Welcome to Marian’s documentation!

Marian is an efficient and self-contained Neural Machine Translation framework with an integrated automatic differentiation engine based on dynamic computation graphs, written entirely in C++.

This is developer documentation. User documentation is available at https://marian-nmt.github.io/docs/

Contents:

• Code Organisation
  • Operating Modes
  • Data
  • Expression Graph
  • Model
  • Utility
  • External Libraries
  • Tests and Examples
• Expression graphs
  • Graph construction
  • Node types
  • Graph execution
• Operations in the expression graph
  • Operator Structure
  • Expression Operator
  • Node Operator
  • Functional Operator
  • Tensor Operator
  • Declared Specialization
• Layers
  • Convolution layer
  • MLP layers
  • RNN layers
  • Embedding layer
• Using Marian with factors
  • Requirements:
  • Define factors
  • Data preparation
  • Create the factored vocabulary
  • Training options
  • Examples
• Library API
  • Page Hierarchy
  • Full API
• How to contribute to Marian
  • Reporting a bug/asking questions
  • Submitting changes
  • Coding conventions
• Writing documentation
  • Code documentation with Doxygen
  • User documentation with Sphinx

Indices and tables

• Index