One of the major announcements from I/O 2017 was TensorFLow Lite created for machine learning on mobile devices. Developers were already informed about the launch of TensorFlow in May itself.
TensorFlow was in use for many devices starting right from the servers to loT devices and varied platforms, but the demand of adapting machine learning models have increased in the past which has lead to an increase in the need of deploying them on mobile embedded devices.
On device machine learning models, TensorFlow Lite enables low-latency inference. With an aim of creating lightweight machine learning solution for smartphone and embedded devices the company came with the invention of TensorFLow for both Android and iOS app developers.
More emphasis will be laid on introducing low-latency inference from machine learning models to less robust devices and not on training models. To put it in simple words, TensorFlow will ensure that existing capabilities of models are being applied to the new data.
TensorFlow has been effectively trained for:
Inception v3 which is image recognition model offering higher accuracy and also larger size.
MobileNet which has got the capability of identifying 1000 different object classes and have been designed for mobile and embedded devices.
Smart Reply is on- device conversational model which enables one-touch reply to the incoming chat messages.
Google has mentioned that while they were designing TF Lite they laid much emphasis on lightweight product as that can help in initializing quickly and will lead to an improvement in the model on various mobile devices.
While redesigning TensorFlow Lite from the scratch focus was majorly laid on 3 areas:
Cross Platform, where in order to run many different platforms including both Android and iOS runtime was designed.
Lightweight enables inference of on-device machine learning models to be developed with a small binary size featured with fast startup.
Fast - To improve the model loading times and so as to support hardware acceleration, focus was laid on making the optimisation of mobile devices a little faster.
Google even mentioned that the full release is yet to come as more things will be added to the bucket. At present TensorFLow Lite is active for models like MobileNet, Inception v3 and Smart Reply.
Google also stated that considering the needs of the developers, constrained platforms have been started so as to ensure effective performance of most important common models.
Further they added, in future they have planned to prioritize functional expansion depending on the needs and demands of the users.
They will also be working on the goal of simplifying the developers experience and also model deployment for various mobiles and embedded devices.
No comments:
Post a Comment