Posted by the TensorFlow Team
2019 was an exciting year for TensorFlow. From releasing TensorFlow 2.0 and several product updates to hosting a global roadshow in 11 different countries and the first ever TensorFlow World, this year highlighted how TensorFlow is helping to empower developers, researchers, and enterprises around the world to solve challenging, real-world problems with ML. With TensorFlow turning 4 years old this year, we’re taking a look back at some of our efforts, and things we’re looking forward to in 2020.
TensorFlow Dev Summit 2019 & TensorFlow 2.0
We kicked-off the year with our
third annual Developer Summit in Sunnyvale, CA. With over 1,000 attendees and thousands more via
livestream, it was our largest Dev Summit yet. We announced the alpha release of
TensorFlow 2.0, the biggest release of the platform to-date that makes building ML systems easier.
We also announced new educational resources and partnerships with
deeplearning.ai on Coursera and
Udacity to create courses that will help train the next generation of ML users.
On-device ML with TensorFlow
2019 was a big year for TensorFlow running on-device. It powered everything from an app that can help measure
air quality to
Coral Dev Boards, and even a live demo on-stage at
I/O. For TensorFlow Lite updates, we added post-training
float16 quantization and a new
pruning API as part of the Model Optimization Toolkit as well as releasing a
guide for TensorFlow Lite on
microcontrollers.
With ML running on many different platforms and devices, we were excited to announce
MLIR, a flexible infrastructure that addresses the complexity caused by growing software and hardware fragmentation and makes it easier to build AI applications.
TensorFlow Roadshows
To connect with local communities around the world, we visited over 11 different cities to host the TensorFlow Roadshows, a gathering of the TensorFlow community to discuss developments and share community highlights. We hope you were able to attend one of them, but if not, be sure to get involved through
forums,
user groups, and
Special Interest Groups (SIGs). Keep an eye out on our
Twitter to see if the TensorFlow Roadshow will stop by in your city next year!
Community Updates
The TensorFlow community is what makes TensorFlow one of the most popular ML platforms in the world and in 2019 - they did a lot to contribute to the ecosystem! From
answering questions on Stack Overflow, to actively engaging with users on
Twitter (@tensorflow), to helping to
translate documentation and creating
Special Interests Groups, the community contributed significantly to all aspects of the TensorFlow ecosystem this year. We also wanted to make sure we supported their efforts by improving the developer experience by redesigning
tensorflow.org, translating documentation, and launched a new
blog.
|
#TFWorld 2.0 Challenge on Devpost |
For the first time, we are participating in
Google Code-in, a global, online contest that introduces teenagers to the world of open source development. We are also launching a new
2.0 hackathon on DevPost, to let you share your latest and greatest projects, and win prizes. Additionally, we partnered with Kaggle to launch a
contest that challenges you with a question answering task. It’s open for entry until January 22nd, so best of luck, and we will feature winners at
TensorFlow Developer Summit 2020.
TensorFlow 2.0 general release
As a follow up to the alpha release announcement at Dev Summit 2019, in September, we were excited to
officially release TensorFlow 2.0! It was a culmination of the community telling us what they wanted in an easy-to-use platform that works on any device.
Add-ons and extensions are an important part of the TensorFlow ecosystem, which is why we wanted to make sure they are also 2.0 compatible. You can now use popular libraries like
TF Probability,
TF Agents, and
TF Text with 2.0. We’ve also introduced a host of new libraries to help researchers and ML practitioners in more useful ways such as
Neural Structure Learning and the new
Fairness Indicators add-on.
Using TensorFlow 1.x and want to convert to 2.0? Our team has created a
migration guide and a guide on
how to be effective with TensorFlow 2.0.
TensorFlow World
We ended the year with the inaugural TensorFlow World at the Santa Clara Convention Center, in partnership with O’Reilly Media. The event was attended by over 1,000 machine learning enthusiasts and watched over livestream by thousands more.
The event was comprised of two days of technical training and two days of keynote programming. At the event, we made some exciting announcements, including:
Updates to TensorFlow Hub and TensorBoard.dev
We updated the
TensorFlow Hub experience so that it’s more intuitive where you can find all of the pre-trained models in the TensorFlow ecosystem, like
BERT. This means you can find models related to image, text, video, and more that are ready to use with TensorFlow Lite and TensorFlow.js.
TensorBoard, TensorFlow’s visualization toolkit, helps anyone analyze their ML experiments in detail. At TensorFlow World, we announced
TensorBoard.dev, a managed TensorBoard experience that lets you upload and share your ML experiment results with anyone. You’ll now be able to host and track all your ML experiments and share them easily for free - no setup required. Just simply upload your logs and create a URL!
New educational resources
The new
Learn ML page on
TensorFlow.org features books, courses, and videos to help users improve knowledge of MLand learn how to apply it to their projects. Additionally, we announced a new 4-course specialization
“TensorFlow: Data and Deployment”, which is now available on Coursera with
deeplearning.ai.
Trusted Partner Pilot Program
For enterprises, finding the right resources to implement ML solutions can be difficult. In an effort to help, we launched the
TensorFlow Trusted Partner Pilot Program. The TensorFlow Trusted Partner Pilot connects system integrators verified by the TensorFlow team with enterprises that are getting started with ML. Our current Trusted Partners include Accenture, Cognizant, Quantiphi, and Wipro.
TensorFlow 2.1 first release candidate now available
In November, we announced the RC of TensorFlow 2.1, which continues the momentum of TensorFlow 2.0 with major improvements and bug fixes. For more details, you can find the
release notes on GitHub. Within 2.1, TensorFlow now has Cloud TPU support. This means for high-performance training scenarios, you can now use the
Distribution Strategy API to distribute training with minimal code changes and attain great out-of-the box performance now on Cloud TPUs. Check out the distributed training
guide for more details.
2.1 also includes a pip package for GPU support by default on Linux and Windows machines with and without NVIDIA GPUs. Additionally, TensorFlow now has CUDA 10.1 support, and
experimental support for mixed precision is available on both GPU and Cloud TPUs.
TensorRT 6.0 is now supported and enabled by default which will add more support for TensorFlow ops.
Looking to 2020
With such a great year behind us, what’s next for TensorFlow in 2020? If you want to provide feedback, please fill out this
short form so that we can contact you about feature needs and relevant parts of the ecosystem.
We have more great events coming up and would love to meet you in-person! Remember to sign up for our monthly
newsletter to stay updated on product, events, and community updates. Important event dates are listed below:
From everyone on the TensorFlow team, thank you! From your cutting-edge research to your ambitious projects, and real-world applications the community continues to advance and improve TensorFlow for all. Without everyone’s contributions, our vision to inform, inspire, and empower ML users around the world would not be possible. We cannot wait to see how the global machine learning community improves TensorFlow in the years to come. Here’s to another great 4 years!