Laurence Maroney here.

I'm at TensorFlow World, and it's a great privilege

that I have to chat with Megan Kacholia, vice

president of engineering working on TensorFlow.

And you just gave a great keynote about TensorFlow

and about TensorFlow 2 and some of the new things in it.

Could you tell us a little bit about it?

MEGAN KACHOLIA: Yeah, we've been working on TensorFlow 2

for a little while, all right?

We talked about it at the Dev Summit

earlier this year in the spring, and then

just finalized the release in September, so just last month.

One big thing with TensorFlow 2 is

we're focused a lot on just usability, all right,

making it easier for users to be able to do what they need

to do with cleaner, more streamlined APIs, better

debugability, with eager mode, just try and make things

simpler and easier for folks.

LAURENCE MARONEY: One of the things I really

liked in your keynote was where you called out

like three audiences, right?

There was researchers, there was data scientists,

and there was developers.

I spend most of my time working with developers,

but what I'd love to really drill down on

is researchers and some of the great stuff

that's in TensorFlow 2 for researchers.

Could you tell us about that?

MEGAN KACHOLIA: Sure.

I think sometimes people worry that if they

want a lot of control--

that's one thing researchers really like,

is they want to be able to try out-- especially ML

researchers-- they want to be able to try out new things.

And they want to make sure they have that control

and be able to do that.

I think one thing to emphasize is that researchers still have

that control with TensorFlow 2.

Yes, we have done a lot to try and streamline

the high-level APIs, because we want

to make it easier for folks to commit at that point.

But I know there's a lot of folks who just want

to be able to go in deeper.

They want to go under the covers.

They want to go under the hood and be able to figure out,

well, I have this new model type I want to try out.

I have some new model architecture

I want to play with.

And all the things they loved about TensorFlow

and being able to do that before, that's all still there.

We're just trying to make sure there's also

a nice clean high-level surface as well.

It's not an either/or situation.

There's still both.

LAURENCE MARONEY: Yeah.

And then, all that great work that we've

done to make it easier for software developers who

aren't necessarily researchers doesn't preclude that.

MEGAN KACHOLIA: Exactly.

LAURENCE MARONEY: Everything the researchers

loved is still there.

MEGAN KACHOLIA: Exactly.

LAURENCE MARONEY: So-- and with TensorFlow 2,

there's also some state-of-the-art research has

gone on, right?

MEGAN KACHOLIA: Yes, definitely.

I mean, just even showing today the example that I went through

from Hugging Face.

Hugging Face has done so many cool things

in the NLP, the natural language processing space.

Everyone's very excited about BERT models right now.

I think you hear it everywhere.

Everyone's talking about BERT.

LAURENCE MARONEY: They just like saying BERT.

MEGAN KACHOLIA: It's a fun name.

Everyone's talking about BERT right now, and transformers.

And you can see even they were able to implement

some of the more advanced models in TensorFlow 2.0.

And recently, they did it.

And it really exciting for us to be

able to highlight that and show that no, it

wasn't us, the TensorFlow team who

has all the intricate knowledge of things.

It was the external community being

able to use this for really advanced research use cases.

LAURENCE MARONEY: And that's a really important thing, right?

So it's like we can build what we know, but it's

when people are able to take that platform

and bring their expertise to it--

MEGAN KACHOLIA: Yes, exactly.

LAURENCE MARONEY: --and make changes.

MEGAN KACHOLIA: And do what they need to do with it.

LAURENCE MARONEY: That's what I find particularly inspiring,

really, really cool.

And one of the other things that you drilled into

was performance.

You've been working really hard to tweak performance

on Tensorflow as well.

MEGAN KACHOLIA: Yes.

LAURENCE MARONEY: Could you talk a little bit about that?

MEGAN KACHOLIA: Yeah, so performance

is something that we always pay attention to.

I mean, one big thing that came with TensorFlow 2.0 as well was

distribution strategy.

Make it easy for people to be able to scale things up and not

have to worry about how do I set things?

How do I do things?

Like no, let us handle more for you.

So that's something that we focused a lot on,

is just how do you get that performance

and get it in an easy way so that way users don't

have to worry?

Now, again, if people want to go under the covers

and tinker and tweak every last thing, that's fine.

That's all still there, but we want

to make sure that it's easy for them

to get the performance they need.

So I talked some numbers for GPUs,

looking at different types of Nvidia GPUs, the 3x performance

improvement that we're able to get there with 2.0.

LAURENCE MARONEY: Which is amazing.

MEGAN KACHOLIA: Which is great.

Yeah, it's great, those things.

And then in the upcoming releases,

we'll have more things around TPUs,

as well as Mixed Precision support.

LAURENCE MARONEY: Wow.

OK, so still working hard.

MEGAN KACHOLIA: Yes.

It's never-- I mean, it's just like anything else.

The work keeps going.

LAURENCE MARONEY: Yes, exactly.

MEGAN KACHOLIA: There's always new things happening.

LAURENCE MARONEY: So I'm preventing

the advancement of this performance

by having you here to talk to us.

Thanks for taking the time.

MEGAN KACHOLIA: No, that's all right.

LAURENCE MARONEY: But a couple of other things

as well for developers, and for tooling around developers,

there's TensorBoard, right?

MEGAN KACHOLIA: Yes.

LAURENCE MARONEY: There's been advances in that.

MEGAN KACHOLIA: Yes.

LAURENCE MARONEY: Could you--

MEGAN KACHOLIA: One big thing with TensorBoard

that we announced here at TensorFlow World

is hosted TensorBoard.

So the whole idea here is that, again, people

love to be able to share their experiments.

Sometimes it's because they need help.

Let me show you.

Can you help me see what's going on?

Sometimes it's because like, oh, my gosh, look.

Look at what I found.

Look at the results I got.

Let me show someone else.

And right now, folks are generally

sharing with screenshots.

They'll take screenshots of TensorBoard

just showing what's going on with their experiment

or their current set-up.

And instead, we want to make it so folks can actually

share their TensorBoard instead of showing

a picture of some snapshot in time of their TensorBoard.

And that's where hosted TensorBoard comes in.

So this is something that we're just starting

to preview release for.

There's a lot of features we'll be adding

over the coming months as we stabilize it and get it to more

of the general availability.

But I think it's really exciting for folks

to be able to take that and then make use of it

to more easily share with others,

like, look what I'm doing.