Name: TensorFlow--移動平臺的深度學習解決方案（TensorFlow遇見）。 (TensorFlow - the deep learning solution for mobile platforms (TensorFlow Meets))
Uploaded: 2021-01-14T10:33:24.000Z
Duration: 8 min 10 s
Description: 【看影片學英語】數萬部 YouTube 影片，搭配英漢字典即點即查，輕鬆掌握單字發音與用法，長久累積看電影不必再看字幕。

We're here with the TENSORFLOW Developers Summit Have Intensive Flow Cafe and it's my owner right now to meet with Pete.

Awarded on Pete is the tech lead for tensorflow on Mobile, which spends a whole lot of things.

Oh, there's so much interesting stuff that people are doing.

So here now, at the summit, we've been showing a lot of mobile stuff, including TF Lights, Yes, yes.

So what's new and exciting into your flight?

So over the last year, you know, tensorflow on mobile, taking the whole of the tensorflow fame work on fitting it down to fit our mobile has been really useful for a lot of people.

On dhe, we've done some seen, some really interesting applications of people built.

But what we've wanted to do is build something that's even smaller, even faster and even easier to use.

Its mission is to be that fame work that people are looking for.

We've been seeing some really cool stuff done.

A mobile like the cassava demo in the kino.

Yes, yes way We're just chatting off camera.

It's very hard to convey when we all have great connectivity.

Why do you need it to be isolated on a mobile?

But that example I mean, yeah, how did that work?

So we actually have been working with part village for think up to a couple of years now on Dhe.

We've been working with them because they had such interesting problems, like they were trying to get stuff into people's hands who didn't have data connections on.

Also, even if they had data connections, they wanted answers fast.

So getting that interactivity and allowing people to actually move the camera around and focus it on the leaves and kind of get that instant feedback makes a massive difference for them application.

And we've seen that across a whole bunch of applications.

It's not just oh, I don't have a date a connection.

It's like I need to work with video audio on.

I need that instant interactivity that you can only get by running on device makes sense.

And yet, so you got that speed video might be a big uploads of the oh, I see where you're going, That's actually pretty cool beyond, like, mobile, like android and IOS.

Now you have t f light on tensorflow light coming on raspberry pi way Saw cools M o of yes, and Andy has been doing amazing work with the raspberry pi and actually getting the speed that we've got on Android and IOS actually ported over to these wonderful sort of $25 devices on these things that you can put almost anywhere.

And do I think we were talking earlier about donkey cart?

Yeah, which is this fantastic example of using tensorflow running on a raspberry pi to power a self driving toy racing car.

And if you've seen the videos, their incredible how they I actually bought When I'm building it myself, I don't know what it's gonna be like driving it home, but we'll see actual self driving car even though it's a toy.

Think about the compute power of a raspberry pi.

Now you could build a self driving car thanks to and the fact that this could be something that almost anybody can actually buy and build themselves now for just you know, it's so much fun seeing this Gettinto.

You know, I have a lot of kids actually playing around with this stuff.

A lot of high school students as well, building all of these crazy sort of, you know, trash sorters using last week pies.

You know, all of these other, like, really, really interesting projects You can only build if you have this cheap, almost disposable computer that you can play with.

But when you can take a model that was typically only restricted, toe high ends the offensive supercomputers.

Nowadays, with tensorflow, you can train that on your desktop machine.

You can flatten it, and you can deploy it quickly.

Thio an android or in IOS are now a raspberry pi.

One thing I would like to give a shout out to is actually tensorflow for poets, the Etoile, which doesn't require any coding, and we actually have the ability to very easily get that onto IOS, Android or the last three pie.

So if you don't feel confident in coding at all, you can still create your own image model.

Go through the code lab on it on a laptop with nog.

You know, for many without what that does is it takes the last layer and it's and it uses right for image classification Mobile net.

Yes, it takes the last layer of that on in the Poets Code lab.

So instead of classifying general emergency training on a bunch of flowers, recognizes those flowers.

And like you said, code Leslie, you could do it way.

Generally having a conference is like Google i o.

And I think it was the most popular code lab, but I owe last year.

It's super cool factor Raspberry pi for a second.

So, like, you know, the damage that we showed earlier.

Wrong was tensorflow light running on a raspberry.

Yes, which was inference only, but people will also want to be able to train on something like a raspberry pi.

So the vast with pies, this really interesting beast because it's half like a limericks desktop, and it's half like a mobile phone.

So when all you want to do is run influence Rana model.

But if you actually want to use Python on, do you want to program and use the full programming environment of tensorflow Do training of models?

You can actually get binary Pippen stools on.

Unfortunately, we don't have it up on the tense photo or website.

Put a Lincoln Way will add a link that you can actually just do a Pippen stool and get it on your arm.

Because to be able to do that on a small embedded system because they may be scenarios like we're talking about earlier where you don't want to Yes, on especially doing things like personalization or modifying, according to data that you don't want to leave the device right.

That could be a really interesting application.

So from a developer on, I want to get started and enjoying all this magic other than the tensorflow for poets code lab, you have any recommendations that's That's a really good question.

I would actually look through the examples that are in tensorflow on tensorflow Light.

You should be able to get up and running pretty easily and X Code or Android studio on really play around with them on dhe.

Start to get ideas on, learn to modify the code that's already dead so much easier to start with code that's already working on trying to build something from scratch.

So I actually put the demo on my my pixel and I put the android demo on here.

Going around and classifying on on the speed is actually really good because I can point it like, for example, at a mug.

And it's generally less than 100 milliseconds.

And with the North neural network, a p I coming up in Android Yeah, Android P.

Well, we actually have a preview of it out already.

You should see even better performance, especially using the Qualcomm hdx accelerator, but in your pixel.

Look out for maybe I'll have to put a P preview on here.

Thank you so much, Pete, that it's been a blast.