This repository provides the Lua-based fairseq toolkit for sequence-to-sequence learning, specifically tailored for Neural Machine Translation (NMT). It implements convolutional and LSTM-based models, offering multi-GPU training and fast beam search generation. The toolkit is intended for researchers and practitioners in NLP and NMT.

How It Works

Fairseq utilizes Torch (Lua) for its backend, enabling efficient tensor operations and GPU acceleration. It implements state-of-the-art NMT architectures, including convolutional sequence-to-sequence models and standard LSTM-based models. The toolkit supports multi-GPU training for faster model development and includes optimized generation routines for both CPU and GPU, facilitating efficient inference.

Quick Start & Requirements

• Install: Clone the repository and run luarocks make rocks/fairseq-scm-1.rockspec. For CPU-only translation, use luarocks make rocks/fairseq-cpu-scm-1.rockspec.
• Prerequisites: macOS or Linux, Torch installation (LuaJIT and Intel MKL recommended), and a recent nn package (from May 5th, 2017 or later). NVIDIA GPU and NCCL are required for training.
• Resources: Training requires significant GPU resources. Pre-trained models are available for English-French, English-German, and English-Romanian translation.
• Docs: fairseq-py (PyTorch version) is the focus of new development.

Highlighted Details

• Implements Convolutional Sequence to Sequence Learning and A Convolutional Encoder Model for Neural Machine Translation.
• Supports multi-GPU training on a single machine.
• Features fast beam search generation on CPU and GPU.
• Provides pre-trained models for several language pairs.

Maintenance & Community

This Lua version is preserved but provided without support, with new development focusing on the PyTorch version. Community links include a Facebook group and Google group.

Licensing & Compatibility

BSD-licensed, including pre-trained models, with an additional patent grant. Compatible with commercial use.

Limitations & Caveats

This Lua version is no longer actively developed or supported, with all new efforts directed towards the PyTorch implementation. Users should be aware of potential compatibility issues with modern hardware or operating systems due to its age.