TensorFlow Dev Summit And 1.0 Release

Today was the first TensorFlow Dev Summit and also TensorFlow 1.0 release (yep, no more alpha).

Improvements in performance (XLA compiler), TFDBG (TensorFlow debugger) and new Android demos were announced among the new stuff coming out with the 1.0 release.

Some good special news: TensorFlow PIP packages are now PYPI compliant, which means you can install TensorFlow in one line (pip install tensorflow or tensorflow-gpu for GPU support) on Ubuntu, MacOS and Windows, hassle free.

On the Dev Summit one of the coolest things introduced was the TensorBoard Embedding Projector, a tool for high-dimensional data visualization (that I had way too much fun playing with) that runs in the browser. A standalone version was also released and you can try it out here.

MNIST on Embedding Projector

Among the amazing talks given today, from distributed computing to applications in medicine (check out Lily Peng’s talk and how she’s using Machine Learning to diagnose diabetic retinopathy above human accuracy), one of my favorites was Chris Leary and Todd Wang’ talk on XLA, a domain-specific compiler for linear algebra that optimizes TensorFlow computations. Besides the improvement on execution speed and memory usage it also generates smaller binaries improving portability on mobile devices (yep, you can run TensorFlow models on a smartphone, iOS and Android).

XLA: TensorFlow, Compiled!

I am very excited about the advancements of machine intelligence software that are so fundamental for the improvements we’ve been experiencing in Deep Learning, and looking forward to what will be developed further [looking at you too PyTorch ;)].

A few useful links:

Say hi on Twitter.

Happy hacking :)