TensorFlow Newsletter | July, 2017

[Alina Shinkarsky]

TL;DR: Welcome to your TensorFlow update! We'll be sending periodic newsletters to keep you apprised of what’s coming down the pipe. Sundar announced Cloud TPUs, TPU pods, and the TensorFlow Research Cloud at Google I/O, and Fei-Fei encouraged the world to sign up to learn more via g.co/tpusignup. TensorFlow 1.3.0 has been released!

 

High-Level APIs The first round of canned estimators will be moved into core as of TensorFlow 1.3.x, which includes: DNNClassifier, DNNRegressor, DNNLinearCombinedClassifier, and DNNLinearCombinedRegressor.  The new canned estimators are backward compatible and can be used by changing from tf.contrib.learn.estimators.DNNRegressor to tf.estimator.DNNRegressor.

 

News from Performance In case you missed it, in May we rolled out a new set of benchmarks, along with a Performance Guide - check out the blog post. In addition, the following performance improvements were added in the past month: Fused batch norm added to tf.layers.batch_normalization (default: False).  20-30% speedup over non-fused. Intel MKL added as a compile option for open source and may provide significant speedups in some cases.

 

Recent contributions Yahoo contributed IBverbs-based RDMA support for distributed TensorFlow. PR (merged): https://github.com/tensorflow/tensorflow/pull/8943 Minds.ai implemented an MPI-based communication path which complements Yahoo’s work in this space. PR (merged): https://github.com/tensorflow/tensorflow/pull/9864 Our collaboration with Intel produced a PR, primarily focusing on performance improvements in their high-end CPU families (Knights Landing and Broadwell). PR (merged): https://github.com/tensorflow/tensorflow/pull/9580 Intel also published a blog post about these efforts

 

Usability Datasets API for input pipeline: Is more convenient than and typically faster (and getting more so) than queue-based pipelines. Please consider using this going forward.

 

Events TensorFlow Events We've Done Since Last Report RE-Work (Amsterdam) - ML Learning in Amsterdam with 200 attendees ML Camp Jeju - ML teaching program for 20 students in South Korea AI Conference (NYC) - Big Google Brain presence for keynote & workshops Strata (Beijing) - Data conference with TensorFlow session and workshop Cloud + Data NEXT Con (Santa Clara) - NLP session and TensorFlow workshop PLDI & MAPL (Barcelona) - Research conf on Programming Language design Curry On (Barcelona) - New paradigms in software development IIT (Santa Clara) - TensorFlow demos for Big data & ML developers In addition to those, we've talked to a number of companies and universities! Upcoming TensorFlow Events Google Developer Day (Krakow) - Google Developer focused event Deep Learning Indaba (Johannesburg) - Brain & DeepMind teams will be there AI Conference (San Francisco) - A big Google ML team will be there

 

This newsletter was curated by,

The TensorFlow Team