NNEF Parser Project

This repository contains C++ and Python source code for a sample NNEF parser.

Introduction

The code consists of a C++ library that contains two example parsers (one for flat and one for compositional NNEF syntax). This library can be used to build tools that require parsing NNEF files. It requires a C++11 compatible compiler. The sample.cpp contains a minimal example that showcases the use of the parser.

The Python code wraps the C++ parser and adds some further utilities to load and save NNEF documents easily. It also contains a script to validate NNEF documents (validate.py) and optionally print a lowered version of the graph. If the tool encounters an invalid document, it prints the first error and stops parsing. Type python validate.py -h to show the usage help.

Building the C++ library

The C++ library can be compiled with cmake.

Example of build commands under Linux:

$ cd NNEF-Tools/parser/cpp
$ mkdir build && cd build
$ cmake ..
$ make

Using the C++ library

Using the C++ parser is as simple as follows:

#include "nnef.h"

nnef::Graph graph;
std::string error;
bool success = nnef::load_graph("path/to/NNEF/folder", graph, error);

Upon succeess, the graph structure is filled, while in case of an error, the error string is filled. The fields inside the graph structure, and further parameters to the load_graph function are documented in nnef.h. After the graph is successfully loaded, shape inference can be performed in a subsequent call if required:

success = nnef::infer_shapes(graph, error);

Upon success, the shape fields of tensors are filled in.

Building the Python module

The python folder contains a Python wrapper for the C++ parser code. To build the python module, move into the python folder and run

cd NNEF-Tools/parser/python
python setup.py install

This invokes the system compiler for C++ (e.g. gcc, g++, clang depending on the operating system), builds and installs an 'nnef' python module. If that command succeeds, the nnef module can be used within the Python interpreter.

Using the Python module

In the python interpreter, type

import nnef
graph = nnef.load_graph('example.nnef')

If the path (example.nnef) points to a folder (with a graph.nnef in it), the whole model with weights is loaded. If it points to a file, it is interpreted as the graph description only, and it is loaded without weights.

Alternatively, the methods

graph = nnef.parse_file("graph.nnef", quantization = "graph.quant")

and

graph = nnef.parse_string("version 1.0; graph ...", quantization = "...")

can be used to parse a graph and optional quantization info from files or strings.

After invocation, graph is a data structure (named tuple) containing the name, tensors, operations, inputs and outputs of the graph. See nnef.py and python/sample.py for more details. If shape information is also required, it can be obtained by calling nnef.infer_shapes(graph), which updates the shape information on the graph structure in place.