Review of TensorFlow Announcements at Google I / O – 2021
As many of you know, we recently held a conference Google I / O Is Google’s premier developer event. This year, for the first time, the conference was held entirely in an online format, and although we did not manage to hold the conference in a standard format, we hope that we were able to make it accessible to everyone. At the conference, a lot of interesting things for developers of almost all directions were announced. In this article, we’d like to give you an overview of what’s new and updated in the various machine learning product families and what the TensorFlow team has introduced. At the end of the article, you will find a list of all materials.
If you haven’t watched the introductory talk from the TensorFlow team yet, we recommend that you do so. The video is shown below. The records of all performances are posted on the official TensorFlow YouTube channel…
TensorFlow for mobile and web applications
TensorFlow Lite Runtime will ship with Google Play Services
The TensorFlow Lite runtime will be part of the Google Play Services suite. You no longer need to ship it bundled with your application, which greatly reduces the size of applications. Now you can distribute models without worrying about the runtime. Register now you can join the early access program now, and a full-scale update will take place before the end of the year.
TensorFlow Lite models can now be run in the web interface
Run TensorFlow Lite models directly in your browser using the new TFLite API for Web Applications… These APIs are framework compatible TensorFlow.js and support all models from TFLite task librariesdesigned for image classification and segmentation, object recognition, and various natural language processing tasks. New intuitive and user-friendly APIs compatible with TensorFlow.js also allow you to run custom TFLite models… As a result, you can develop machine learning models for websites and mobile devices using a single technology stack.
New site about machine learning on devices
It is not always obvious how to develop an application more efficiently for browser, Android and iOS. Therefore, we created a new machine learning on devices site… He will help you make a choice: use a ready-made solution or develop your own model? Create a cross-platform mobile or browser application? On the site, you will learn about all the stages on the way from an idea to a working application.
We are working on additional tools to enable Android developers to track application performance. IN TensorFlow Lite there is built-in Systrace support and easy integration with Perfetto for Android 10.
The performance analytics updates aren’t just for Android. For iOS developers, TensorFlow Lite now has built-in support for profiling based on signpost technology. If you are using tracing while developing your application, you can launch the Xcode profiler and watch the signpost events to examine the application in detail down to the individual operations.
TFX 1.0: Build Machine Learning Models for Enterprise Use
To turn a prototype machine learning model into a working version requires a strong infrastructure. We needed a robust framework for machine learning products and services. Therefore, we created the TFX platform, and then opened access to its source code so that everyone can use the platform’s capabilities. TFX supports model training for mobile, server and web applications.
We have successfully conducted beta testing with partners and now we present TFX 1.0 – a version that allows you to tailor machine learning models for corporate use. The TFX framework has all the benefits enterprises need, including enterprise-grade support, security updates, bug fixes, and guaranteed backward compatibility with all 1.X releases. TFX is powered by Google Cloud and enables the creation of natural language processing solutions, as well as mobile and web applications.
If you want to use machine learning models for production purposes, TFX is waiting for you. More information can be found at platform website…
We are introducing several new tools that allow you to integrate the Responsible AI framework into the development of any machine learning solution.
Know Your Data
With a tool Know Your Data machine learning researchers and development teams can analyze large datasets of visual and textual data to improve the model and the data itself, and identify and resolve validity issues and bias. Follow the link to find an interactive demo of the tool.
People + AI Guidebook 2.0
When designing AI solutions, it is important to remember that they are for humans. Therefore, we have released version 2.0 of the manual. People + AI Guidebook… The new version will help to put into practice the recommendations for creating human-centered AI. You will find many new resources throughout the tutorial, including code, design patterns, and more.
Get to know also a set of Responsible AI toolsto easily practice Responsible AI using the TensorFlow platform.
Decision forests in Keras
Support for random forests and gradient boosting trees
Machine learning isn’t just about neural networks. Starting with TensorFlow 2.5, you can train powerful algorithm-based models using the familiar Keras APIs. decision forests, including popular variants such as random forest and gradient boosting trees. The new version supports many advanced algorithms for training, interpreting and operating models for regression, classification and ranking of data. TF Serving will power the decision forest and any other model trained with TensorFlow. Explore our guides and see video recording of the report…
TensorFlow Lite for microcontrollers
New board with pre-installed software, experiments and competition
Platform TensorFlow Lite for microcontrollers allows you to run machine learning models on microcontrollers and other devices where memory is limited to a few kilobytes. Now you can buy an Arduino board with preinstalled software and connect to it using Bluetooth and a browser. Try to spend on such a board Google experiments, where it is suggested to set up gesture recognition. You can also create your own classifier or run your TensorFlow model. We have organized a competition for projects using TensorFlow Lite for microcontrollers. Details can be found here… Be sure to watch the video of the TinyML workshop.
Vertex AI: New Managed Machine Learning Platform on Google Cloud
A machine learning model is only useful if it works. Do you know how difficult it can be to create an effective turnkey solution at the right scale. Therefore, Google Cloud is releasing Vertex AI Is a new managed machine learning platform that helps you experiment and deploy AI models faster. The Vertex AI interface has tools for all stages of development. The platform can be used to markup data, work with notebooks and models, forecasting, and continuous monitoring. New MLOps features such as Vertex Pipelines and the Vertex Feature Store differentiate the Vertex AI platform from similar offerings. They simplify support and ensure reproducibility of self-service models.
TensorFlow Cloud: From Local Model Building to Distributed Learning in the Cloud
Library TensorFlow Cloud Provides APIs to help you seamlessly move from building and debugging your model locally to distributed training and hyperparameter tuning in Google Cloud. You will be able to submit a model for customization or training to Google Cloud directly using Colab notebook, Kaggle or local script, without using the Cloud Console. Read about updates on our new site…
New TensorFlow Forum
We have created a new TensorFlow forumwhere you can ask questions and communicate. This is where developers, authors and users can exchange opinions with each other and with the TensorFlow experts. Register and join the community on the page discuss.tensorflow.org…
List of reports
This is just an incomplete list of topics that were discussed at the Google I / O – 2021 conference. All talks about TensorFlow can be viewed in this playlist… Direct links to each report are provided below for your convenience.
Machine Learning News (introductory talk)
Machine Learning for Next Generation Web Applications with TensorFlow.js
Does your application use machine learning? Release a working version with TFX
ML Kit: How to use machine learning on devices with ready-made APIs in mobile apps
Easily implement cross-platform computer vision using the Model Maker library
How to Easily Deploy TF Lite Models Online (demonstration)
Cloud training of TensorFlow models with TensorFlow Cloud (demonstration)
More than an estimate: improving confidence with the Model Remediation tool (demonstration)
Want to learn more about TensorFlow? Visit the site tensorflow.org, read other articles in blog, follow our publications in social networks and subscribe to our YouTube channel… You can also join the one closest to you TensorFlow user community…