At Google’s first ever TensorFlow Developer Summit, the company announced it has updated its machine learning software to version 1.0. This update contains three key improvements: performance, compatibility via APIs and flexibility. Depending on the platform and hardware TensorFlow runs on, the new version 1.0 can be considerably quicker to execute. Google’s blog on the subject explains how the performance improvement comes from using XLA, which is a domain-specific compiler for TensorFlow graphs. XLA targets CPUs and GPUs and is currently “rapidly evolving.” Google explains that users should “expect to see more progress in upcoming releases.” In terms of numbers, Google details how Inception v3 code running on systems with 8 GPUs now runs 7.3 times quicker. For distributed Inception v3 on 64 GPU hardware, the software is now an impressive 57 times quicker. On the compatibility and flexibility stage, TensorFlow 1.0 introduces a high level API including tf.layers, tf.metrics, and tf.losses. The Python APIs have been redesigned and are now closer in structure to NumPY, which improves stability and simplifies the process of incorporating TensorFlow with existing Python applications – in Google’s words, this makes TensorFlow 1.0 “more production ready.” As there have been a number of changes here, and how some are described as “backwards-incompatible” by Google, for developers seeking more information there’s a handy application compatibility and porting guide linked from the source, below.
Google have also included a Keras compatibility API, allowing TensorFlow to work with this high level neural networks library. There are experimental APIs for Java and Go and Google have also introduced the TensorFlow Debugger, tfdbg, described as “a command-line interface and API for debugging live TensorFlow programs.” And finally, Google have included new Android demos for object detection, localization and on-camera image stylization. The installation process has been streamlined too, which makes it easier to use TensorFlow.
TensorFlow, which has been available for around a year, and is also ready used by a number of Google products including Google Photos and Google Translate, plus external applications in the healthcare sector, where it’s used for cancer detection amongst other applications. Overall, Google is aware of over 6,000 online open source repositories containing TensorFlow. A number of hardware designers and vendors have taken note of TensorFlow and are designing the technology into their products, such as Qualcomm engineering TensorFlow into the Snapdragon 835.