Google has announced the developer preview of TensorFlow Lite, a solution for enabling on-device machine learning inference that has a small binary size and low latency. The solution supports hardware acceleration via the Android Neural Networks API which developers can use to build applications on Android devices that are capable of running computationally intensive operations for machine learning. The API is to be used with machine learning libraries like TensorFlow that allow models to be trained off-device and then deployed on Android devices.
The official product documentation provides a lengthy list of highlights included in the developer preview. A few examples of these highlights include a set of core operators that can be used to build and run custom models, a TensorFlow converter for converting models trained on TensorFlow into a format compatible with TensorFlow Lite, and a new model file format based on FlatBuffers. The developer preview also features several pre-trained models that work out-of-the-box including Inception-V3, MobileNets, and On-Device Smart Reply. These pre-trained models can be downloaded from GitHub, and they have been trained and optimized for mobile devices.
Google already provides an API that allows models to be deployed on mobile and embedded devices, the TensorFlow Mobile API. Google explains in the announcement post how developers should view the two services: “TensorFlow Lite should be seen as the evolution of TensorFlow Mobile, and as it matures it will become the recommended solution for deploying models on mobile and embedded devices. With this announcement, TensorFlow Lite is made available as a developer preview, and TensorFlow Mobile is still there to support production apps.”
For more information about TensorFlow and TensorFlow Lite, visit TensorFlow.org.