Placeholder Image

Subtitles section Play video

  • [MUSIC PLAYING]

  • EDD WILDER-JAMES: Hey, everybody.

  • How you doing?

  • Good.

  • Good.

  • Excellent.

  • That was an amazing set of demos, wasn't it?

  • So I work with TensorFlow, helping

  • build community and collaboration

  • around the open source project.

  • And usually, I put the thank you slide at the end,

  • for listening to me.

  • But actually, I want to thank you for contributing and being

  • part of the TensorFlow project.

  • Whether you're here in this room or on the livestream--

  • there's been some amazing talk on YouTube from people

  • in China, and India, and Japan, and all over the world joining

  • us virtually today--

  • thank you for your contributions.

  • The project is where it is today because of you.

  • Of course, in core TensorFlow alone, we've

  • had this many commits.

  • Now this figure is out of date-- over 50,000

  • from 1,800 contributors, and much more than just code

  • commits.

  • There's been 39,000 Stack Overflow

  • questions about TensorFlow.

  • We have 66 machine learning Google Developer Experts,

  • many of whom are here with us today.

  • So welcome, and thank you guys.

  • [APPLAUSE]

  • It's really great to have you with us.

  • And thank you for everything you do helping

  • teach people about TensorFlow.

  • And we've had 14 guest posts to the TensorFlow blog,

  • and that keeps going up.

  • There are so many ways that people are contributing.

  • Whether you're organizing a meetup, whether you're

  • teaching other people, whether you're speaking at conferences,

  • thank you.

  • You're really helping build out the TensorFlow ecosystem.

  • So in this talk, what I want to do

  • is discuss how we're growing the ecosystem

  • and report back on some of the changes

  • that we've made over the last year.

  • So I'm going to cover how we're making it easier

  • to get involved in TensorFlow.

  • How, also, we're trying to consult better

  • with the users in the community and be

  • more transparent about our development.

  • I'm going to cover how we're empowering everybody

  • to get involved, and to do more, and increasing

  • the number of contact points where you can

  • get involved in the project.

  • Finally, I'm going to go into a bit more depth

  • about the conference that was announced this morning,

  • the TensorFlow World.

  • So let's talk about how we're making contribution easier

  • to TensorFlow.

  • One of the most important things to help

  • people contribute to the project is increasing its modularity.

  • You heard Martin talk, this morning,

  • about the low-level APIs.

  • And with the move to TensorFlow 2.0,

  • we're trying to make it less of a monolith, both

  • in terms of code and in terms of people organization.

  • When you come and you want to contribute to an open source

  • project, it helps to be able to find where to contribute

  • and who to work with.

  • By splitting things out, we're creating more surface area

  • where it's easy to start building and creating

  • new projects.

  • And our special interest groups play a big part in this,

  • and I'll talk a bit more about them later.

  • But it's not just code.

  • There's so many more places to contribute this year, compared

  • to where we were last year.

  • So I'm going to talk briefly about our documentation

  • groups, the groups getting involved

  • in testing, people who are blogging, and on YouTube,

  • and more.

  • I was super excited to see, last week,

  • that we have published a TensorFlow tutorial now

  • in Korean.

  • And that's not a translation that we've done on our team,

  • but that has come from the community.

  • So thank you so much to Hasin Park for the Korean work.

  • Similarly, we're able, also, to publish it in Russian.

  • Thank you to Andrew Steppen.

  • This is just so exciting, to see that TensorFlow

  • is being taken to more areas around the world,

  • thanks to you.

  • I'm also really excited about the TensorFlow 2.0 testing

  • group.

  • Led by Paige Bailey, this is a bunch

  • of contributors and Google Developer

  • Experts who are working to give TensorFlow 2.0 a thorough test.

  • And you see, on the screen, an example of a friction log.

  • And so what's happening here is that folks

  • are going through ML workflows with TensorFlow 2.0,

  • documenting what they find delightful and awesome,

  • and also things that could be a little bit better.

  • If you'd like to join in this work,

  • this group meets weekly and often has guests talks

  • from maintainers, and SIG leaders, and so on,

  • and is really helping bring TensorFlow 2.0

  • from the cutting edge into something that is thoroughly

  • tested and ready for use.

  • Already mentioned, we have over 14 posts

  • from guests on the TensorFlow blog.

  • This is from a great post about realtime person segmentation

  • in the browser with TensorFlow.js.

  • It comes from a grad student and researcher and ITP.

  • So whether it's testing, whether it's documentation,

  • whether it's blogs and conference talks, thank you.

  • Now I want to talk a little bit about TensorFlow RFCs.

  • As you probably know, RFC means Request For Comments.

  • This time last year, we weren't that organized

  • about how we evolved TensorFlow's design,

  • in terms of communicating it.

  • And I stood on this stage and told you

  • about how we were going to launch the RFC process.

  • Well, now we've accepted 21 RFCs over the period

  • of the last year.

  • This is our key way to communicate design, where

  • before code gets landed in the project,

  • we post an RFC about the design and consult widely.

  • This isn't just about code that's

  • coming in from the TensorFlow core team outwards.

  • They can be created and commented on by anyone.

  • We've had several RFCs that come from the broader community.

  • And I expect to see so many more of those in the future.

  • We have several, for instance, from. the SIG groups already.

  • One of the things I'm most proud about

  • is how the RFC process is underpinning

  • the 2.0 transition.

  • This was mentioned earlier, but all the major changes

  • in TensorFlow 2.0 have been proposed and consulted

  • with in RFCs.

  • This isn't just a great way of consulting and getting

  • information feedback.

  • Going forward, you now have a big repository

  • of technical documentation about why design choices were made

  • a certain way in TensorFlow.

  • And it's a great educational resource, as well,

  • for people who are coming on and want to get involved

  • in contributing to the project.

  • So I really want to give a big thanks

  • to anyone who has authored or reviewed an RFC.

  • You've played a vital role in making TensorFlow better.

  • Now let's talk a bit about the social structure of TensorFlow.

  • Last year I talked about how coming to a large project

  • can be a little bit daunting.

  • You don't know where people are, where

  • the people that have your interests in common are.

  • And so we created the Special Interest Groups, or SIGs,

  • as a way of organizing our work.

  • There are so many uses of TensorFlow,

  • so many environments, so many architectures.

  • And many of them are outside of the scope

  • that the core team can resource.

  • And what we wanted to do was enable TensorFlow

  • to grow and be more sustainable by creating

  • a way for like-minded people to collaborate

  • around well-defined projects.

  • So this is why SIGs exist.

  • They're groups of people who are working together

  • for a defined project focus.

  • We started last year with SIG Build,

  • and now we have six of them up and running.

  • I'm going to give you a quick state of the SIGs.

  • Many-- in fact, most-- of all the SIG leaders

  • are here with us today as well, so I'll

  • give a shout-out to them.

  • And hopefully, you'll also be able to talk to them

  • in the lunch and tomorrow.

  • So SIG Addons first--

  • thank you to Shaun Morgan and Amanda Fandango

  • for leading this group.

  • Martin mentioned, at the beginning of the day,

  • that tf.contrib is no longer a part of TensorFlow

  • going into TensorFlow 2.0.

  • And SIG Addons is a place where a lot of that code is going.

  • So these are parts of TensorFlow that don't fall into the core,

  • but do conform to these well-defined APIs--

  • so more losses, ops, layers, and so on.

  • Now there's already an RFC published

  • about where you can find things that you

  • used to find in contrib that have gone into addons.

  • And addons are also going to publish another RFC real soon

  • to say, well, how can you get involved,

  • if you have your favorite app or whatever,

  • that you want to step up and be a maintainer,

  • and maintain it for everybody, how

  • you can join in the project.

  • So I'd encourage you to take a look at that.

  • SIG Build-- SIG Build really is where TensorFlow

  • meets the outside world.

  • And it's not always the most glamorous piece of work,

  • but building TensorFlow, and packaging it,

  • and distributing it is tough.

  • And so thank you so much to Jason Ziman and Austin

  • Anderson, who lead that SIG.

  • SIG Build has achieved a lot in the last few months.

  • One thing, it's the home for third-party

  • contributed builds for architectures

  • that we don't ship out as part of core--

  • so IBM Power, Intel MKL optimized builds.

  • And SIG Build works on improving the TensorFlow build

  • and helps us be a better neighbor in the Python

  • ecosystem as well.

  • As you can imagine, machine learning

  • generates a lot of extreme situations

  • that need changes in ways we evolve in packaging

  • and distributing software.

  • SIG IO is a fantastic group that helps connect TensorFlow

  • to other systems.

  • Out in the real world, your data exists somewhere.

  • You're using other systems in other formats.

  • So this group is led by Yong Tang and Anton Dimitriev.

  • IO really ships support for extra file systems, extra file

  • formats.

  • So if you're using any of these things in the Apache ecosystem

  • or any of these file formats, you

  • can use the SIG Addons module to use that data in TensorFlow.

  • This group is prolific.

  • They've already dropped four releases.

  • Last week, they just created their 0.4 release.

  • And they also ship R integration with their module too.

  • SIG Networking is where a lot of the alternative networking

  • schemes that are available in contrib are going to.

  • This is led by Byron Yi and Jeron Bedoff.

  • So if you're using GDR, VERBS, MPI,

  • this is where you can find that.

  • SIG Rust, led by Adam Crume, is developing idiomatic language

  • bindings for the Rust programming language.

  • If you're interested in this, or any of the other SIGs,

  • please talk to the leaders.

  • They really do want more help.

  • And now they're up and running, they're in a great place

  • to bring people on.

  • If you're looking for a way to get

  • involved in contributing to TensorFlow,

  • this is an ideal one.

  • Finally, let me touch on SIG TensorBoard.

  • We've rebooted SIG TensorBoard this year

  • to really work closely with the community.

  • And so this is a great time to be involved, as the TensorBoard

  • team are starting to consult and figuring out how we can best

  • enable people who are using TensorBoard,

  • both in terms of creating plugins or using it at scale.

  • If you go to the demo area above and go to the TensorBoard

  • stand, you'll find Mani and Gal there,

  • who will be happy to talk to you.

  • So this is the URL for anything you

  • want to do with joining the TensorFlow