Tensorflow, Google’s opensource machine learning library, has toddled into whole figures, just over a year after the first release - with the proviso that version 1.0 might not be entirely backwards-compatible with early adopters’ programmes.
Using just the sort of understatement which characterises the machine learning space right now, technical program manager for TensorFlow, Amy McDonald Sanjideh, said TensorFlow 1.0 is “incredibly fast” and “more production ready than ever”.
McDonald Sanjideh said it would publish implementations showing a 7.3x speedup on 8 GPUs for Inception v3 and 58x speedup for distributed Inception v3 training on 64 GPUs. Earlier versions were marked down for performance.
A key addition is XLA (Accelerated Linear Algebra), an (experimental) domain specific compiler targeting CPUs and GPUs, which the project promises will deliver improvements in “speed, memory usage, and portability on server and mobile platforms”. Specifically, it currently supports just in time compilation on x86-64 and NVIDIA GPUs, and ahead of time compilation on x86-64-64 and ARM.
However, the team says “developers targeting new hardware accelerators are especially encouraged to try out XLA” and said that XLA improved portability to “make it relatively easy to write a new backend for novel hardware.”
Sharing top billing in new features is a high level API, with f.layers, tf.metrics, and tf.losses modules, together with compatibility with Keras.There are experimental APIs for Java and Go.
Python PAIs have been given a scrub up, and the release notes bracket these under “other backwards-incompatible changes made to support API stability” that mean “programs that worked on TensorFlow 0.n won’t necessarily work on TensorFlow 1.0”.
A “handy” migration guide and conversion script is provided, and the project team have said it doesn’t intend to make backwards breaking changes throughout the 1.N lifecycle. Quite how long that lifecycle will be isn’t clear to us - possibly not terribly given the speed with which the 1.0 milestone has been reached.
Other additions include the TensorFlow Debugger, a command line interface and API for debugging TensorFlow programmes live.
You may not be surprised that object detection and localisation and camera-based image stylization is handled by new Android demos.
The release coincided with the TensorFlow Dev Summit. Not surprising, this gave other outfits the chance to hang off this own announcements. Yahoo opensourced its TensorFlow on Spark project, which it described as “latest open source framework for distributed deep learning on big-data clusters.” Last month IBM added TensorFLow support to its Power AI platform. ®
Have you been playing around with TensorFlow, or other similar frameworks? We'd love to hear about your experiences so far at M3, our AI/ML conference, particularly when it comes to production ready applictions. The call for papers is here.