About the talk
Are you enthusiastic about sharing your knowledge with your community? InfoQ.com is looking for part-time news writers with experience in artificial intelligence or machine learning.
Earn money! Build recognition!
To learn more, contact email@example.com with a link to your LinkedIn profile.
Magnus Hyttsten talks about TensorFlow and gives an overview of the different products & APIs and the best practice.
Magnus Hyttsten is a Developer Advocate for TensorFlow at Google.
This video was recorded at QCon.ai 2018: https://bit.ly/2piRtLl
More videos from QCon.ai 2018 on InfoQ: https://bit.ly/2rNAT8z
The InfoQ Architects' Newsletter is your monthly guide to all the topics, technologies and techniques that every professional or aspiring software architect needs to know about. Over 200,000 software architects, team leads, CTOs are subscribed to it. Sign up here: https://bit.ly/2KqYfrs
Excellent. So we have a 10-minute lightning talk. I would say here 10 minutes. If not the lot of time to cover tensorflow. And after that they have a 00:04 10-minute break. And after that I'm going to do a 15-minute huge feet on TP use and what is machine learning and how do we use machine learning in to 00:11 Google then? Also, how do we do this tribute to tensorflow? So if you're really interested in understanding more than nuts and bolts compensatory 00:20 should definitely catch that section because right now or just going to focus on getting you started the resources that you can use. So what is 00:28
tensorflow About 10% So I hope it's going to be beefed up a bit after today session 00:37 open source machine learning library blah blah blah blah blah section is a machine-learning Highbury that we use at school before all 00:46 machine learning purposes and is exactly the same thing we have open sores as we have inside of Google. There is no difference people think that we're 00:56 keeping something Secret inside a Google and we can only use that thing. But if that's not true that actually built on the same kind of infrastructure 01:04
is only one single difference and that's at the RPC protocol between processes is a big difference in a Google set data center as opposed to just 01:11 today or Center outside. So the RPC protocol is a bit different. That's the only thing no magic sauce from a machine learning perspective. Tensorflow 01:19 is used how I built it to be usable for both research as far as production environment. There's rigorous testing going on between each release of 01:27 the chance of snow and we put out the new release approximately every 6 to 8 weeks we have about the Thousand contributor. So it's not a Google thing. 01:37
It's actually an open-source everyone can participate in this project. And also the one thing that we've recognized is that I'll muesli training on 01:47 TCP UDP you and TPU that we'll talk about more in the next session is really important, but when it comes to the employment, we gotta support the 01:56 smaller device might as well Android iOS and also Raspberry Pi it's all released under Apache 2.0. So it's a very open 02:05 we are open source disk back in 2014 to stall as you see here in a number machine learning models in Google slowly grew and then when 02:13
we really sensitive, I'm just lie there is this hockey stick. Kind of traction up words every single service your son. Should I use from Google 02:23 nowadays is built a small dose machine learning models which are built on tensorflow. And if you've heard Google is now an AI first company, we used 02:32 to be mobile first company an hour and a half her company. So this is very very important to us. So therefore tensorflow as a lot of investment going 02:40 on in making this work actually just one minute talk about what machine learning is. So we have a social experiment that's going to be very awkward 02:48
for me. Hopefully not so awkward for you if I say it ends with o An awkward silence, right because you have no idea what you were going to respond to 02:58 that. Correct. Very good at what we actually did. This is an introduction to machine learning side said than example to you 03:07 which was the training data supposed to say tensorflow here today and potatoes * 55.2 your model you responded with nothing. Right? And that was 03:17 awkward for me. And then I show the next Picture Rocks and now you know, you take the loss function between nothing and Roxy said, oh there was a 03:27
mistake I should have responded with rocks instead. So the next time you have to put out roxtec, so everyone get that what kind of have something we 03:34 feed into the model and then we get a right answer and then we adapt a model to be able to go to work Kool. How do I get started with that? One of the 03:43 best thing is to go to ten Sapporo get started here. You have all of the stuff. We recently converted to use proposition 2 03:51 days. So, I don't know if I should do this to use collab so you can actually go in here and if you 04:01
Go into town so far or develop getting started with eager execution. Are you can see these things and then you can click on these notebook so you can 04:12 actually get an executable notebook directly from the website. So there is no installation requirements know nothing we're converting all of the 04:22 things to comply to this kind of format. So it's really easy to get started. Cool. The other thing you should know is that tensorflow is built on 04:29 many different layers in the end. It's architecture the hardware the distributor execution engine the python front end when you decode tensorflow 04:39
nowadays, you should use one of these high-level apis and that's exactly what this getting started experience. It's also teaching you so that's really 04:48 important because there are many ways to configure in machine learning work clothes sometimes Advanced research that I need to go to this one down to 04:55 this level, but that's definitely not a place where you should start to build a simple example and carrots you put a sequential only had a couple of 05:02 lawyers and they can train and 5th and evaluate this model and then you can build his kind of cool a classifier to can take pixel inputs and then 05:11
through these weights and layers determine if it's a cat or a dog training is one thing done deployment. That's another thing and to deploy on these 05:19 mobile devices. We have a component of tensorflow called tensorflow Lite. Why is it called light because these mobile devices 05:28 they are not as big as their traditional desktops or whatever data center computers. He we might use for training because that takes a lot of power. 05:38 So tell if light is a very optimize library for embedded devices and mobile devices easy to use of course, everything is his to use right 05:45
it's fast and is really really small. It's really complex to use. I would be weird if I had that and the way that it works if that you do your 05:55 classical training using thousands of computers and TVs that we'll talk about the next session and then you'll train your thing and when you're 06:05 finished with your mother, you convert it into a light format and that form a done on the mobile device through and interpret that gets loaded and he 06:13 can use to make use of all of the hardware accelerators that are available on the mobile device on Android that will be 06:21
developed in the future and on iOS Okay, it has the capability of going 06:27 through Cora now. That's one of the options you can also ask you could all the operations on the CPU. Of course, if you're not a developer who is not 06:37 a developer. I thought so it's cute. But anyhow, I mean should you not want to code for some reason one day we have all of these cool Services 06:44 also in the cloud will you can upload videos and they can automatically detect stuff for you and your sex is pretty cool for developers Fallout 06:54
because you can get this kind of messed up tagging immediately from did you start your clothes even though you have no idea what a video contains if 07:00 your uploaded to the cloud you can actually get actionable information on how to process that video or talk to you sir with respect to that 07:07 children quotes, whatever how you upload 07:15 just images to this thing and using the latest technologies will look at some of these models in the next week. Are you actually get a 07:24
state-of-the-art machine learning models directly without any Home machine learning whatsoever. Which is kind of boring because machine learning is 07:31 quite fun. So so as I said, we got one minute left get started go to tech support orange and get underscore 07:39 started getting started with eager execution. You can open up the cold out there working on making bee stings Cole s at 12, so they should appear it's 07:49 cold outside right now. Unfortunately, they're HTML, but we're working hard on fixing. Thank you so much. 07:56
Buy this talk
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.