And welcome to the TENSORFLOW developer.

Someone.

I'm Laurence Maroney here in the Tensorflow Cafe on I'm chatting with Clements Me walls and Clements.

You work on tensorflow extended or T f X.

Could you tell us all about it?

Seems really cool.

Yeah, it is.

Um, Texas Unturned Machinery platform that we pulled around Tensorflow.

Okay, what this really means is that internally it could or a lot of developers actually want to deploy tensorflow models in production.

Okay.

And it's very hard, because in addition to the machine learning piece, you need many, many more pieces, right intensive extent.

It is exactly this and 20 machinery from the takes care of all of this because one of the things in your talk that I really like the beginning of the talk was like, You have this machine learning little box, and you've got all this stuff surrounding it.

That kind of really hit home, because it makes it really so this is actually taken from a paper from one of my colleagues, De Scully and a couple of co authors.

And in this paper, he was making this point that machine learning systems are extremely complex systems that have all of these other components, and it's extremely hard to integrate all of them and make them work well together.

And there's actually a lot of technical debt that was incurred by gluing things together at Hawk.

Okay, so when we tried with the effects is actually to provide a solution for this problem, provide one well integrated platform that had all of these pieces integrated.

And it provided an easy way for developers to use all of these pieces and 20 and get started really fast.

Wow.

So there were a whole bunch of things in the eco system, but one of them was like model analysis.

And I know you've just published the blood post on that.

Could you could you tell us a little bit about it?

So model analysis.

And we just open sources today as tense formal analysis is TFM a future mate just sounds perfect, refused to technology internally for a while now, and we use it for a couple off purposes.

One off them, really is that once a pencil from always trained, you really want to find out how it performs on your entire data set on slices off the data set for subgroups of your data, and you will really want to drill down into these metrics because in some cases, or in many cases people only look at an aggregate quality metric.

And it may tell you that the model works well on average.

Right?

But every rich is not always enough.

And it could be some false positives and false negatives hidden right Exactly what you're doing so tense.

If Amal analysis helps with both computing this matrix or a large amounts of data and also for sliced off those data and then we provide some front and tooling an interactive components to drill down into these metrics, okay, and find some areas of data where the model minute you performing well or to find error cases, our cases where the model is actually not performing as expected.

And then developers can use this information to either improve the model or go fix their data.

If there's a mistake.

Okay, sounds good.

So sure love the time it is a it's a case of Ri factoring your data to make it as efficient as possible.

Yes, OK, yeah.

So but this will help you find that because otherwise there's a lot of trial and error.

Exactly.

Okay, so there is one of the other things that I've seen is a T f serving right?

And that's why I get a lot of questions from people like, you know, howto I expose my model to the world.

You know, that kind of thing.

It's like it's all very well if I have it on my own machine and I could call it So how does he have several actually work?

City of serving.

It's very interesting.

The court business.

We're actually open.

Sourcing is exactly the corpus that we use In turn little serve our machine learning models.

Okay.

And the basic concept behind it is that you don't want to heart chord your mission and model into an application.

Okay?

You would rather have a server that loads a model and a new can make requests to That's over.

I see which has a lot of benefits intensively serving his executive solution that, um, that provides the community with this technology.

And we just announced today that we're working on arrested P.

I okay, for it's over because this is the number one future request from the community and referred to the community.

And we're working very hard on releasing this.

You read my mind cause that's exactly what's gonna ask.

What is the interface?

Is it arrest a P?

I like So you know, what is the interface today?

Well, apparently it's Ah, it's Gervasi.

Okay, which works really well, this is what we use internally, but obviously externally, the standard in the community and things that a lot of developers asked us is for arrested.

Okay, Cool.

Yeah, I would love to see arrested here around this myself.

So I'll plus one that boat.

Perfect.

So now if I want to get started with any of this stuff, I'm a developer, and I really want to get into T FX and see what it gives me.

Maybe model analysis are serving or any of the other tools.

Where should I go to get started?

Great questions of desert.

There's a couple of components that we've already released, namely transform mal analysis and serving.

Okay, All of these are under the tensorflow repo on get up.

But most importantly, when we released Moloch analysis today, the example that we released with mal analysis shows all of thes and how to use together.

Nice.

And that's in your covered in your blood post.

Yes, okay, that's being discussed in a block post.

And there's also a link where people can actually go and follow the example to see how to use all of these components together.

Cool.

Nice.

And then there's more more components on the way.

Exactly.

A za pointed out in the depths of my talk.

This is only a very small part of defects.

Were working very hard to release more, to gift this benefit to the community and see what the community can do with it.

Okay, well, sounds good.

I have to check it out.

Where can I learn about this again?

Right now, the best place is really to start with the block post intense form of analysis, and we are actually working on a block post, specifically 40 if x into part of project.

Sounds good.

I believe I'm working on that with you.

So on.

So we'll put links to all of these in the description below, so, you know, for people watching So thank you so much, Clements.

I've learned so much about T fx and I hope that you have to.

It's been a real pleasure having you on the show.

Thanks for having me.

So thanks everybody for watching this episode.

If you've got any questions for me, if you have any questions for Clements, please leave him in the comments below.

On.

As always, don't forget to hit that subscribe button.