We're here at the Tensorflow Developer Summit at the Tensorflow Cafe.

And I'm chatting with Andrew Gasparovic from Tensorflow Hub on you announce Tensorflow hub today.

Can you tell us all about it?

Yes.

Oh, Tensorflow Hub is the new way to share what we're calling modules, which is meant to be a reusable component.

So it's a little bit less than a model.

It's just the reusable portion.

But it is the graph, the weights and potentially any assets that come with it.

So it's all packaged up into a component that makes it easy to share in just one line of code.

One line of code.

Yeah.

Yeah, that was a requirement from the beginning.

Actually, no more than one line of code.

Where do you go from here?

Yeah.

Yeah, Well, I guess the natural 10 But that's next year.

Now.

One things you mentioned in your talk was about these modules being compose herbal, reusable, re trainable.

Yes, Tapestry.

That composer.

Yes.

Oh, compose herbal just means that, um, you can do things like add your own classifications.

So if you're talking about an image model module, um, it doesn't include the classification from the model.

So that is something that you can compose with what?

Your building in the case of text classification that maybe something like an embedding module.

And in other cases, we also wanna have just general purpose modules that you can compose around almost like functions that you can call together and make building blocks for a new entire model.

I say.

So if I have a model that I don't know, for example, that doing OCR and an image on.

But it can also tag things within that image.

What sayings?

I could break out the OCR part.

I could break out the tagging part, compose them into something new.

Yeah, absolutely absolutely.

And that's gonna be something really interesting to see what develops in the community over time.

Just the ways of putting things together and sort of modules.

People end up.

Yeah, so it's We're providing building all the communities, providing building blocks, todo eso we have.

We have a number of modules to start with, and they're very general purpose things doing, you know, image classifications, doing text classification.

We have those inventing modules, but I think that over time, what the community can contribute is what really?

It will be interesting.

Sounds great.

So we've mentioned that there compose a ble and then reusable reusable just means that you can basically take something that already exists and apply it to your own particular problem.

Makes perfect sense.

And then I'm really interested in re trainable way.

Speak a lot about the tensorflow for poets where you can retrain the final layer toe, make it flowers instead of genuine images.

Is it too?

Do a similar thing.

Go.

Do you hear that?

Yeah, definitely.

So there is, You know, the classic sort of transfer learning case there where you would take an image classifications module just up to the feature vectors and then retrain your own classifications on top to do things.

But you know, you can also go deeper than that.

You can start retraining, text embedding modules.

And the thing about TF hub is that you can actually do fine tuning and retraining inside the module itself.

So if you have enough data, you can actually turn on retraining inside the module because it's just a graph with the weights.

So, um, you know you can really get better results because it's not just something that static.

It's something that you can really include in your own model.

Sounds getting into your talk.

You showed this for, like, classifying rabbits.

Yes.

Yeah, you had that one picture of a rabbit, which I think is the best picture of and in the entire day.

So we have to cut.

So that Joe, it's It's a beautiful Oh, yeah, Yeah.

Tell us what?

What did you do in that?

Damn it.

In the demo, we basically just It's the same ideas the tensorflow.

For poets, we start with a general purpose image classifications model just up to the feature vectors.

And in the particular demo example, we're classifying rabbits, maybe including the Easter Bunny.

And, uh, you know, we added our own classifications on top, and we fed in all of our own training examples.

And the end result is that you get something that is special to your task, but includes all of the benefits of the general purpose model.

Nice.

That's kind of fun.

Yeah.

How how accurate was detecting rabbits?

It's very accurate.

We'll have to see how it works with the Easter Bunny.

But you mean he didn't put the Easter Bunny in your taxes next year.

So?

So if I'm a developer and I want to build, like maybe I'm an expert in rabbit detection and you know, and I want to build, like, a custom model and I want to contribute this inter tensorflow, how would I go about it?

Well, that's something that we're really excited to work on over the next few months.

Right now, a module is accessible via any girl, and so you can host a module, and you can use it in that one line of code.

But we want to create a platform where people can go and find modules for all sorts of different topics and have that one central place that you always know that you can get really high quality stuff in a variety of different areas.

Is that like some kind of curation?

Um, that's something it remains to be, you know, seeing specifically what we're going to dio.

But we want to be able to open it up to as wide a group of people.

A ce possible.

Sounds good.

So where can I find this again?

It's tensorflow dot org's slash hub tensorflow dog slash hub I keep saying tensorflow dot hub slash or way should register main tensorflow dot org slash you go check it out.

So thank you so much, Andrew.

Thank you.

It's been really fun, and I love that slide.

Okay, so thanks everybody for watching this episode.

If you've got any questions for me, if you have any questions for Andrew, please leave him in the comments below And all of the links that we spoke about today, we'll put in the text description, so thank you so much.

And don't forget to hit that subscribe button, okay?