flashlight

Flashlight is a fast, flexible machine learning library written entirely in C++
from the Facebook AI Research and the creators of Torch, TensorFlow, Eigen and
Deep Speech. Its core features include:

• Total internal modifiability including internal APIs for tensor computation.
• A small footprint, with the core clocking in at under 10 MB and 20k lines of C++.
• High-performance defaults featuring just-in-time kernel compilation with modern C++ via the ArrayFire
  tensor library.
• An emphasis on efficiency and scale.

Native support in C++ and simple extensibility makes Flashlight a powerful research framework that enables fast iteration on new experimental setups and algorithms with little unopinionation and without sacrificing performance. In a single repository, Flashlight provides apps for research across multiple domains:

• Automatic speech recognition (formerly wav2letter project) — Documentation | Tutorial
• Image classification
• Object detection
• Language modeling

Project Layout

Flashlight is broken down into a few parts:

• flashlight/lib contains kernels and standalone utilities for audio processing and more.
• flashlight/fl is the core tensor interface and neural network library using the ArrayFire tensor library by default.
• flashlight/pkg are domain packages for speech, vision, and text built on the core.
• flashlight/app are applications of the core library to machine learning across domains.

Quickstart

First, build and install Flashlight and link it to your own project.

Sequential forms a sequence of Flashlight Modules for chaining computation.

#include <flashlight/fl/flashlight.h>

Sequential model;

model.add(View(fl::Shape({IM_DIM, IM_DIM, 1, -1})));
model.add(Conv2D(
    1 /* input channels */,
    32 /* output channels */,
    5 /* kernel width */,
    5 /* kernel height */,
    1 /* stride x */,
    1 /* stride y */,
    PaddingMode::SAME; /* padding mode */,
    PaddingMode::SAME; /* padding mode */));
model.add(ReLU());
model.add(Pool2D(
    2 /* kernel width */,
    2 /* kernel height */,
    2 /* stride x */,
    2 /* stride y */));
model.add(Conv2D(32, 64, 5, 5, 1, 1, PaddingMode::SAME, PaddingMode::SAME));
model.add(ReLU());
model.add(Pool2D(2, 2, 2, 2));
model.add(View(fl::Shape({7 * 7 * 64, -1})));
model.add(Linear(7 * 7 * 64, 1024));
model.add(ReLU());
model.add(Dropout(0.5));
model.add(Linear(1024, 10));
model.add(LogSoftmax());

auto output = model.forward(input);
auto loss = categoricalCrossEntropy(output, target);
loss.backward();

See the MNIST example for a full tutorial including a training loop and dataset abstractions.

Variable is a tape-based abstraction that wraps Flashlight tensors. Tape-based Automatic differentiation in Flashlight is simple and works as you’d expect.

auto A = Variable(fl::rand({1000, 1000}), true /* calcGrad */);
auto B = 2.0 * A;
auto C = 1.0 + B;
auto D = log(C);
D.backward(); // populates A.grad() along with gradients for B, C, and D.

Building and Installing

Install with vcpkg | With Docker | From Source | From Source with vcpkg | Build Your Project with Flashlight

Requirements

At minimum, compilation requires:

• A C++ compiler with good C17 support (e.g. gcc/g >= 7)
• CMake — version 3.10 or later, and make
• A Linux-based operating system.

See the full dependency list for more details if building from source.

Instructions for building/installing Python bindings can be found here.

Flashlight Build Setups

Flashlight can be broken down into several components as described above. Each component can be incrementally built by specifying the correct build options.

There are two ways to work with Flashlight:

1. As an installed library that you link to with your own project. This is best for building standalone applications dependent on Flashlight.
2. With in-source development where the Flashlight project source is changed and rebuilt. This is best if customizing/hacking the core framework or the Flashlight-provided app binaries.

Flashlight can be built in one of two ways:

1. With vcpkg, a C++ package manager.
2. From source by installing dependencies as needed.

Installing Flashlight with vcpkg

Library Installation with vcpkg

Flashlight is most-easily built and installed with vcpkg. Both the CUDA and CPU backends are supported with vcpkg. For either backend, first install Intel MKL. For the CUDA backend, install CUDA >= 9.2, cuDNN, and NCCL. Then, after installing vcpkg, install the libraries and core with:

./vcpkg/vcpkg install flashlight-cuda # CUDA backend, OR
./vcpkg/vcpkg install flashlight-cpu  # CPU backend

To install Flashlight apps, check the features available for installation by running ./vcpkg search flashlight-cuda or ./vcpkg search flashlight-cpu. Each app is a “feature”: for example, ./vcpkg install flashlight-cuda[asr] installs the ASR app with the CUDA backend.

Below is the currently-supported list of features (for each of flashlight-cuda and flashlight-cpu):

flashlight-{cuda/cpu}[lib]      # Flashlight libraries
flashlight-{cuda/cpu}[nn]       # Flashlight neural net library
flashlight-{cuda/cpu}[asr]      # Flashlight speech recognition app
flashlight-{cuda/cpu}[lm]       # Flashlight language modeling app
flashlight-{cuda/cpu}[imgclass] # Flashlight image classification app

Flashlight app binaries are also built for the selected features and are installed into the vcpkg install tree’s tools directory.

Integrating Flashlight into your own project with is simple using vcpkg’s CMake toolchain integration.

From-Source Build with vcpkg

First, install the dependencies for your backend of choice using vcpkg (click to expand the below):

./vcpkg install \
    cuda intel-mkl fftw3 cub kenlm                \ # if building flashlight libraries
    arrayfire[cuda] cudnn nccl openmpi cereal stb \ # if building the flashlight neural net library
    gflags glog                                   \ # if building any flashlight apps
    libsndfile                                    \ # if building the flashlight asr app
    gtest                                           # optional, if building tests

./vcpkg install \
    intel-mkl fftw3 kenlm                              \ # for flashlight libraries
    arrayfire[cpu] gloo[mpi] openmpi onednn cereal stb \ # for the flashlight neural net library
    gflags glog                                        \ # for the flashlight runtime pkg (any flashlight apps using it)
    libsndfile                                         \ # for the flashlight speech pkg
    gtest                                                # optional, for tests

Build Using the vcpkg Toolchain File

To build Flashlight from source with these dependencies, clone the repository:

git clone https://github.com/flashlight/flashlight.git && cd flashlight
mkdir -p build && cd build

Then, build from source using vcpkg’s CMake toolchain:

cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DFL_BUILD_ARRAYFIRE=ON \
    -DCMAKE_TOOLCHAIN_FILE=[path to your vcpkg clone]/scripts/buildsystems/vcpkg.cmake
make -j$(nproc)
make install -j$(nproc) # only if you want to install Flashlight for external use

To build a subset of Flashlight’s features, see the build options below.

Building from Source

To build from source, first install the below dependencies. Most are available with your system’s local package manager.

Some dependencies marked below are downloaded and installed automatically if not found on the local system. FL_BUILD_STANDALONE determines this behavior — if disabled, dependencies won’t be downloaded and built when building Flashlight.

Once all dependencies are installed, clone the repository:

git clone https://github.com/flashlight/flashlight.git && cd flashlight
mkdir -p build && cd build

Then build all Flashlight components with:

cmake .. -DCMAKE_BUILD_TYPE=Release -DFL_BUILD_ARRAYFIRE=ON [...build options]
make -j$(nproc)
make install

Setting the MKLROOT environment variable (export MKLROOT=/opt/intel/oneapi/mkl/latest or export MKLROOT=/opt/intel/mkl on most Linux-based systems) can help CMake find Intel MKL if not initially found.

To build a smaller subset of Flashlight features/apps, see the build options below for a complete list of options.

To install Flashlight in a custom directory, use CMake’s CMAKE_INSTALL_PREFIX argument. Flashlight libraries can be built as shared libraries using CMake’s BUILD_SHARED_LIBS argument.

Flashlight uses modern CMake and IMPORTED targets for most dependencies. If a dependency isn’t found, passing -D<package>_DIR to your cmake command or exporting <package>_DIR as an environment variable equal to the path to <package>Config.cmake can help locate dependencies on your system. See the documentation for more details. If CMake is failing to locate a package, check to see if a corresponding issue has already been created before creating your own.

Minimal setup on macOS

On MacOS, ArrayFire can be installed with homebrew and the Flashlight core can be built as follows:

brew install arrayfire
cmake .. \
      -DFL_ARRAYFIRE_USE_OPENCL=ON \
      -DFL_USE_ONEDNN=OFF \
      -DFL_BUILD_TESTS=OFF \
      -DFL_BUILD_EXAMPLES=OFF \
      -DFL_BUILD_SCRIPTS=OFF \
      -DFL_BUILD_DISTRIBUTED=OFF
make -j$(nproc)

Dependencies

Dependencies marked with * are automatically downloaded and built from source if not found on the system. Setting FL_BUILD_STANDALONE to OFF disables this behavior.

Dependencies marked with ^ are required if building with distributed training enabled (FL_BUILD_DISTRIBUTED — see the build options below). Distributed training is required for all apps.

Dependencies marked with † are installable via vcpkg. See the instructions for installing those dependencies above for doing a Flashlight from-source build.

Build Options

The Flashlight CMake build accepts the following build options (prefixed with -D when running CMake from the command line):

Building Your Own Project with Flashlight

Flashlight is most-easily linked to using CMake. Flashlight exports the following CMake targets when installed:

• flashlight::flashlight — contains flashlight libraries as well as the flashlight core autograd and neural network library.
• flashlight::fl_pkg_runtime — contains flashlight core as well as common utilities for training (logging / flags / distributed utils).
• flashlight::fl_pkg_vision — contains flashlight core as well as common utilities for vision pipelines.
• flashlight::fl_pkg_text — contains flashlight core as well as common utilities for dealing with text data.
• flashlight::fl_pkg_speech — contains flashlight core as well as common utilities for dealing with speech data.
• flashlight::fl_pkg_halide — contains flashlight core and extentions to easily interface with halide.

Given a simple project.cpp file that includes and links to Flashlight:

#include <iostream>

#include <flashlight/fl/flashlight.h>

int main() {
  fl::init();
  fl::Variable v(fl::full({1}, 1.), true);
  auto result = v + 10;
  std::cout << "Tensor value is " << result.tensor() << std::endl; // 11.000
  return 0;
}

The following CMake configuration links Flashlight and sets include directories:

cmake_minimum_required(VERSION 3.10)
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

add_executable(myProject project.cpp)

find_package(flashlight CONFIG REQUIRED)
target_link_libraries(myProject PRIVATE flashlight::flashlight)

With a vcpkg Flashlight Installation

If you installed Flashlight with vcpkg, the above CMake configuration for myProject can be built by running:

cd project && mkdir build && cd build
cmake .. \
  -DCMAKE_TOOLCHAIN_FILE=[path to vcpkg clone]/scripts/buildsystems/vcpkg.cmake \
  -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

With a From-Source Flashlight Installation

If using a from-source installation of Flashlight, Flashlight will be found automatically by CMake:

cd project && mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

If Flashlight is installed in a custom location using a CMAKE_INSTALL_PREFIX, passing -Dflashlight_DIR=[install prefix]/share/flashlight/cmake as an argument to your cmake command can help CMake find Flashlight.

Building and Running Flashlight with Docker

Flashlight and its dependencies can also be built with the provided Dockerfiles; see the accompanying Docker documentation for more information.

Contributing and Contact

Contact: [email protected], [email protected], [email protected], [email protected], [email protected], [email protected],
[email protected], [email protected], [email protected], [email protected]

Flashlight is being very actively developed. See
CONTRIBUTING for more on how to help out.

Acknowledgments

Some of Flashlight’s code is derived from
arrayfire-ml.

Citing

You can cite Flashlight using:

@misc{kahn2022flashlight,
      title={Flashlight: Enabling Innovation in Tools for Machine Learning},
      author={Jacob Kahn and Vineel Pratap and Tatiana Likhomanenko and Qiantong Xu and Awni Hannun and Jeff Cai and Paden Tomasello and Ann Lee and Edouard Grave and Gilad Avidov and Benoit Steiner and Vitaliy Liptchinsky and Gabriel Synnaeve and Ronan Collobert},
      year={2022},
      eprint={2201.12465},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

License

Flashlight is under an MIT license. See LICENSE for more information.