.win implements. In most cases you won't need to modify this file.

R API

Now the Rcpp wrappers are implemented and exported you have now access to lltm_forward in the R side.

• R/lltm.R: In this package we wanted to provide a new autograd function and a nn_module that uses it and we implemented it in this file. This is normal R code and we won't discuss the actual implementation.

Packaging

It's not trivial to package torch extensions because they can't be entirely built on CRAN machines. We would need to include pre-built binaries in the package tarball but for security reasons that's not accepted on CRAN.

In this package we implement a suggested way of packaging torch extensions that makes it really easy for users to install your package without having to use custom installation steps or building libraries from source. The diagram below shows an overview of the packaging process.

• R/package.R: implements the suggested installation logic - including downloading from GitHub Releases and dynamically loading the shared libraries.

• .github/workflows/R-CMD-check.yaml: the job called Build-Libs implements the logic for building the binaries from csrc for each operating system and uploading to GH Releases.

Installation

~~You can install the released version of lltm from CRAN with:~~

install.packages("lltm")

And the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("mlverse/lltm")

mlverse/lltm documentation built on Jan. 28, 2022, 8:41 p.m.