Placeholder Image

Subtitles section Play video

  • Hello, everybody.

  • This is Laurence Maroney.

  • I'm here at the TENSORFLOW.

  • Developer Summit.

  • We're in the Tensorflow Cafe.

  • I'll have Jeremiah with me all the way from Zurich.

  • That's right.

  • It's been an exciting trip.

  • So all the way from zero Can you announced TF hub TENSORFLOW.

  • Huh?

  • But your talk.

  • So tell me all about it.

  • Tensorflow Hub is something we're really excited about.

  • We realized that we want to help people reuse machine Learning machine is really hard to do.

  • You need a lot of stuff.

  • So we're letting people package up good machine learning and share it with the world.

  • Okay, Now, you felt this in Zurich Way did we did.

  • This is our first big infrastructure project there.

  • We've been growing Google Research Europe.

  • We have a brain team there that's been grilling and are applied machine intelligence team eyes also growing quick.

  • Nice.

  • Nice now so that this tea of hope, it's almost like a natural extension to some of the things we've been doing on.

  • Get help.

  • Right where we have, like, model repositories, get help.

  • But you're taking a lot for right, right?

  • Certainly, sharing models isn't anything new.

  • The big difference here is that we realize that when you share a model, either fits your problem exactly.

  • Okay?

  • Or it doesn't, and you can't use it.

  • Okay?

  • You can either give it the inputs it wants and take the outputs.

  • If you want to do anything outside that, you're sunk so tensor full hub because you're smaller pieces, so it's more likely you can reuse them.

  • Think the way to think about it is a model is like a binary.

  • Okay, Module is like a library.

  • Okay, Model is a binary, and a module is a library.

  • That's right.

  • We need that on a T shirt.

  • That's right.

  • That's right.

  • So where can I find all these?

  • Where is tense?

  • Where does that leave tensorflow?

  • Check out tensorflow dot org's slash hub tensorflow.

  • That's easy to remember.

  • What kind of models are on their Well, we've got all kinds of things we certainly have once for image processing things that will let you help our help you build your own image.

  • Classifier.

  • Ziff, you've used tensorflow for poets.

  • We've really taken that process and streamlined it down to really a single line of python.

  • Um, so we've got all kinds of, uh, kind of models.

  • Some of the newer ones include the neural architecture search models that are really state of the art.

  • As Andrew mentioned in his talk.

  • Those are incredible because with that one line of python, you actually get over 60,000 hours of GPU training.

  • When in discovering the architecture er and training, it was pretty powerful.

  • So one line give me 60,000 hours worth of training 60,000 GPU hours, yet is the image ones.

  • We've also got some really exciting text ones where you can give it the entire sentence.

  • It will give you back of Vector that characterizes it.

  • And again we could do that same tensorflow proponents trick where we build a classifier on top of it on, but the ones we're really excited about eyes the universal sentence encoder.

  • This is a paper that just came out on archive last night, so it's just hot off the presses, and it's great to see people that can use it instantly just by clicking on the public that's in the paper.

  • So what is it all about?

  • The universal sentence and go to.

  • So this is some new work that is doing things like letting you understand the semantic differences between different sentences.

  • It actually works on a sentence level, which is a little bit different than words to back.

  • Ah, lot of those.

  • You just look up a single word.

  • But this will take the entire sentence.

  • And consider that as it builds this characterization in the form of embedding.

  • Remember always that.

  • What was it?

  • The fruit flies like bananas.

  • Have you heard of that?

  • That's right.

  • That's right.

  • That's the exact problem solving.

  • Oh, cool.

  • I have to look that up.

  • So what was it called again?

  • That's the universal sentence encoder, and that's on T f.

  • I've already it's they're ready for you to use it.

  • Oh, wow.

  • Gonna check it out.

  • So now if I'm like a specialist in building models in some scenario, I don't know what the scenario is and I want to contribute.

  • How would I go about doing that?

  • Yeah, we're really excited.

  • We're working very quickly to make it possible for other people to upload and share these things.

  • I think that's something that's really important to us, is certainly as talented as we are inside Google Answer was really a global community under the wisdom of crowds, right?

  • This is a lot of people who know stuff that we don't know when they can.

  • That's right.

  • That's right.

  • Cool.

  • So thank you so much.

  • This has been so much fun, and I've learned so much already.

  • So I really want to try some of these models that you told me about.

  • So thank you, Jeremiah.

  • That's been fun.

  • And thanks everybody for watching this episode.

  • If you've got any questions for me, if you've got any questions for Jeremiah, please leave him in the comments below on.

  • If you're looking for any of the lengths we spoke about, we'll put them in description on whatever you do.

  • Don't forget to hit that subscribe on.

  • Thank you so much.

Hello, everybody.

Subtitles and vocabulary

Operation of videos Adjust the video here to display the subtitles

A2 jeremiah sentence machine encoder machine learning model

TensorFlow Hub: reusing machine learning modules (TensorFlow Meets)

  • 0 0
    林宜悉 posted on 2020/03/25
Video vocabulary