Introduction to TensorFlow Lite

TensorFlow Lite is TensorFlow’s lightweight solution for mobile and embedded devices. It enables on-device machine learning inference with low latency and a small binary size. TensorFlow Lite also supports hardware acceleration with the Android Neural Networks API.

TensorFlow Lite uses many techniques for achieving low latency such as optimizing the kernels for mobile apps, pre-fused activations, and quantized kernels that allow smaller and faster (fixed-point math) models.

Most of our TensorFlow Lite documentation is on Github for the time being.

What does TensorFlow Lite contain?

TensorFlow Lite supports a set of core operators, both quantized and float, which have been tuned for mobile platforms. They incorporate pre-fused activations and biases to further enhance performance and quantized accuracy. Additionally, TensorFlow Lite also supports using custom operations in models.

TensorFlow Lite defines a new model file format, based on FlatBuffers. FlatBuffers is an open-sourced, efficient cross platform serialization library. It is similar to protocol buffers, but the primary difference is that FlatBuffers does not need a parsing/unpacking step to a secondary representation before you can access data, often coupled with per-object memory allocation. Also, the code footprint of FlatBuffers is an order of magnitude smaller than protocol buffers.

TensorFlow Lite has a new mobile-optimized interpreter, which has the key goals of keeping apps lean and fast. The interpreter uses a static graph ordering and a custom (less-dynamic) memory allocator to ensure minimal load, initialization, and execution latency.

TensorFlow Lite provides an interface to leverage hardware acceleration, if available on the device. It does so via the Android Neural Networks library, released as part of Android O-MR1.

Why do we need a new mobile-specific library?

Machine Learning is changing the computing paradigm, and we see an emerging trend of new use cases on mobile and embedded devices. Consumer expectations are also trending toward natural, human-like interactions with their devices, driven by the camera and voice interaction models.

There are several factors which are fueling interest in this domain:

  • Innovation at the silicon layer is enabling new possibilities for hardware acceleration, and frameworks such as the Android Neural Networks API make it easy to leverage these.

  • Recent advances in real-time computer-vision and spoken language understanding have led to mobile-optimized benchmark models being open sourced (e.g. MobileNets, SqueezeNet).

  • Widely-available smart appliances create new possibilities for on-device intelligence.

  • Interest in stronger user data privacy paradigms where user data does not need to leave the mobile device.

  • Ability to serve ‘offline’ use cases, where the device does not need to be connected to a network.

We believe the next wave of machine learning applications will have significant processing on mobile and embedded devices.

TensorFlow Lite developer preview highlights

TensorFlow Lite is available as a developer preview and includes the following:

  • A set of core operators, both quantized and float, many of which have been tuned for mobile platforms. These can be used to create and run custom models. Developers can also write their own custom operators and use them in models.

  • A new FlatBuffers-based model file format.

  • On-device interpreter with kernels optimized for faster execution on mobile.

  • TensorFlow converter to convert TensorFlow-trained models to the TensorFlow Lite format.

  • Smaller in size: TensorFlow Lite is smaller than 300KB when all supported operators are linked and less than 200KB when using only the operators needed for supporting InceptionV3 and Mobilenet.

  • Pre-tested models:

    All of the following models are guaranteed to work out of the box:

    • Inception V3, a popular model for detecting the dominant objects present in an image.

    • MobileNets, a family of mobile-first computer vision models designed to effectively maximize accuracy while being mindful of the restricted resources for an on-device or embedded application. They are small, low-latency, low-power models parameterized to meet the resource constraints of a variety of use cases. They can be built upon for classification, detection, embeddings and segmentation. MobileNet models are smaller but lower in accuracy than Inception V3.

    • On Device Smart Reply, an on-device model which provides one-touch replies for an incoming text message by suggesting contextually relevant messages. The model was built specifically for memory constrained devices such as watches & phones and it has been successfully used to surface Smart Replies on Android Wear to all first-party and third-party apps.

  • Quantized versions of the MobileNet model, which runs faster than the non-quantized (float) version on CPU.

  • New Android demo app to illustrate the use of TensorFlow Lite with a quantized MobileNet model for object classification.

  • Java and C++ API support

Getting Started

We recommend you try out TensorFlow Lite with the pre-tested models indicated above. If you have an existing mode, you will need to test whether your model is compatible with both the converter and the supported operator set. To test your model, see the documentation on GitHub.

Retrain Inception-V3 or MobileNet for a custom data set

The pre-trained models mentioned above have been trained on the ImageNet data set, which consists of 1000 predefined classes. If those classes are not relevant or useful for your use case, you will need to retrain those models. This technique is called transfer learning, which starts with a model that has been already trained on a problem and will then be retrained on a similar problem. Deep learning from scratch can take days, but transfer learning can be done fairly quickly. In order to do this, you'll need to generate your custom data set labeled with the relevant classes.

The TensorFlow for Poets codelab walks through this process step-by-step. The retraining code supports retraining for both floating point and quantized inference.

TensorFlow Lite Architecture

The following diagram shows the architectural design of TensorFlow Lite:

Starting with a trained TensorFlow model on disk, you'll convert that model to the TensorFlow Lite file format (.tflite) using the TensorFlow Lite Converter. Then you can use that converted file in your mobile application.

Deploying the TensorFlow Lite model file uses:

  • Java API: A convenience wrapper around the C++ API on Android.

  • C++ API: Loads the TensorFlow Lite Model File and invokes the Interpreter. The same library is available on both Android and iOS.

  • Interpreter: Executes the model using a set of kernels. The interpreter supports selective kernel loading; without kernels it is only 100KB, and 300KB with all the kernels loaded. This is a significant reduction from the 1.5M required by TensorFlow Mobile.

  • On select Android devices, the Interpreter will use the Android Neural Networks API for hardware acceleration, or default to CPU execution if none are available.

You can also implement custom kernels using the C++ API that can be used by the Interpreter.

Future Work

In future releases, TensorFlow Lite will support more models and built-in operators, contain performance improvements for both fixed point and floating point models, improvements to the tools to enable easier developer workflows and support for other smaller devices and more. As we continue development, we hope that TensorFlow Lite will greatly simplify the developer experience of targeting a model for small devices.

Future plans include using specialized machine learning hardware to get the best possible performance for a particular model on a particular device.

Next Steps

For the developer preview, most of our documentation is on GitHub. Please take a look at the TensorFlow Lite repository on GitHub for more information and for code samples, demo applications, and more.