字幕表 動画を再生する 英語字幕をプリント hi and welcome to ask tensorflow the show where you get to ask questions about Tensorflow and we will try to answer them. My name is Magnus. It's done. I work as a developer advocate on the Tensorflow team, and I'm Laurence Maroney, also a developer advocate on the team. So let's get started. The first question is from Bern, versed on Twitter. And the question is, how about advice on a suitable data set for a learning experience on Tensorflow? That's a great question. Yes, it is. There's a few that are grateful. Starting there's always CAG. Oh, who has slots off open data sets that you can use. So if you don't have a use for a count of cattle, you should definitely check it out. Also, models that are bundled will tensorflow. You can find these in the Tensorflow flash model skit Hub Directory, and many of them are from well known a. I stash ml problems, including things like the M NUS data set, which is really the hello world or machine learning thio that allows you to help classify handwritten digits. It could even detect my handwriting that's doing something. There is also a tutorial called wide and deep that used the census data from the UC Irvine machine learning repositories. Indeed, you can find lots of data sets in that repositories to play with. I actually found one for a tutorial that I was writing, and it was from the West Wisconsin breast cancer research data, and that was in the UC Irvine machine learning repositories. So I wrote a tutorial around that, and you can see it in the links underneath. So the next question comes from that Kali on Twitter. I'm not office. Can I retrain a model to recognize an object of something else, for example, if it sees a sunflower, and I want to classify that as a weed instead? So this all falls into something that we called retraining on. The concept of retraining removes the last layer, which contains the class labels on allows you to relearn it with your own labels. Now it's too long for me to explain it here, but fortunately, we have something called the Tensorflow for Poets Code lab. I've put a link in the description below. In that cold lab, you'll take a look at the standard mobile neck cold lab and retrain it specifically for flowers and strongly recommend looking into that and having a play with it. Now it's a very common technique. In fact, you can train a model quite easily with not so many example using this kind of methodology, which speeds things up to right now. And the next question is from Carmen Twitter on The question is, what's your best advice for a computer science student to begin in a i slash m o the top three things to do. Top three. Hard to think of just three, right? Yeah, but three it is. Well, I think a clear number one is to learn Python not just for ML, but also because it's a stag, a legal and useful language. I can't believe the things I've been able to achieve our since I learned, and it's used everywhere in machine learning and data science. So my number two after I've learned Python would be toe studies. Some of the math that's used in ML on it might be the girl who don't wanna learn math, but it's really not that bad. And the two parts of math that I really look into Number one would be regression. So regression is all around fitting a line two points into dimension or plain two points and three dimensions, as well as getting into higher order dimensions. And then the other one is matrices and understanding how matrices fit the picture for handling this higher order dimensional math. So how to multiply and transpose matrices and all that kind of thing about number three Number three. My number three would then be to practice on simple examples in there, do some great data sets out there. As we mentioned at the top of the show, you should download them, play with them and a spreadsheet and do correlations of stuff. And then he should load them into a machine learning model to see if you can do some classifications and above all, have fun with it on. Dhe is so money meeting things you can do with even quite simple models, eh? So you should really go out there and just try things out. And if you love what you do, you'll do what you love. Yeah, right. So next question, this one comes from Mitzi Key in Tokyo. Um, it's key us at Io. Last year, we heard about tp use. When and how will they be available while we gotta meet Sookie? That's a fabulous question and fortune. It's really easy to answer because they're now available for you to play with in beta. There are details of the U. R L at the block post that I've linked underneath now with some of the things we've have. There's some really, really impressive power available. So the teepee is that that are available on the cloud platform will pack up to 180 teraflops of floating point performance and about 64 gigabytes of high bandwith memory on a single board 180 teraflops tariff tariff flops. That's a that's a giga Giga far right. That's a lot of flops. So we actually did a test on one of these on a trained, resonant 50 in less than a day on when we actually posted the first blood post. About this, we said it costs less than $200 in total. We've actually been doing retraining since then, and it's come closer to $100 in total to train a complete resident 50. So how does this work? Well, TPU usage is paid for by the second. So the faster you go effectively, the cheaper it is on. We have a tutorial about using T. P. Used to train residents. That's at a lake underneath. I'd love to hear what you're planning to do in the type of machine learning you're going to be. So thanks, everybody again, a lot of really, really fun questions. Some great stuff that we've been able to look into would hope we've been ableto help you with answers to these questions. I'm Lawrence and this is Magnus. And remember, if you've got any more questions for us, please leave them in the comments below or better still, posting on social media with the hashtag tensorflow, we'll be here to answer them for you. Thank you so much.