Google today released a tool that converts AI models produced for mobile devices using its TensorFlow Lite tool into Apple’s Core ML .mlmodel file format for use with iOS devices.
The converter was made in collaboration with Apple, according to a Google developers blog post. TensorFlow Lite was designed to produce lighter-weight machine learning models that run quickly on mobile devices, while still allowing developers to build using Google’s popular TensorFlow open source framework.
Core ML is designed to provide an optimized execution environment for deploying AI services like object identification or natural language processing to iOS apps. Like TensorFlow Lite, it’s supposed to help tackle one of the key problems with machine learning computation on mobile devices: while models can produce intelligent results, they often require a great deal of computation power that can run slowly on devices without the full firepower of a server, and consume a great deal of precious battery, to boot.
As part of the release of the converter and an update to Core ML Tools, TensorFlow is now featured on the Core ML for developers webpage.
Also part of the Core ML update: developers can now create custom layers for models running on devices running iOS 11.2 or higher, and 16-bit Floating Point support is now available for neural networks, computing which can greatly reduce the size of AI models, an Apple spokesperson told VentureBeat in an email.
TensorFlow Lite developer preview for makers of iOS and Android apps was first made available last month. Created especially for the deployment of AI in mobile devices, TensorFlow Lite comes prepared to deploy some premade models, such as MobileNet and Inception-v3 for object identification with computer vision and Smart Reply for suggested responses.
Google will still support the creation of cross-platform models that run on both iOS and Android through TensorFlow Lite and its custom .tflite file format.
Updated 3:20 p.m. to include additional Core ML information from an Apple spokesperson.
Correction 12:00 Pacific: This story has been corrected to clarify that the tool released today converts from TensorFlow Lite’s .tflite file format to CoreML’s .mlmodel format, not the other way around.