See all Lex Fridman transcripts on Youtube

Speaker 1

00:00

The following is a conversation with Rajat Manga. He's an engineering director at Google, leading the TensorFlow team. TensorFlow is an open source library at the center of much of the work going on in the world in deep learning, both the cutting edge research and the large scale application of learning based approaches. But it's quickly becoming much more than a software library.

Speaker 1

00:20

It's now an ecosystem of tools for the deployment of machine learning in the cloud, on the phone, in the browser, on both generic and specialized hardware, TPU, GPU, and so on. Plus, there's a big emphasis on growing a passionate community of developers. Raja, Jeff Dean, and a large team of engineers at Google Brain are working to define the future of machine learning with TensorFlow 2.0, which is now in alpha. I think the decision to open source TensorFlow was a definitive moment in the tech industry.

Speaker 1

00:51

It showed that open innovation can be successful and inspire many companies to open source their code to publish and in general engage in the open exchange of ideas. This conversation is part of the Artificial Intelligence Podcast. If you enjoy it, subscribe on YouTube, iTunes, or simply connect with me on Twitter at Lex Friedman, spelled F-R-I-D. And now, here's my conversation with Rajat Manga.

Speaker 2

01:17

You were involved with Google Brain since its start in 2011 with Jeff Dean. It started with Disbelief, the proprietary machine learning library, and turned into TensorFlow in 2014, the open source library. So what were the early days of Google Brain like?

Speaker 2

01:39

What were the goals, the missions? How do you even proceed forward once there's so much possibilities before you?

Speaker 3

01:47

It was interesting back then, you know, when I started, or when we were even just talking about it, the idea of deep learning was interesting and intriguing in some ways. It hadn't yet taken off, but it held some promise. It had shown some very promising and early results.

Speaker 3

02:08

I think the idea where Andrew and Jeff had started was, what if we can take this, what people are doing in research and scale it to what Google has in terms of the compute power. And also put that kind of data together, what does it mean? And so far the results have been if you scale the compute, scale the data, it does better and would that work? And so that was the first year or 2, can we prove that out, right?

Speaker 3

02:35

And with Disbelief, when we started the first year, we got some early wins, which is always great.

Speaker 2

02:40

What were the wins like? What was the wins where you were, there's some problems to this, this is gonna be good?

Speaker 3

02:46

I think there are 2 early wins where 1 was speech that we collaborated very closely with the speech research team who was also getting interested in this. And the other 1 was on images where we, you know, the cat paper as we call it, that was covered by a lot of folks.

Speaker 2

03:03

And the birth of Google Brain was around neural networks. That was, so it was deep learning from the very beginning. That was the whole mission.

Speaker 2

03:10

So what would, in terms of scale, what was the sort of dream of what this could become? Were there echoes of this open source TensorFlow community that might be brought in? Was there a sense of TPUs? Was there a sense of like, machine learning is now gonna be at the core of the entire company, is going to grow into that direction?

Speaker 3

03:36

Yeah, I think, so that was interesting, and like if I think back to 2012 or 2011, and first was can we scale it, and in the year or So we had started scaling it to hundreds and thousands of machines. In fact, we had some runs even going to 10,000 machines. And all of those shows great promise.

Speaker 3

03:53

In terms of machine learning at Google, the good thing was Google's been doing machine learning for a long time. Deep learning was new. But as we scaled this up, we showed that yes, that was possible, and it was going to impact lots of things. Like we started seeing real products wanting to use this.

Speaker 3

04:11

Again, speech was the first. There were image things that photos came out of and then many other products as well. So that was exciting. As we went into that a couple of years, externally also academia started to, there was lots of push on, okay, deep learning's interesting, we should be doing more, and so on.

Speaker 3

04:30

And so by 2014, we were looking at, okay, this is a big thing, it's gonna grow. And not just internally, externally as well. Yes, maybe Google's ahead of where everybody is, but there's a lot to do. So a lot of this started to make sense and come together.

Speaker 2

04:46

So the decision to open source, I was just chatting with Chris Glattner about this, the decision to go open source with TensorFlow, I would say, for me personally, seems to be 1 of the big seminal moments in all of software engineering ever. I think that's when a large company like Google decides to take a large project that many lawyers might argue has a lot of IP, just decide to go open source with it, and in so doing, lead the entire world in saying, you know what, open innovation is a pretty powerful thing, and it's okay to do. That was, I mean, that's an incredible moment in time.

Speaker 2

05:26

So do you remember those discussions happening? Whether open source should be happening? What was that like?

Speaker 3

05:32

I would say, I think, so the initial idea came from Jeff, who was a big proponent of this. I think it came off of 2 big things. 1 was, research-wise, we were a research group.

Speaker 3

05:46

We were putting all our research out there. If you wanted to, we were building on others' research and we wanted to push the state of the art forward. And part of that was to share the research. That's how I think deep learning and machine learning has really grown so fast.

Speaker 3

06:01

So the next step was, okay, now, would software help with that? And it seemed like there were existing a few libraries out there, Tiano being 1, Torch being another, and a few others. But they were all done by academia, and so the level was significantly different. The other 1 was from a software perspective, Google had done lots of software that we used internally, and we published papers.

Speaker 3

06:29

Often there was an open source project that came out of that, that somebody else picked up that paper and implemented, and they were very successful. Back then it was like, okay, there's Hadoop, which has come off of tech that we've built. We know the tech we've built is way better for a number of different reasons. We've invested a lot of effort in that.

Speaker 3

06:52

And turns out we have Google Cloud and we are now not really providing our tech, but we are saying, okay, we have Bigtable, which is the original thing. We are going to now provide HBase APIs on top of that, which isn't as good, but that's what everybody's used to. So there's like, can we make something that is better and really just provide? Helps the community in lots of ways, but also helps push a good standard forward.

Speaker 2

07:18

So how does Cloud fit into that? There's a TensorFlow open source library. And how does the fact that you can use so many of the resources that Google provides and the Cloud fit into that strategy?

Speaker 3

07:31

So TensorFlow itself is open, and you can use it anywhere. And we want to make sure that continues to be the case. On Google Cloud we do make sure that there's lots of integrations with everything else and we want to make sure that it works really really well there.

Speaker 3

07:46

So

Speaker 2

07:47

You're leading the TensorFlow effort. Can you tell me the history and the timeline of TensorFlow project in terms of major design decisions, so like the open source decision, but really what to include and not? There's this incredible ecosystem that I'd like to talk about.

Speaker 2

08:04

There's all these parts, but what if you just, some sample moments that defined what TensorFlow eventually became through its, I don't know if you're allowed to say history when it's just, but in deep learning everything moves so fast in just a few years, is already history.

Speaker 3

08:23

Yes, yes. So looking back, we were building TensorFlow, I guess we open sourced it in

Speaker 1

08:31

2015,

Speaker 3

08:32

November 2015. We started on it in summer of 2014, I guess. And somewhere like 3 to 6, late 2014, by then we had decided that, okay, there's a high likelihood we'll open source it.

Speaker 3

08:47

So we started thinking about that and making sure we're heading down that path. By that point, we had seen a few, lots of different use cases at Google. So there were things like, okay, yes, you want to run in at large scale in the data center. Yes, we need to support different kind of hardware, we had GPUs at that point, we had our first GPU at that point, or was about to come out, you know, roughly around that time.

Speaker 3

09:15

So the design sort of included those. We had started to push on mobile, so we were running models on mobile. At that point, people were customizing code, so we wanted to make sure TensorFlow could support that as well, so that that sort of became part of that overall design.

Speaker 2

09:35

When you say mobile, you mean like pretty complicated algorithms running on the phone?

Speaker 3

09:39

That's correct. So when you have a model that you deploy on the phone and run it the right way.

Speaker 2

09:45

So already at that time there was ideas of running machine learning on the phone.

Speaker 3

09:48

That's correct. We already had a couple of products that were doing that by then. And in those cases, we had basically customized handcrafted code or some internal libraries that we're using.

Speaker 2

10:00

So I was actually at Google during this time in a parallel, I guess, universe, but we were using Theano and Caffe. Was there some degree to which you were balancing, like trying to see what Caffe was offering people, trying to see what Theano was offering, that you want to make sure you're delivering on whatever that is, perhaps the Python part of things, maybe, did that influence any design decisions?

Speaker 3

10:27

Totally, so when we built this belief, And some of that was in parallel with some of these libraries coming up. I mean, Tiano itself is older. But we were building Disbelief focused on our internal thing because our systems were very different.

Speaker 3

10:42

By the time we got to this, we looked at a number of libraries that were out there. Tiano, there were folks in the group who had experience with Torch, with Lua. There were folks here who had seen Caffe. I mean, actually, Yang Jing was here as well.

Speaker 3

10:58

There's... What other libraries? I think we looked at a number of things. Might even have looked at JNR back then.

Speaker 3

11:06

I'm trying to remember if it was there. In fact, yeah, we did discuss ideas around, okay, should we have a graph or not? And they were so supporting all these together was definitely, you know, they were key decisions that we wanted. We had seen limitations in our prior disbelief things.

Speaker 3

11:28

A few of them were just in terms of research was moving so fast, we wanted the flexibility, the hardware was changing fast, we expected to change that, so that those probably were 2 things. And yeah, I think the flexibility in terms of being able to express all kinds of crazy things was definitely a big 1 then.

Speaker 2

11:46

So what, the graph decisions, so that was moving towards TensorFlow

Speaker 1

11:51

2.0.

Speaker 2

11:52

There's more, by default, it'll be eager execution. So sort of hiding the graph a little bit because it's less intuitive in terms of the way people develop and so on. What was that discussion like in terms of using graphs?

Speaker 2

12:06

It seemed, it's kind of the Theano way, did it seem the obvious choice? So I think where

Speaker 3

12:13

it came from was our disbelief had a graph-like thing as well. A much more, it wasn't a general graph, it was more like a straight line thing. More like what you might think of cafe, I guess in that sense.

Speaker 3

12:26

But the graph was, and we always cared about the production stuff. Like even with Disbelief, we were deploying a whole bunch of stuff in production. So, graph did come from that when we thought of, okay, should we do that in Python? And we experimented with some ideas where it looked a lot simpler to use, but not having a graph meant, okay, how do you deploy now?

Speaker 3

12:48

So that was probably what tilted the balance for us and eventually we ended up with a graph.

Speaker 2

12:52

And I guess the question there is, did you, I mean, so production seems to be the really good thing to focus on, but Did you even anticipate the other side of it where there could be, what is it, what are the numbers? Something crazy, 41 million downloads? I mean, was that even a possibility in your mind that it would be as popular as it became?

Speaker 3

13:19

So I think we did see a need for this a lot from the research perspective and like early days of deep learning in some ways.

Speaker 1

13:32

41

Speaker 3

13:32

million, No, I don't think I imagined this number then. It seemed like there's a potential future where lots more people would be doing this and how do we enable that? I would say this kind of growth, I probably started seeing somewhat after the open sourcing where it was like, okay, you know, deep learning is actually growing way faster for a lot of different reasons.

Speaker 3

13:59

And we are in just the right place to push on that and leverage that and deliver on lots of things that people want.

Speaker 2

14:07

So what's changed once you open source? Like how, you know, this incredible amount of attention from a global population of developers, what, how did the project start changing? I don't even actually remember during those times.

Speaker 2

14:22

I know looking now, there's really good documentation, there's an ecosystem of tools, there's a community, there's a YouTube channel now, right? It's very community driven. Back then, I guess 0.1 version. Is that the version?

Speaker 3

14:39

I think we called it 0.6 or 5, something like that.

Speaker 2

14:43

What changed leading into

Speaker 1

14:45

1.0?

Speaker 3

14:47

It's interesting. I think we've gone through a few things there. When we started out, when we first came out, people loved the documentation we have, because it was just a huge step up from everything else, because all of those were academic projects, people doing, you know, would don't think about documentation.

Speaker 3

15:04

I think what that changed was, instead of deep learning being a research thing, some people who were just developers could now suddenly take this out and do some interesting things with it, right? Who had no clue what machine learning was before then. And that, I think, really changed how things started to scale up in some ways and pushed on it. Over the next few months, as we looked at, you know, how do we stabilize things?

Speaker 3

15:31

As we look at not just researchers, now we want stability, people want to deploy things. That's how we started planning for 1.0. And there are certain needs for that perspective. And so again, documentation comes up, designs, more kinds of things to put that together.

Speaker 3

15:49

And so that was exciting to get that to a stage where more and more enterprises wanted to buy in and really get behind that. And I think post 1.0 and you know, with the next few releases, that enterprise adoption also started to take off. I would say between the initial release and 1.0, it was, okay, researchers, of course, then a lot of hobbies and early interest, people excited about this who started to get on board, and then over the 1.x thing, lots of enterprises.

Speaker 2

16:19

I imagine anything that's below 1.0, gets pressure to be, enterprise probably wants something that's stable.

Speaker 3

16:28

Exactly.

Speaker 2

16:29

And Do you have a sense now that TensorFlow is state, like it feels like deep learning in general is extremely dynamic field, so much is changing. TensorFlow has been growing incredibly. You have a sense of stability at the helm of it?

Speaker 2

16:46

I mean, I know you're in the midst of it, but.

Speaker 3

16:48

Yeah, I think in the midst of it, it's often easy to forget what an enterprise wants and what some of the people on that side want. There are still people running models that are 3 years old, 4 years old. So Inception is still used by tons of people.

Speaker 3

17:05

Even ResNet-50 is, what, a couple of years old now or more? But there are tons of people who use that, and they're fine. They don't need the last couple of bits of performance or quality. They want some stability in things that just work.

Speaker 3

17:20

And so there is value in providing that with that kind of stability and making it really simpler because that allows a lot more people to access it. And then there's the research crowd which wants, okay, they want to do these crazy things exactly like you're saying, right? And not just deep learning in the straight up models that used to be there. They want RNNs and even RNNs are maybe old, they are transformers now, and now it needs to combine with RL and GANs and so on.

Speaker 3

17:48

So there's definitely that area that, like the boundary that's shifting and pushing the state of the art. But I think there's more and more of the past that's much more stable and even stuff that was 2, 3 years old is very, very usable by lots of people. So that part makes it a lot easier.

Speaker 2

18:07

So I imagine, maybe you can correct me if I'm wrong, 1 of the biggest use cases is essentially taking something like ResNet-50 and doing some kind of transfer learning on a very particular problem that you have. It's basically probably what majority of the world does. And you want to make that as easy as possible.

Speaker 3

18:27

So I would say, for the hobbyist perspective, that's the most common case, right? In fact, the apps on phones and stuff that you'll see, the early ones, that's the most common case. I would say there are a couple of reasons for that.

Speaker 3

18:40

1 is that everybody talks about that. It looks great on slides. Yeah, that's a part of the presentation. Yeah, exactly.

Speaker 3

18:49

What enterprises want is that is part of it, but that's not the big thing. Enterprises really have data that they want to make predictions on. This is often what they used to do with the people who were doing ML was just regression models, linear regression, logistic regression, linear models, or maybe gradient booster trees and so on. Some of them still benefit from deep learning, but they weren't that, that's the bread and butter, like the structured data and so on.

Speaker 3

19:16

So depending on the audience you look at, they're a little bit different.

Speaker 2

19:19

And they just have, I mean, the best of enterprise probably just has a very large data set, or deep learning can probably shine.

Speaker 3

19:28

That's correct, that's right. And then I think the other pieces that they want, again, with 2.0, the developer summit we put together is the whole TensorFlow extended piece, which is the entire pipeline. They care about stability across doing their entire thing.

Speaker 3

19:43

They want simplicity across the entire thing. I don't need to just train a model, I need to do that every day again, over and over again.

Speaker 2

19:51

I wonder to which degree you have a role in, I don't know, so I teach a course on deep learning, I have people like lawyers come up to me and say, you know, say, when is machine learning gonna enter legal, the legal realm? The same thing in all kinds of disciplines, immigration, insurance. Often when I see what it boils down to is these companies are often a little bit old school in the way they organize the data.

Speaker 2

20:20

So the data is just not ready yet. It's not digitized. Do you also find yourself being in the role of an evangelist for like, let's get, organize your data folks and then you'll get the big benefit of TensorFlow. Do you get those, have those conversations?

Speaker 3

20:38

Yeah, yeah, I get all kinds of questions there from, okay, what do I need to make this work, right? Do we really need deep learning? I mean, there are all these things.

Speaker 3

20:52

I already use this linear model, why would this help? I don't have enough data, let's say, you know, or I want to use machine learning, but I have no clue where to start. So it varies, back to all the way to the experts who wise for very specific things, so it's interesting.

Speaker 1

21:08

Is there a good answer?

Speaker 2

21:09

It boils down to oftentimes digitizing data. So whatever you want automated, whatever data you want to make prediction based on, you have to make sure that it's in an organized form. Like within the TensorFlow ecosystem, there's now, you're providing more and more datasets and more and more pre-trained models.

Speaker 2

21:28

Are you finding yourself also the organizer of datasets?

Speaker 3

21:32

Yes, I think with TensorFlow datasets that we just released, that's definitely come up where people want these data sets. Can we organize them and can we make that easier? So that's definitely 1 important thing.

Speaker 3

21:45

The other related thing I would say is I often tell people, you know what, don't think of the most fanciest thing, that the newest model that you see. Make something very basic work, and then you can improve it. There's just lots of things you can do with it.

Speaker 2

21:58

Yeah, start with the basics, sure. 1 of the big things that makes it, makes TensorFlow even more accessible was the appearance, whenever that happened, of Keras. The Keras standard, sort of outside of TensorFlow.

Speaker 2

22:12

I think it was Keras on top of Tiano at first only, and then Keras became on top of TensorFlow. Do you know when Keras chose to also add TensorFlow as a backend, was it just the community that drove that initially? Do you know if there was discussions, conversations?

Speaker 3

22:37

Yeah, so Francois started the Keras project before he was at Google. And the first thing was Theano. I don't remember if that was after TensorFlow was created or way before.

Speaker 3

22:49

And then at some point when TensorFlow started becoming popular, there were enough similarities that he decided to create this interface and put TensorFlow as a backend. I believe that might still have been before he joined Google. So we weren't really talking about that. He decided on his own and thought that was interesting and relevant to the community.

Speaker 3

23:12

In fact, I didn't find out about him being at Google until a few months after he was here. He was working on some research ideas and doing Keras on his nights and weekends project.

Speaker 2

23:24

Oh, interesting. So he wasn't like part of the TensorFlow. He didn't join initially.

Speaker 3

23:29

He joined research and he was doing some amazing research. He has some papers on that and research. He's a great researcher as well.

Speaker 3

23:38

And at some point we realized, oh, he's doing this good stuff. People seem to like the API and he's right here. So we talked to him and he said, okay, why don't I come over to your team and work with you for a quarter and let's make that integration happen. And we talked to his manager and he said, sure, my quarter's fine.

Speaker 3

24:00

And that quarter's been something like 2 years now. And so he's fully on this.

Speaker 2

24:05

So Keras got integrated into TensorFlow, like in a deep way. And now with 2.0, TensorFlow 2.0, sort of Keras is kind of the recommended way for a beginner to interact with TensorFlow. Which makes that initial sort of transfer learning or the basic use cases even for enterprise super simple.

Speaker 3

24:29

That's correct, that's right.

Speaker 2

24:30

So what was that decision like? That seems like a... That's kind of a bold decision as well.

Speaker 3

24:38

We did spend a lot of time thinking about that 1. We had a bunch of APIs some built by us. There was a parallel layers API that we were building.

Speaker 3

24:49

And when we decided to do Keras in parallel, so there were like, OK, 2 things that we are looking at. And the first thing we was trying to do is just have them look similar, like be as integrated as possible, share all of that stuff. There were also like 3 other APIs that others had built over time because we didn't have a standard 1. But 1 of the messages that we kept hearing from the community, okay, which 1 do we use?

Speaker 3

25:13

And they kept saying like, okay, here's a model in this 1, and here's a model in this 1, which should I pick? So that's sort of like, okay, we had to address that straight on with 2.0. The whole idea was we need to simplify, and we had to pick 1. Based on where we were, we were like, okay, let's see what are the people like.

Speaker 3

25:35

And Keras was clearly 1 that lots of people loved. There were lots of great things about it. So we settled on that. Organically, that's kind

Speaker 2

25:45

of the best way to do it. It was great. It was surprising, nevertheless, to sort of bring in an outside, I mean, there was a feeling like Keras might be almost like a competitor in a certain kind of, to TensorFlow, and in a sense it became an empowering element of TensorFlow.

Speaker 3

26:02

That's right. Yeah, it's interesting how you can put 2 things together which can align. And in this case, I think Francois, the team, and a bunch of us have chatted.

Speaker 3

26:14

And I think We all want to see the same kind of things. We all care about making it easier for the huge set of developers out there, and that makes a difference.

Speaker 2

26:23

So Python has Guido van Rossum, who until recently held the position of benevolent dictator for life. Does a huge, successful open source project like TensorFlow need 1 person who makes a final decision? So you've did a pretty successful TensorFlow Dev Summit just now, last couple of days.

Speaker 2

26:47

There's clearly a lot of different new features being incorporated in an amazing ecosystem and so on. How are those design decisions made? Is there a BDFL in TensorFlow or is it more distributed and organic?

Speaker 3

27:05

I think it's it's somewhat different I would say. I've always been involved in the key design directions But there are lots of things that are distributed where there are a number of people, Martin Wick being 1 who has really driven a lot of our open source stuff, a lot of the APIs, and there are a number of other people who've been pushed and been responsible for different parts of it. We do have regular design reviews over the last year.

Speaker 3

27:38

We've really spent a lot of time opening up to the community and adding transparency, we're setting more process in place. So RFCs, special interest groups really grow that community and scale that. I think the kind of scale that ecosystem is in, I don't think we could scale with having me as the stone point of decision maker. So,

Speaker 2

28:02

yeah. I got it. So yeah, the growth of that ecosystem. Maybe you can talk about it a little bit.

Speaker 2

28:07

First of all, when I, it started with Andrej Karpathy when he first did ConvNetJS. The fact that you can train in your own network in the browser in JavaScript was incredible. So now TensorFlow.js is really making that a serious, like a legit thing, a way to operate, whether it's in the back end or the front end. Then there's the TensorFlow Extended, like you mentioned.

Speaker 2

28:32

There's TensorFlow Lite for mobile. And all of it, as far as I can tell, it's really converging towards being able to save models in the same kind of way. You can move around, you can train on the desktop, and then move it to mobile, and so on. So there's that cohesiveness.

Speaker 2

28:52

So can you maybe give me, whatever I missed, a bigger overview of the mission of the ecosystem that's trying to be built and where is it moving forward?

Speaker 3

29:02

Yeah, so in short, the way I like to think of this is our goal is to enable machine learning. And in a couple of ways. 1 is we have lots of exciting things going on in ML today.

Speaker 3

29:16

We started with deep learning, but we now support a bunch of other algorithms too. So 1 is to, on the research side, keep pushing on the state of the art. How do we enable researchers to build the next amazing thing? So BERT came out recently.

Speaker 3

29:31

It's great that people are able to do new kinds of research. There's lots of amazing research that happens across the world. So that's 1 direction. The other is how do you take that across all the people outside who want to take that research and do some great things with it and integrate it to build real products, to have a real impact on people.

Speaker 3

29:51

And so that's the other axis in some ways. You know, at a high level, 1 way I think about it is there are a crazy number of compute devices across the world. We often used to think of ML and training and all of this as, okay, something you do either in the workstation or the data center or Cloud. But we see things running On the phones, we see things running on really tiny chips.

Speaker 3

30:17

I mean, we had some demos at the developer summit. And so the way I think about this ecosystem is, how do we help get machine learning on every device that has a compute capability? And that continues to grow. And so in some ways, this ecosystem has looked at various aspects of that and grown over time to cover more of those.

Speaker 3

30:42

And we continue to push the boundaries. In some areas, we've built more tooling and things around that to help you. I mean, the first tool we started was TensorBoard. You wanted to learn just the training piece.

Speaker 3

30:56

TFX or TensorFlow Extended to really do your entire ML pipelines if you care about all that production stuff, but then going to the edge, going to different kinds of things. And it's not just us now. If you're at a place where there are lots of libraries being built on top, So there are some for research, maybe things like TensorFlow agents or TensorFlow probability that started as research things or for researchers for focusing on certain kinds of algorithms, but they're also being deployed or used by production folks. And some have come from within Google, just teams across Google who wanted to build these things.

Speaker 3

31:36

Others have come from just the community because there are different pieces that different parts of the community care about. And I see our goal as enabling even that, right? It's not, we cannot and won't build every single thing. That just doesn't make sense.

Speaker 3

31:54

But if we can enable others to build the things that they care about, and there's a broader community that cares about that, and we can help encourage that. And that's great. That really helps the entire ecosystem, not just those. 1 of the big things about 2.0 that we are pushing on is, OK, we have these so many different pieces, right?

Speaker 3

32:14

How do we help make all of them work well together? So there are a few key pieces there that we are pushing on, 1 being the core format in there and how we share the models themselves through SaveModel and TensorFlow Hub and so on, and a few of the pieces that we really put this together.

Speaker 2

32:33

I was very skeptical that that's, you know, when TensorFlow.js came out, it didn't seem, or Deep Learning.js, as it was earlier.

Speaker 3

32:40

Yeah, that was the first.

Speaker 2

32:41

It seems like technically a very difficult project. As a standalone, it's not as difficult, but as a thing that integrates into the ecosystem, it seems very difficult. So, I mean, there's a lot of aspects of this you're making look easy, but on the technical side, how many challenges have to be overcome here?

Speaker 2

33:00

A lot. And still have to be overcome. That's the question here too.

Speaker 3

33:04

There are lots of steps to it. I mean, we've iterated over the last few years, so there's a lot we've learned. I, yeah, often when things come together well, things look easy, and That's exactly the point.

Speaker 3

33:16

It should be easy for the end user, but there are lots of things that go behind that. If I think about still challenges ahead, there are, you know, We have a lot more devices coming on board, for example, from the hardware perspective. How do we make it really easy for these vendors to integrate with something like TensorFlow? So there's a lot of compiler stuff that others are working on.

Speaker 3

33:46

There are things we can do in terms of our APIs and so on that we can do. TensorFlow started as a very monolithic system. And to some extent, it still is. There are lots of tools around it, but the core is still pretty large and monolithic.

Speaker 3

34:02

1 of the key challenges for us to scale that out is how do we break that apart with clearer interfaces? In some ways, it's software engineering 101, but for a system that's now 4 years old, I guess, or more, and that's still rapidly evolving and that we're not slowing down with, it's hard to change and modify and really break apart. It's sort of like, as people say, right? It's like changing the engine with a car running or fix, that's exactly what we're trying to do.

Speaker 2

34:35

So there's a challenge here because the downside of so many people being excited about TensorFlow and coming to rely on it in many of their applications is that you're kind of responsible, like it's the technical debt, you're responsible for previous versions to some degree still working. So when you're trying to innovate, I mean, it's probably easier to just start from scratch every few months.

Speaker 3

35:05

Absolutely.

Speaker 2

35:07

So do you feel the pain of that? A 2.0 does break some back compatibility, but not too much. It seems like the conversion is pretty straightforward.

Speaker 2

35:18

Do you think that's still important, given how quickly deep learning is changing? Can you just, the things that you've learned,

Speaker 3

35:26

can you just start over, or is there pressure to not? It's a tricky balance. So if it was just a researcher writing a paper who a year later will not look at that code again, sure, it doesn't matter.

Speaker 3

35:41

There are a lot of production systems that rely on TensorFlow, both at Google and across the world. And people worry about this. I mean, these systems run for a long time. So it is important to keep that compatibility and so on.

Speaker 3

35:57

And yes, it does come with a huge cost. We have to think about a lot of things as we do new things and make new changes. I think it's a trade-off, right? You can, you might slow certain kinds of things down, but the overall value you're bringing because of that is much bigger because it's not just about breaking the person yesterday, it's also about telling the person tomorrow that, you know what, this is how we do things, we're not going to break you when you come on board because there are lots of new people who are also going to come on board.

Speaker 3

36:32

You know, 1 way I like to think about this, and I always push the team to think about it as well, when you want to do new things, you want to start with a clean slate. Design with a clean slate in mind, And then we'll figure out how to make sure all the other things work. And yes, we do make compromises occasionally, but unless you design with the clean slate and not worry about that, you'll never get to a good place.

Speaker 2

36:58

Oh, that's brilliant. So even if you are responsible in the idea stage, when you're thinking of new, to put all that behind you. That's really well put.

Speaker 2

37:09

So I have to ask this, because a lot of students, developers ask me, how I feel about PyTorch versus TensorFlow. So I've recently completely switched my research group to TensorFlow. I wish everybody would just use the same thing and TensorFlow is as close to that, I believe, as we have. But do you enjoy competition?

Speaker 2

37:31

So TensorFlow is leading in many ways, in many dimensions, in terms of ecosystem, in terms of number of users, momentum, power, production level, so on, but you know, a lot of researchers are now also using PyTorch.

Speaker 3

37:45

Do you

Speaker 2

37:46

enjoy that kind of competition or do you just ignore it and focus on making TensorFlow the best that it can be? So

Speaker 3

37:52

just like research or anything people are doing, right, it's great to get different kinds of ideas. And when we started with TensorFlow, like I was saying earlier, it was very important for us to also have production in mind. We didn't want just research, right?

Speaker 3

38:08

And that's why we chose certain things. Now, PyTorch came along and said, you know what? I only care about research. This is what I'm trying to do.

Speaker 3

38:16

What's the best thing I can do for this? And it started iterating and said, OK, I don't need to worry about graphs. Let me just run things. I don't care if it's not as fast as it can be, but let me just make this part easy.

Speaker 3

38:30

And there are things you can learn from that, right? They again had the benefit of seeing what had come before, but also exploring certain different kinds of spaces and they had some good things there, building on say things like JNR and so on before that. So competition is definitely interesting. It made us, you know, this is an area that we had thought about, like I said, you know, very early on.

Speaker 3

38:53

Over time, we had revisited this a couple of times, should we add this again? At some point we said, you know what, here's, it seems like this can be done well, so let's try it again, and that's how we started pushing on eager execution, how do we combine those 2 together, which has finally come very well together in 2.0, but it took us a while to get all the things together and so on.

Speaker 2

39:16

So let me, I mean, ask, put another way, I think eager execution is a really powerful thing that was added. You think it wouldn't have been, you know, Muhammad Ali versus Fraser, right? You think it wouldn't have been added as quickly if PyTorch wasn't there?

Speaker 3

39:34

It might have taken longer. No longer. Yeah, it was, I mean, we had tried some variants of that before, so I'm sure it would have happened, but it might have taken longer.

Speaker 2

39:42

I'm grateful that TensorFlow responded in the way they did. It's doing some incredible work last couple years. What other things that we didn't talk about are you looking forward in 2.0 that comes to mind?

Speaker 2

39:53

So we talk about some of the ecosystem stuff, making it easily accessible through Keras, Eker execution, Is there other things that we miss?

Speaker 3

40:02

Yeah, so I would say 1 is just where 2.0 is, and you know, with all the things that we've talked about. I think as we think beyond that, there are lots of other things that it enables us to do, and that we're excited about. So what it's setting us up for, OK, here are these really clean APIs.