TensorFlow, opensource software for machine learning has been used many Google products such as Google Translate to make the translation result so better, Gmail for Smart Reply to predict answers email replies, machine learning is used to diagnose disease diabetic retinopathy, machine learning to make artwork illustration of music, machine learning is used to optimize the performance of processor Qualcomm, and many more.
Now TensorFlow has release version 1.0. The release of version 1.0 was announced as part of the event TensorFlow Developer Summit which was held yesterday. Google says there are three main things for this release, faster, more flexible and better equipped for the needs of production. This version got a lot of improvements from the new API module performace, as part of a high-level API for TensorFlow Python API stability and promising.
It's faster: TensorFlow 1.0 is very fast! XLA puts the basis for improved performance in the future, and tensorflow.org now contains tips and tricks to perform tuning you models to achieve the maximum speed. We will soon publish the renewal of implementation of some popular models to demonstrate how to take advantage of TensorFlow 1.0 including speeding up 7.3 x 8 GPU for Inception v3 58x and accelerate distribution of Inception v3 64 on the GPU!
It's more flexible: TensorFlow 1.0 introduces the high level API for TensorFlow, with tf. layers, tf. metrics, and module tf. losses. We've also announced the inclusion of a module tf. new hardware that provides full compatibility with hard, high nervous network library-another popular level.
It's more production-ready than ever: TensorFlow 1.0 is promising stability Python API (details here), making it easier to implement new features without the worry of damaging code you have there.
There are some other changes in this version, Google says "Python API has been modified more akin to NumPy", there is a new experimental API for Java and Go, and also a new demo of Android for the object detection and localization.
No comments:
Post a Comment