nGraph (the compiler for the systems of machine learning)

2018: nGraph exit

In March, 2018 the Intel company submitted the compiler for the systems of machine learning nGraph. Its code is written in language C ++ and  is distributed  under the license Apache 2.0.

Within the nGraph project the open compiler, a set of libraries and runtime environment of programs for compilation in executable code of models were released.

Intel submitted the compiler for the systems of machine learning nGraph

By means of nGraph it is possible to create prototypes of the systems of machine learning and to test them without the need for a binding to a certain framework and to adapt the developed models of neural network to a training and accomplishment on different devices. 

Direct support of frameworks of TensorFlow,  Apache MXNet  and  Neon and also indirect —  Caffe2,  PyTorch  and  CNTK  (Cognitive Toolkit) — through intermediate  conversion  to the universal  Open Neural Network Exchange format is implemented. In nGraph are used additional optimization for Intel processors and other platforms.

One of the main advantages of the compiler Intel was intermediate representation of a computation graph (nGraph IR) which abstracts parts of sale of specific devices and allows the developer to focus on scientific work, algorithms and models, without caring for formation of optimum machine code. Each node or transaction in this graph represent one step of calculations which produces at the exit zero or several tensors (multidimensional data arrays) on the basis of zero or more tensors on an input. Each transaction nGraph IR is the assembly block from which it is possible to group more difficult transactions provided by different frameworks of machine learning. 

By March 23, 2018 Graph works only in operating systems of Ubuntu 16.04 and Centos 7.4 and also in Intel Clear Linux — in the experimental mode. There is no support of Windows yet.^[1]

Notes