Tensorflow RC 1.0 Released, Android Optimizations Among New Features

Tensorflow RC 1.0 Released, Android Optimizations Among New Features

Feature Image Displays Picture A in the Style of Famous Paintings B,C, and D – Image Credit: Google Research Blogs

Tensorflow – an open-source neural network platform from the Google Brain team – has made available the release candidate for version 1.0 of its increasingly popular machine learning platform. Some of the most exciting new features include pre-made neural networks for Android cameras (person/object detection as well as artistic style transfer), a Java API, and Accelerated Linear Algebra (XLA) integration – a compiler which aims to lessen resource load and optimize applications for mobile use.

Tübingen Neckarfront, Germany by Andreas Praefcke in the style of “Head of a Clown” by Georges Rouault – Image Credit: Google Research Blogs

Improvements in Python and Java

In this version, Python interaction has been upgraded, adopting some of Python’s own syntax and metaphors. Unfortunately, this means that previous Python-based applications of Tensorflow will need to be upgraded to continue functioning in 1.0. Although a conversion script has been released, some scripts still may need to be modified manually. Installing via Python is compatible with MacOS, Linux, and Windows – available as a Pip, Anaconda, or Docker install, among others.

An experimental Java API has also emerged, but for now requires a Linux or MacOS environment to be built from source code.

XLA in Mobile and Beyond

The creators of Tensorflow have long been committed to a universal API, and the implementation of XLA can certainly help. XLA, in essence, utilizes the CPU or GPU to optimize the translation of data from layer to layer of the neural network, resulting in reduced resource load, increased speed, less code, and an overall smaller, more lightweight application. This optimization will also facilitate the porting of server run networks to mobile hardware. Providing the structure, and in some applications, the data to create or use a neural network, Tensorflow’s single API can be used for implementation across desktops, servers, or mobile devices – all that’s required is a unique backend. While offering great potential, XLA is still experimental and the team behind it has asked explicitly for developer input, so that they can quickly bring better machine learning performance to mobile platforms – they’re looking at you Snapdragon 835 fans.

Qualcomm, IBM, Raspberry Pi, Snapchat, and of course Google are a few names on the growing list of companies to either add support for or work closely with Tensorflow and its dedicated team. With this release, the team edges ever-closer to delivering freely implemented neural networks to application developers and consumers alike.

Interested in developing? Check out these links:

Source 1: Tensorflow Source 2: Infoworld

About author