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.