Developers of deep learning, machine learning and other artificial intelligence are increasingly adopting the TensorFlow framework.

Though TensorFlow is not an official Apache project, it was open-sourced a few years ago by its developer, Google LLC, which continues to make deep investments in the framework. At last week’s TensorFlow Developer Summit in Mountain View, California, Google made several announcements that show its commitment to TensorFlow’s evolution remains strong.

TensorFlow’s maturation into a truly enterprise-grade AI development tool is now well underway. As more than 500 in-person attendees and many more via livestream saw, Google announced it has made the following fresh investments in growing the tool, the enabling platform and the developer ecosystem around TensorFlow:

• End-to-end TensorFlow development pipeline platform: Google announced a roadmap for a future end-to-end AI pipeline platform called TensorFlow Extended or TFX. This platform will include components that have been released so far, including TensorFlow Model Analysis, TensorFlow Transform, Estimators and TensorFlow Serving, as well as future integrated components to help developers prepare data, train, validate and deploy TensorFlow models in production.
• TensorFlow client support: Google announced TensorFlow.js, which is a new browser-based ML framework for JavaScript developers. TensorFlow.js is an interactive framework for developing client-side ML applications where the data remains entirely in the browser. It supports ML model building and training entirely in the browser. It also supports import of TensorFlow and Keras models that were trained offline in order to support browser-based inference using WebGL acceleration. Google announced updates to TensorFlow Lite, including a lighter-weight, faster core interpreter for deploying trained ML models on mobile and other edge devices, including those running on Raspberry Pi. And it announced that it is open-sourcing TensorFlow for Swift in support of ML/DL development for mobile apps running in iOS.
• TensorFlow developer experience: Google rolled out an intuitive programming model for TensorFlow developers. This new model, called eager execution, provides an imperative Python programming environment that evaluates computational graph operations immediately in TensorFlow, without requiring an extra graph-building step.  Eager execution enables quick iteration on small models and small data. It helps AI developers to get started rapidly and debug TensorFlow models interactively, rather than always having to consrtruct computational graphs for later execution.
• TensorFlow model sharing and reuse: Google launched a new library, TensorFlow Hub, that is designed to help developers share and reuse DL models. TensorFlow Hub encourages the publication and discovery of self-contained modular pieces of TensorFlow graphs for reuse across similar tasks. The modules have been pretrained on large datasets and may be retrained and used in new applications.
• TensorFlow model evaluation and debugging: Google announced TensorFlow Model Analysis, an open-source library that combines TensorFlow and Apache Beam to compute and visualize model evaluation metrics. It also launched the Debugger Dashboard, which is an interactive graphical debugger plug-in for TensorBoard visualization tool that helps developers to inspect and step through internal nodes of a TensorFlow computation graph in real-time.
• TensorFlow hardware support: Google debuted a new method for running TensorFlow Estimator models on multiple graphics processing units on a single machine, thereby enabling developers to quickly scale their models with minimal code changes. It announced TensorFlow’s integration with NVIDIA’s TensorRT library, which optimizes DL models for inferencing on GPUs and creates a runtime for deployment on GPUs in production environments. Google also announced that it and Intel has delivered integration with a faster, more efficient Intel MKL-DNN open source DL library Furthermore, it said TensorFlow now runs on Google’s Cloud TPUs, which were released in beta in February and will be available to TensorFlow users with the upcoming TensorFlow 1.8 release.
• TensorFlow statistical algorithm support: Google announced that TensorFlow now supports Bayesian analysis via the TensorFlow Probability API and library, which contains building blocks such as probability distributions, sampling methods and new metrics and losses. It also announced new premade high-level classes for training and deployment of boosted decision trees.
• TensorFlow application domains: Google released the Nucleus genomics library for TensorFlow, which supports reading, writing and filtering common genomics file formats. In tandem with DeepVariant, an open-source TensorFlow based tool for genome variant discovery, this is designed to accelerate genomics research and development in TensorFlow.

To build the developer ecosystem around TensorFlow, Google also announced a range of new community resources, including an official TensorFlow blog, a TensorFlow YouTube channel, new TensorFlow mailing lists and TensorFlow Special Interest Groups.

For more insight, here is Google’s Ron Bodkin recently on theCUBE at the recent Big Data SV event. He discussed the company’s strategy and investments in TensorFlow and other components of its AI/DL/ML portfolio and roadmap:

Image: Google