Quickstart | Installation | Python Documentation | Citing

Flashlight Sequence is a library with fast implementations of sequence-based operations. It includes:

• A fast, parallel CPU implementation of the Viterbi algorithm for greedy "argmax-style" decoding
• Fast implementations (CPU and CUDA) of the Wav2letter ASG loss function including the fully-connected and forced-alignment algorithms.

Flashlight Sequence has Python bindings. To install the bindings from source, [optionally install CUDA] then clone the repo and build:

git clone https://github.com/flashlight/sequence && cd sequence
pip install .

To install with CUDA support, set the environment variable USE_CUDA=1 when running the install command. By default, bindings are installed with OpenMP support; to build and install without OpenMP, set the environment to have USE_OPENMP=0 when buildling.

See the full Python binding documentation for examples and more.

From Source (C++) | With vcpkg (C++) | From Source (Python) | Adding to Your Own Project (C++)

At minimum, C++ compilation requires:

• A C++ compiler with good C++17 support (e.g. gcc/g++ >= 7)
• CMake — version 3.16 or later, and make
• A Linux-based operating system.

CUDA Support: If building with CUDA support, CUDA >= 9 is recommended. To toggle CUDA support use the FL_SEQUENCE_USE_CUDA CMake option or the USE_CUDA environment variable when building the Python bindings. To toggle OpenMP support, use the FL_SEQUENCE_USE_OPENMP CMake option or use the USE_OPENMP environment variable when building the Python bindings.

Tests: If building tests, Google Test >= 1.12 is required, but is installed automatically on build if not found. The FL_SEQUENCE_BUILD_TESTS CMake option toggles building tests.

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

Building the C++ project from source is simple:

git clone https://github.com/flashlight/sequence && cd sequence
cmake -S . -B build
cmake --build build --parallel
cd build && ctest && cd .. # run tests
cmake --install build # install at the CMAKE_INSTALL_PREFIX

To enable CUDA while building, pass -DFL_SEQUENCE_USE_CUDA=ON to CMake. To enable building with OpenMP, pass -DFL_SEQUENCE_USE_OPENMP=ON to CMake. To disable building tests, pass -DFL_SEQUENCE_BUILD_TESTS=OFF.

If building with CUDA < 11, NVIDIA cub is required. It will be downloaded automatically if not found; the FL_SEQUENCE_BUILD_STANDALONE build option controls this behavior.

Flashlight Sequence can also be installed and used downstream with the vcpkg package manager. The port contains optional features for building with OpenMP and/or CUDA:

vcpkg install flashlight-sequence # no dependencies, or:
vcpkg install "flashlight-sequence[cuda]" # with CUDA
vcpkg install "flashlight-sequence[openmp]" # with OpenMP
vcpkg install "flashlight-sequence[cuda,openmp]" # with both!

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

#include <iostream>

#include <flashlight/lib/sequence/criterion/cpu/ViterbiPath.h>

int main() {
  auto res = fl::lib::cpu::ViterbiPath<float>::compute(...);
  std::cout << "ViterbiPath result[0] " << res[0] << std::endl;
 return 0;
}

The following CMake configuration links Flashlight and sets include directories:

cmake_minimum_required(VERSION 3.16)
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

add_executable(myProject project.cpp)

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

Contact: [email protected]

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

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}
}

For the AutoSegmentation Criterion (ASG), cite:

@article{collobert2016wav2letter,
  title={Wav2letter: an end-to-end convnet-based speech recognition system},
  author={Collobert, Ronan and Puhrsch, Christian and Synnaeve, Gabriel},
  journal={arXiv preprint arXiv:1609.03193},
  year={2016}
}

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