TensorFlow community announcements

Pengfei Fan and Lingling Jin offer an overview of an efficient and elastic GPU-sharing system for users who do research and development with TensorFlow.

Criteo's real-time bidding of ad spaces requires its TensorFlow (TF) models to make online predictions in less than 5 ms. Nicolas Kowalski and Axel Antoniotti explain why Criteo moved away from high-level APIs and rewrote its models from scratch, reimplementing cross-features and hashing functions using low-level TF operations in order to factorize as much as possible all TF nodes in its model.

Hilbert is an AI framework that works with TensorFlow Extended (TFX) at Yahoo! JAPAN, which provides AutoML to create production-level deep learning models automatically. Hilbert is currently used by over 20 services of Yahoo! JAPAN. Shin-Ichiro Okamoto details how to achieve production-level AutoML and explores service use cases at Yahoo! JAPAN.

MLIR is TensorFlow's open source machine learning compiler infrastructure that addresses the complexity caused by growing software and hardware fragmentation and makes it easier to build AI applications. Chris Lattner and Tatiana Shpeisman explain how MLIR is solving this growing hardware and software divide and how it impacts you in the future.

In many applications where data is generated continuously, combining machine learning with streaming data is imperative to discover useful information in real time. Yong Tang explores TensorFlow I/O, which can be used to easily build a data pipeline with TensorFlow and stream frameworks such as Apache Kafka, AWS Kinesis, or Google Cloud PubSub.

Jack Chung, Chao Liu, and Daniel Lowell explore breaking convolution algorithms into modular pieces to be better fused with graph compilers such as accelerated linear algebra (XLA).