Placeholder Image

字幕列表 影片播放

  • ♪ (music) ♪

  • Hi, everybody, and welcome to TensorFlow Meets.

  • It's my great honor today to be chatting with Megan Kacholia,

  • Engineering Director, right? on TensorFlow.

  • So can you tell us a little bit more about what it is that you do?

  • Sure. I'm an Engineering Director on the TensorFlow team.

  • I've been with the team for a little over two-and-a-half years now.

  • I've gotten to participate in all of the Dev Summits

  • - which has been really exciting. - Nice.

  • Just to get to come, be part of the keynote, be part of the event.

  • In terms of what I actually do,

  • I'm one of the leads for the TensorFlow team.

  • I just help with, obviously, the team itself,

  • make sure we're going in the right direction,

  • and make sure that we're doing the right things by our community.

  • One of the things that you spoke about

  • that I saw that a lot of people are excited about

  • and I know I'm personally excited about

  • is just really how we're doubling down on making it very developer-friendly

  • - in TensorFlow. - Yes.

  • How do that come about, can you tell us a bit more about it?

  • Obviously we've taken a lot of feedback from the community,

  • just to see what are the pain points,

  • what works well, what doesn't, I think that's a big part of it.

  • And just in general TensorFlow has been around--

  • We just had our third birthday in November, right?

  • So, things change, or the industry is really moving quickly,

  • machine learning is advancing in lots of different places

  • than we might have anticipated when TensorFlow was first developed.

  • So we need to make sure that we're making it very, very easy

  • for folks to just come in, get started, and be able to take advantage

  • of all of the cool things happening

  • - in the machine learning space. - Right.

  • I really think it comes from kind of both of those angles,

  • just how things have moved so fast,

  • and just the feedback we've gotten from the community

  • as we've had TensorFlow out there for a few years.

  • I've been working on it for about a year and a half

  • and I came in with a developer background

  • and when I started trying to kick the tires on TensorFlow,

  • there were some things where it was like...

  • it just wasn't quite intuitive to me.

  • But a lot of that has been changing, right?

  • There's stuff like Keras

  • - and other things, - That's correct.

  • and the eager execution by default that you've been adding

  • just hopefully will make it a lot easier for other developers.

  • (Megan) That's correct. We want to make sure, obviously,

  • that the flexibility and the power of TensorFlow is there,

  • but also make sure it's approachable and easy for people to just come in

  • and be like, it's fine, start here on this surface,

  • and if you need to dive down to this part

  • and really get into the guts of it, you can, it's fantastic,

  • but you don't have to if you don't need to.

  • (Laurence) Right. And all of this doesn't come

  • at the cost of performance, right?

  • Because you had that great slide with performance and...

  • Can you tell us a little bit about the performance improvements

  • that you've been seeing with TensorFlow?

  • (Megan) So a lot of those improvements as well

  • while we're looking at just kind of the core TensorFlow, right,

  • kind of you think of like the heart of it,

  • and that's where the majority of the improvements come,

  • meaning that those improvements will be applicable

  • whether we're talking about TensorFlow 2.0

  • or whether we're talking about using Keras or something else.

  • It's that big meaty part kind of under the hood.

  • And a lot of it comes

  • from just better making use of the different types of hardware,

  • making sure that we're using the different types of accelerators appropriately,

  • understanding the limitations and restrictions for things like mobile.

  • We talked about a lot of improvements for performance on TF Lite as well,

  • and again, some of it is just understanding

  • what do the [workflows] really look like,

  • what does it look like on some of these different devices

  • and Edge TPUs and things that we're trying in TensorFlow out more with now,

  • and then kind of always going back and closing the loop and being like,

  • okay, this is how we can make it better and this is how we can make sure

  • that the users get that performance by default,

  • and don't have to necessarily know

  • what magic had to happen under the covers for it to happen,

  • they should just get it.

  • (Laurence) Right. Now, you mentioned a mobile with TensorFlow Lite.

  • We lovingly call it TF Lite.

  • One of the things that I do get a lot of questions about

  • is that there's all of these different almost parts of the family of TensorFlow,

  • and can you give us a quick summary of the different runtimes

  • that are available for developers,

  • between TensorFlow Lite and, what are those other?

  • (Megan) So there's TensorFlow Lite,

  • there's more higher level types of things

  • that we talked about with TensorFlow Extended,

  • so that's not neccessarily a different runtime

  • but it's more kind of just how you put things together end to end,

  • especially if you're looking more towards a production environment.

  • We talked a bit about JavaScript as well, so TensorFlow.js,

  • and I think that one, it's a language-type thing,

  • because you want to make sure that the JavaScript community--

  • it's such a large community-- that they have access

  • to machine learning as well,

  • but there's also just the whole in-browser experience

  • that it plugs in well with because of Node.js.

  • So I think a lot of it depends on what kind of applications

  • people are interested in

  • and kind of where they're coming at machine learning from.

  • There are so many different ways of applying machine learning,

  • whether you're thinking of it like, "Oh, I work in a large company,

  • so I need it for these enterprise use cases,"

  • or "I'm just an app developer, I'm trying things out myself,"

  • or "I'm someone who's really familiar with this language,

  • I'm really familiar with JavaScript,

  • so that means I can use that as kind of my entryway

  • to start doing other things in machine learning."

  • So some of it depends on where you're at,

  • where you're coming from,

  • and then we try and make sure we have the right way for you

  • to get into the machine learning community.

  • The JavaScript stuff is amazing in the fact

  • that you can actually train models in the browser.

  • (Megan) Yes.

  • When I first heard about that, I was like "Nah, come on, really?"

  • - But you actually can do that, so... - (Megan) Yes, it's really cool.

  • And as you mentioned that with Node you're not just limited to in the browser,

  • you can also run it on backend servers and that kind of stuff.

  • One of my personal favorites actually, is with things like Cloud Functions

  • or Cloud Functions for Firebase because they run Node.js,

  • you can actually start putting models in there.

  • Now, TensorFlow 2.0 is currently an alpha release, right?

  • Now, lots of people are asking what comes next,

  • when are we going to see a release candidate or something like that?

  • Yeah, so like we talked about at the Dev Summit,

  • we're going to have our release candidate kind of coming up sometime in Q2.

  • We want to make sure that we're giving ourselves

  • enough time to make sure we have good performance,

  • make sure we have the right fit, finish, and polish for everything,

  • but there is, like you said, an alpha available.

  • We want feedback from the community.

  • We want to understand what are people enjoying,

  • what's working well, what are they concerned about.

  • And you can also follow along on GitHub

  • and kind of see what features are coming out,

  • what status are things at.

  • That way the community knows what's going on

  • and we can make sure that we're engaged appropriately

  • to be building the right things

  • as we finalize and finish up with the 2.0 release.

  • Sound perfect. Thank you!

  • So people can stay tuned to GitHub and keep track of everything

  • - that's going on. - (Megan) Yeah, that's correct.

  • Thank you so much, Megan.

  • And thanks, everybody, for watching this episode of TensorFlow Meets.

  • If you got any questions for me or any questions for Megan,

  • just please leave them in the comments below,

  • and check out the GitHub page for TensorFlow

  • if you want to learn more about the TensorFlow 2.0 release.

  • Thank you so much.

  • ♪ (music) ♪

♪ (music) ♪

字幕與單字

單字即點即查 點擊單字可以查詢單字解釋

A2 初級

TensorFlow是如何不斷改進的(TensorFlow Meet (How TensorFlow keeps improving (TensorFlow Meets))

  • 2 0
    林宜悉 發佈於 2021 年 01 月 14 日
影片單字