My name's Josh Gordon, and I'm here today with J J Lair and J.

J.

Is the founder of our studio, and he recently added support for tensorflow and care.

Us two are so J.

J.

Thank you very much for coming on the show.

Thank you.

And could you tell me a little bit about the work you've done?

Absolutely.

Tensorflow came out, I think, about 2.5 years ago on from the day that it came out.

I've been excited about what we could do with tensorflow from our are the Our community is filled of people who are not.

They don't self identify software developers.

They identify a statisticians or economists or biologists and the interfaces we build to modeling and statistics from our tend to be really fluent, high level interfaces that kind of really well married to the vocabulary of the domain that people are doing working.

And so tens of low brings this awesome capability for neural networks and even just as a general purpose numerical computing framework.

And I thought, Wow, we could do incredible things from our to make all this power available to all these people solving, solving heart problems then, particularly the caress interface which I hope we get to talk about a little bit more is a the really nice high level vocabulary for doing deep learning that fits really well with the way our users work and think on.

So I was excited to both can make the core capabilities of tensorflow available, but then also to make the caress a P I available.

One thing you did to is you contributed some really great educational resource is especially clearly, yes, so we probably spent as much or more time on educational resource is as we did on the actual interfaces.

So I worked on a book along with Francoise Chalet, about deep learning with are adapted his Deepening with Python book.

Francoise is actually the creator of caress so worked on that book.

We created a whole bunch of examples, probably 25 or so examples of using caress and many, many other examples.

We created a gallery of kind of longer form blawg posts that described like really in depth worked examples in different domains.

So we have invested a huge amount and education resource is, and we're gonna continue to do that and this book is excellent.

The book is awesome.

Yes, I read the python version cover to cover, and I taught a class with it.

Students loved it.

I strongly recommend it to all the developers a meet.

What's remarkable about it is that it covers the concepts, the conceptual terrain of deep learning in a way that I think was really intuitive for people to understand was not not a huge number of prerequisites but also has a wealth of practical information about how to actually do deep learning.

What are the small problems?

You come up with one of the failure moods and how to overcome them.

So it's just it's a rare book that combines kind of conceptual material and practical material.

So I I recommend that is the first thing that people who are in the our community and they say, I want to learn more about this.

I recommend that they get that book and read it first.

It's probably a lot of work porting that over our Yeah, it was, but most of the books predominately conceptual.

And then there are these examples, and so I really just changed the examples, but really most of the book is ends up being conceptual, that kind of at a higher level than specific languages.

So I wasn't too bad.

And then another thing you have on the our website, the Our City website, which is really cool, is worked and end examples.

That's right.

We have a gallery where we don't just show you.

Here's the mechanics of solving a certain type of problem, but we go kind of from the beginning.

Here's the motivation.

Here's the data set.

Maybe let's visualize the data set a little bit.

Let's try different approaches to the problem and see what works and doesn't work.

Eso way kind of cover end, end Like what?

Work flow of the data scientists would go through to solve these problems.

So and then we're gonna be investing more of this.

I think that's the way that people can really get engaged with.

This is not just learning the basic mechanics and concepts, but seeing how it actually applies to their field.

I like this combination a lot.

Yes, A really solid set of learning resource is good.

So in addition to supporting care us, you also have support for tensorflow estimated Yes, we have another our package.

A suite of seven our packages that we have for Tensorflow.

There's our package called tensorflow, which is a low level interface to the full tensorflow graph in the full potential FBI.

There's a caress a poetry been talking about, and we have tensorflow estimators, which is another high level framework that Google has for doing models like classifications, regression models, classification and regression using D N ends.

And so there's a really, really high level functions that come with a nice framework for for data, pre processing and things and a nice framework for deploying the models.

We have another package called TF estimators that covers the covers.

The estimators FBI deployment is really important to have a swell it is.

And that's one of the things that I'm most excited about because traditionally with our when you do modeling and then you want to deploy the model in some fashion, you have to bring the R run time along with you, and that could be a challenge.

So what the The design of Tensorflow is that you're writing a program, but that program is creating a graph, and that graph is executed by a run time a C plus plus runtime.

So deploying models doesn't require that you bring our or python or any of these sorts of language.

The language is that used to model into deployment.

And so we have another package called TF deploy that helps people with sort of serializing there save models.

Um, putting rest interfaces on top of their save models.

Getting models, uh, actually brought into Java script.

There's a library card caress J asked you to caress, model and run in your browser.

So we've got lots of tools for that.

And that's actually one of the most exciting things about tensorflow is that is this deployment model, so tensorflow dot Js okay, supports a caress compatible AP Iook.

And I believe what this means is that people will be ableto author models in our That's right, using the caress interface that's right and then deployment.

Deploy them right into a browser.

Yeah, that's gonna be huge.

It's phenomenal.

Yeah, So I think a lot of people in the argument that I've taught I mean all communities I think this is such a fundamentally good idea to take models and serialize them into this runtime format.

That doesn't because the languages and tools that are good for training for exploratory date analysis, training models authoring models that are different than languages that are good for deployment.

It's a really good division of responsibilities that tensorflow kind of enforces.

And so I'm hopeful that our users who want to put their models and production now we'll have a really straightforward way to do that.

I think so, too.

And it's a really good opportunity for data scientists to visualize the results of their experiments.

I actually deployed.

Yeah, yeah.

So salmon are developer knew too deep, learning about how much code is it going to take for me to write my first neural network?

Well, interestingly, if you look at our examples that the hard part about deep learning is not generally writing the code, uh, you know, many of our examples are on the order of 100 lines of code or even even less than that.

Sometimes the hard part is actually figuring out what sort of mild to build and how it should, what the architecture of the mile should be and what the hyper parameters of the mouth should be.

That's where all the time is the actual expression of the model in code is actually quite trivial.

But model and model building really ends up being not about the code, but about about mapping your problem in the data and the tools you have together toe get a good solution.

I completely agree.

It's well put.

I think a lot of people from the art community, especially from stats and data science background, will be natural.

The background and training of people in the our community, the understanding how to do modeling statistics, probability is perfect for deep learning.

And I'm really excited once more people in the argument to get, get, get their hands on these tools of the kind of work that we're going to see.

J.

J Thanks again for coming on the show.

I really appreciate it just to close things out.

Where's the best place for developers to go to learn more about deep learning and are the best place?

We have?

A website called Tensorflow that our studio dot com it has documentation for all the packages I talked about Paris and estimators and and the court tensorflow FBI, as information about deployment has a gallery of in depth examples kind of blawg posts.

It has a whole bunch of additional learning.

Resource is so tense there flowed out.

Our studio dot com is definitely the place to go to get started.

Okay.

Thanks again.

Really appreciate it.

Thanks.