Updated: February 13, 2018
Welcome to the Intel nGraph repo. While we're transitioning our main project from Python and preparing to open-source our C++ code base to the community, you can browse here to learn a bit about the roots of the legacy project.
When Deep Learning (DL) frameworks first emerged as the vehicle for training and inference models, they were designed around kernels optimized for a particular platform. As a result, many backend details were being exposed in the model definitions, making the adaptability and portability of DL models to other or more advanced backends inherently complex and expensive.
The traditional approach means that an algorithm developer cannot easily adapt his or her model to different backends. Making a model run on a different framework is also problematic because the developer must separate the essence of the model from the performance adjustments made for the backend, translate to similar ops in the new framework, and finally make the necessary changes for the preferred backend configuration on the new framework.
We designed the Intel nGraph project to substantially reduce these kinds of engineering complexities. While optimized kernels for deep-learning primitives are provided through the project and via libraries like Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN), there are several compiler-inspired ways in which performance can be further optimized.
Install the nGraph library and write or compile a framework with the library, in order to run training and inference models. Using the command line on any supported system (currently Linux) specify the backend you want to use. Our Intermediate Representation (IR) layer handles all the hardware abstraction details and frees developers to focus on their data science, algorithms and models, rather than on machine code.
At a more granular level of detail:
-
The nGraph core creates a strongly-typed and platform-neutral stateless graph representation of computations. Each node, or op, in the graph corresponds to one step in a computation, where each step produces zero or more tensor outputs from zero or more tensor inputs.
-
We've developed a framework bridge for each supported framework; it acts as an intermediary between the ngraph core and the framework. A transformer then plays a similar role between the ngraph core and the various execution platforms.
-
Transformers handle the hardware abstraction; they compile the graph with a combination of generic and platform-specific graph transformations. The result is a function that can be executed from the framework bridge. Transformers also allocate and deallocate, as well as read and write tensors under direction of the bridge.
You can read more about design decisions and what is tentatively in the pipeline for backends and development in our ARXIV abstract and conference paper: