Apache MXNet

What MXNet is

MXNet is a machine learning library for building deep neural networks, and it was designed to be used from more than one programming language rather than tied to a single one^[1]. Its authors describe it as blending two styles of programming inside the host language: declarative symbolic expressions, in which a whole computation is described first and carried out afterwards, and imperative tensor computation, in which operations take effect as they are written^[1]. Both styles work on the same underlying arrays of numbers; what differs is when the shape of the computation is fixed.

The library also performs automatic differentiation, deriving the gradients that training needs from the computation the user has described instead of requiring them to be written out by hand^[1]. And it is built to run on heterogeneous systems, from mobile devices up to distributed clusters of GPUs^[1], so that the same code can travel across very different hardware.

Where it came from

The MXNet paper was first posted on 3 December 2015 by Tianqi Chen, Mu Li and their colleagues^[1]. The project presents itself as a truly open source deep learning framework, aimed both at research prototyping and at production use^[2]. It is now known as Apache MXNet.

A hybrid front end

MXNet's hybrid front end switches between Gluon imperative mode and symbolic mode^[2]. In imperative mode a line of code runs as soon as it is written, which makes it straightforward to step through a model, print intermediate values and find where something goes wrong. In symbolic mode the same model is treated as a declaration that can be examined as a whole before anything runs, which is what allows the library to rearrange and optimise the work. The hybrid front end exists precisely to move between the two, so that the choice between them is a choice about how to work rather than a choice of a different tool^[2].

Many machines, many languages

Distributed training is supported through both a parameter server and Horovod^[2]. A parameter server keeps the model's parameters on a set of server processes while worker processes compute gradients and send them back, and Horovod is an alternative way of organising the communication among many workers. Either route lets a training job spread over more than one machine.

The library also reaches users through its language bindings: MXNet offers eight of them, with deep integration into Python plus Scala, Julia, Clojure, Java, C++, R and Perl^[2]. Someone working mainly in R or Julia can therefore call the same underlying library as someone working in Python, instead of switching ecosystems.

The ecosystem

Around the library sits a set of companion projects. There is the D2L.ai book, and there are domain-specific toolkits: GluonCV for computer vision, GluonNLP for language and GluonTS for time series^[2]. These packages collect models and building blocks for their respective areas on top of the core library, so that users do not start each project from raw tensor operations.

Where it sits

Taken together, MXNet's design choices point in one direction: a single library that covers a wide range of ways of working. It offers a symbolic style and an imperative style, switched through a hybrid front end^[2]; it derives gradients automatically^[1]; it targets hardware from phones to GPU clusters^[1]; it trains across machines through a parameter server or Horovod^[2]; and it is callable from eight languages^[2]. Its stated goal is to serve both research prototyping and production^[2], which is why the same project carries a book and domain toolkits alongside the core engine.