How TensorFlow keeps improving (TensorFlow Meets)

Key Takeaways

[Music] hi everybody and welcome to tents Flo meets it's my great honor today to be chatting with Megan Cachola engineering director right on tents Flo so can you tell us a little bit more about what it is that you do sure so I'm an engineering director on the tensorflow team I've been with the team for a little over two and a half years now I've gotten to participate in all of the dev summits which has been really exciting nice um just to get to come be part of the keynote be part of the event in terms of what I actually do so I'm one of the leads for the tensorflow team just helped with obviously the team itself make sure we're going in the right direction and make sure that we're doing the right things by our community so one of the things that you spoke about that I saw that a lot of people are excited about and I know I'm personally excited about it's just really how we're doubling down on making it very developer friendly yes flow so how did that come about and can you tell us a bit more about it obviously we've taken a lot of feedback from the community just to see like what are the pain points what works well what doesn't I think that's a big part of it and just in general tensorflow has been around like I said we just had our third birthday in November right so things changed or at the industry is really moving quickly machine learning is advancing in lots of different places than we might have anticipated when tensorflow was first developed so we need to make sure that we're making it very very easy for folks to just come in and get started and be able to take advantage of all of the cool things happening in the machine learning space right so I think really think it comes from kind of both of those angles just how things have moved so fast and just the feedback we've gotten from the community as we've had tensorflow out there for a few years yeah I mean I've been working on it for about a year and a half and I came in with a developer background and when I started trying to kick the tires on tensorflow there were there were some things where it was like it just wasn't quite intuitive to me but a lot of that has been changing right this stuff like Harris and other things and the eager execution by default that you've been adding just hopefully will make a lot easier for other developers that's correct we want to make sure obviously that the flexibility and the power of tensor flow is is there but also make sure it's approachable and easy for people just come in and be like it's fine start here on this surface and if you need to dive down to this part and really get into the guts of it you can and it's fantastic but you don't have to if you don't need to right and all of this doesn't come at the cost of performance right because you had that great slide with performance and can you tell us a little bit about the important performance improvements that you've been seeing with this yeah so a lot of those improvements as well we're looking at just kind of the core tensorflow all right kind of if you think of like the heart of it and that's where the majority of the improvements come meaning that those improvements will be applicable whether we're talking about tensorflow to NATO whether we're talking about using carrots or something else it's it's that big meaty part kind of under the hood and a lot of it comes from just better making use of the different types of hardware making sure that we're using the different types of accelerators appropriately understanding the limitations and restrictions for things like mobile all right we had talked about a lot of improvements for performance on TF light as well and again some of it is just understanding what are the work clothes really look like what does it look like on some of these different devices and edge TB use and things that we're trying in tensorflow out more with now and then kind of always going back and closing the loop and being like okay this is how we can make it better and this is how we can make sure that the users get that performance by default and don't have to necessarily know what magic had to happen under the covers for it to happen they should just get it right now you've mentioned like on mobile with tens flow lights and we lovingly call it CF light and the one of the things that I do get a lot of questions about is that there's all of these different almost parts of the family of tensorflow and can you give us a quick summary of like the different runtimes that are available for developers like between sense of low light and what others are other so there's tensorflow light there's more higher level types of things that we talked about what tensorflow extended so that's not necessarily a different runtime but it's more kind of just how do you put things together into end especially if you're looking more towards like a production environment we talked a bit about JavaScript as well so tensorflow j/s and I think that one it's a language type thing right because you want to make sure that the JavaScript community it's such a large community that they have access to machine learning as well but there's also just the hole in browser experience that it plugs in well with because of nodejs so I think a lot of it depends on what kind of applications people are in arrested in and kind of where they're coming at machine learning from there's so many different ways of applying machine learning whether you're thinking of it from like oh you know I work in a large company so I needed for these enterprise use cases Errol I'm just an app developer I'm trying things out myself right or I'm someone who's really familiar with its list language I'm really familiar with JavaScript so that means I can use that as kind of my entry way to start doing other things in machine learning so I think some of it depends on where you're at where you're coming from and then we try and make sure we have the right way for you to get into the machine learning community that the JavaScript stuff is amazing in the fact that you can actually train models in the browser yes you know when I first heard about that I was like nah come on but you actually can do that it's really cool and as you mentioned them with node you're not just limited to in the browser you can also run it on backend servers and that kind of stuff and one of my personal favorites actually is with things like cloud functions or cloud functions for firebase cuz they run node.js yes you can actually start putting models in there no tense flow to is currently an alpha release right now lots of people are asking what comes next when are we going to see a release candidate or something like that yeah so like we talked about at the dev summit we're going to have our release candidate kind of coming up sometime in q2 we want to make sure that we're giving ourselves enough time to make sure we have good performance make sure we have the right fit finish and polish for everything but there is like like you said an alpha available we want feedback from the community we want to understand what are people enjoying what's working well what are they concerned about and you can also follow along on github and kind of see like what features are coming out what status are things at that way the community knows what's going on and we can make sure that we're engaged appropriately to be building the right things as we finalize and finish up with the two data release sounds perfect thank you so people can stay tuned to github and keep track of everything that's going on yes so that's correct okay so thank you so much Megan and thanks everybody for watching this episode of tensorflow meats if you've got any questions for me or any questions for Megan just please leave them in the comments below and check out the github page for tensorflow if you want to learn more about the tensorflow 2.0 release thank you so much [Music] you [Music]