TensorFlow 1.0: Unlocking machine learning on smartphones

[51CTO.com Quick Translation] As a deep learning framework created by Google, TensorFlow has released a complete set of 1.0 candidate versions.

[[182958]]

Version 1.0 not only increases the number of machine learning functions in the framework, but also simplifies the development process for Python and Java users and enhances the original debugging mechanism. The new compiler can optimize the TensorFlow computing mechanism, which means that more new machine learning applications will run on smartphone-level hardware.

Focus on Python and Java

As Python has become one of the main platforms for building and using machine learning applications, TensorFlow 1.0 has also improved its ability to interact with Python. The TensorFlow Python API has been upgraded, which means that the syntax and metaphors used by TensorFlow can better match Python, ultimately improving the consistency between the two.

Unfortunately, these changes break support for existing Python applications. TensorFlow developers have released scripts that can upgrade existing TensorFlow API scripts to the new format, but it does not solve all problems yet; you may still need to manually adjust your scripts.

TensorFlow is now available in a Python 3-compatible Docker image, and allows Python users to install it using Python's native package manager pip. This will help more general users (rather than those specializing in data science) quickly get started with TensorFlow.

Java is another major language platform for machine learning, but TensorFlow has not been bound to it before. Version 1.0 of the framework introduced a Java API, but it is not yet complete and may change at any time. In addition, to use the Java API, you need to cooperate with Linux or Mac OS environment. (Obviously, Windows is still a second-class citizen in the TensorFlow camp.)

Going mobile with XLA

The biggest change in TensorFlow 1.0 may not come from new support capabilities or new algorithms, but from an experimental linear algebra compiler, Accelerated Linear Algebra (XLA for short). It can generate machine code that can run on the CPU or GPU, thereby significantly improving the efficiency of mathematical operations. At present, XLA only supports NVIDIA GPUs, but it will provide more GPU support for machine learning applications in the future.

XLA also improves the portability of TensorFlow, so existing TensorFlow programs can run on new hardware platforms without modification. This is mainly due to IBM's inclusion of TensorFlow support on its PowerAI hardware solution, which combines GPUs with Power8 CPUs.

Engineers on the TensorFlow project have reduced its overall memory usage and application size. While these optimizations have general benefits, they are particularly evident on mobile. Previous versions of TensorFlow have supported Android, iOS, and Raspberry Pi hardware platforms, and now it can perform operations such as image classification on such devices.

When it comes to machine learning, we often think of all kinds of high-end hardware, including custom CPUs, GPU arrays, FPGAs, and cloud environments. But the new theory is that building executable machine learning models on ordinary smartphones will bring more new application categories. Even if this part of the goal cannot be fully achieved, this effort will still be able to give a strong boost to the development of TensorFlow.

Original title: TensorFlow 1.0 unlocks machine learning on smartphones

Original author: Serdar Yegulalp

[Translated by 51CTO. Please indicate the original translator and source as 51CTO.com when reprinting on partner sites]