Marian: Fast Neural Machine Translation in C++

TL;DR

Marian is a C++-based neural machine translation framework that offers high-speed training and translation, featuring an integrated automatic differentiation engine and a flexible encoder-decoder design.

Contribution

The paper introduces Marian, a fast, self-contained NMT toolkit in C++ with an integrated automatic differentiation engine and a flexible architecture.

Findings

Achieves high training and translation speed

Supports research-friendly development environment

Fully implemented in C++ with dynamic computation graphs

Abstract

We present Marian, an efficient and self-contained Neural Machine Translation framework with an integrated automatic differentiation engine based on dynamic computation graphs. Marian is written entirely in C++. We describe the design of the encoder-decoder framework and demonstrate that a research-friendly toolkit can achieve high training and translation speed.