Laurence is a developer advocate at Google working on machine learning and artificial intelligence. He's the author of dozens of programming books, and hundreds of articles. When not Googling, he's author of a best-selling Science Fiction book series, and a produced screenwriter.View the profile
About the talk
In this special live episode from TensorFlow Dev Summit ‘19, Paige (@DynamicWebPaige) and Laurence (@lmoroney) answer your #AskTensorFlow questions!
Learn about using callbacks to cancel training once you’ve reached your desired accuracy, how to get started with TensorFlow 2.0 if you’re new to machine learning, and we’ll show you a fun example of image classification in the browser.
Hi, everybody exciting event so far and it's great to see all the questions that have been asked on social media. Please keep asking them and put in hashtag tensorflow. We're here today to answer them live on the live stream Laurence Maroney and I'm Paige Bailey. And remember it is hashtag ask tensorflow because we're the tensorflow that you should be asking if it's the mall, but we'll do our best. So should we take the first question? Absolutely. Let's go for one of the first questions are on today. I get it almost every time I meet
is it's great to be able to do training and like I usually do my training for fixed number of epochs. But what happens when I reach my desired accuracy mastering, how do I cancel training confused? If you've already gotten to a point where your boss would say? Okay cool 99% accuracy, that's fine for us today. A collab hear that you can see where I'm using call backs on call back to the way that you would actually achieve this. So at the top of my collab here, you can see I have class my
call back on on this one then and when Annie Hawkins in training I'm able to take a look at the logs. And if the accuracy log for example in this case is greater than 60% really nice far. He's really happy when my training 60% accurate cancel the training and then to be able to set that up. I'll just say I'm going to create an object called callbacks which is an instance of my callback class. And I love the way and call Dad when I double click it actually highlights just a little thing that I like and then down here on call backs. I'll just say call backs equals call backs. And then when I
actually do the training, you know what I'm going to really show off. I'm going to make my run time type to be a GPS or goes nice and fast. So this is fashion mnist. Let's do a little bit of training on fashion mnist with this one. I'm with doing this live. So I'm connecting up to the to the VM and then we go now, it's actually it's raining. It's getting ready to stop. Thunder first epochs we're on the first Epoch. The first Epoch is progressing away 60,000 images being trained and boom I hit accuracy of 83 and you can see now that it actually reach
60% accuracy. So it's cancel the training so callbacks are your friends if you doing this, certainly when you're learning when you're experimenting I used to before I learned about cult got before I learned about call backs before I learned about callbacks. It's like I would like set something up to train for like a hundred bucks and then go and go to sleep and then wake up the next morning and found like after 3 job and I wasted my time. So use call back an electric stove could be
useful for your particular workflow. And this code also would work and tensorflow to. Out so it's it's Karis. Gosh, I love cares. Clever fair right now things about Cara ask that you may not realize and we've been talkin bout to go to the same code that you're right for 10 slow One. X is the same code for 2. Oh what's going on behind the scenes is that is executing eagerly into. Out instead of a graph mode. So even though this collab, I think I'm running 1.13. This code will actually still run into. Without you my Define to go. Absolutely. All right. So should we take
the next question next question is from Twitter. It looks like and what about all the web developers to renew to does the version 2.0 help them get started Center play jazz team about all of the cool new demos that they've seen a rise from the community and its really such a vibrant ecosystem of artists and creators that are using browser-based or even server-based. How to create how to create these machine learning models training and running them for a little bit for a I said so
Tell me more. I'm intrigued to work on the fire by steam. So shout out to all my friends and Firebase, but I'm so it's for mobile developers and for web developers and one of the things that Firebase gives you these things called Cloud functions for Firebase. I called up the webpage here with you around putting in summary. What they do is that they allow you to execute functions on a back-end without you needing to maintain a server infrastructure and allow you to execute these in response to a trigger. So trigger might be for
example, an analytic event or signing event process it Bingo man. I should try this out for machine learning for me that's amazing web developers that has lots of great options for you. However, did you want to do it in the browser on the back end and mobile that kind of stuff? Hopefully this about the great stuff that you'll be able to get started. And then the question about tensorflow too. Oh and whether it gives additional tools for for application developers. I think it would mostly be in terms of those codes in tutorials that were mentioned before we've also
really some courses. So it's it's easier than ever to get started and models that you create and save model can be deployed to TS light to 10. About Kara's but I think that we love to find your layers and an AR developer like they have carrots for our which is awesome. It was created by JJ Allaire and Friends wash away. They have a they have a book out about it. So web developers lots of options for you. Should we take the next question and this looks like it also came from Twitter and its are there any 10 slow.
JS is popular, right we have learned so object detection. So, how do I answer this? It depends on what you mean by object in Google we talked about object detection. We use a specific term for an image where you got lots of objects me for bounding boxes around them right now. There are no samples for that. But and lovely thing is that the community is incredibly Adept at creating tensorflow JS examples for expert had a great recent example, I'm with using it to track hand movements and a proud.
I'd like to show a demo transfer learning with being able to detect a single item with an A-frame took to do we call that an image classification. So can I I love I'm a child of the 80s. I love Pac-Man if you look carefully it says actually loading mobilenet now. So what's happening is it that just downloaded the mobilenet model? So what I'm going to do is I'm going to add some new classes to the mobilenet model and then use transfer money to get that one down left and right you're trying to run away from the go. So I'm going to try and train
it to move up when I point up like that. So I'm holding it down and I'm gathering a bunch of samples are about 50 samples. And then when I go right like this, I didn't really think this one through though because then turning left is going to be hard, but bear with me I'm like 15, maybe I'll do the last like this and get my head out of the way, you know, she samples like that. I'm in town will look like this. Hopefully these aren't rude gestures some country and something like that in the browser with Aulani so if you look over in the
last year my learning right my back size. I'm just going to try to eat box. I'm going to start training and we'll see it starts going quickly. You see my lost started at 4 and I went down down down down down. That's a zero-sum a digit be on the six digits is never actually at zero but we say we have a very low loss and that we can actually give it a try. So let me see if I can avoid getting eaten by ghost. So I'm going to stop playing the game and I'm going to move laughed and you can see that bounding box around it kind of shows that I'm up. I'm
question then just images to so so make sure to check out all of the great examples that we've got listed on the website like Twitter. I love Twitter. Are you going to publish the updated version of tensorflow for poets tutorial from Pete Worden implementing TF2. O t a flight to go and a lot of others. We have the neural network API Android I didn't get to see it. So do you want to take a shower or so at some point? We will update it. I don't think there's an updated version available right now, but one of the
things that I really liked about the pencil football, what score was it got me building a model very quickly that I could then use on a mobile device. But the downside of that was it was a bunch of scripts that I ran and now we're going on with them. So when the things that we've been doing is that we've decided to get a whole bunch of new tensorflow Lite examples and put them online on the site and I have them on here. So this for new ones gesture recognition image classification object detection and speech recognition. What's nice about these are all open source, they're both Android
and iOS and they include full instructions on how to build a bear sound. Is it going to try to run that in my website? I don't want that late. I want to try and run that in my Android emulator so we can see it running an emulated environment. Let me get that started and we can see it being hashed out here. I'm actually running to text it to water bottle. Now. This is running in the Android emulator. Okay. This is using tensorflow lights. This is the sample is on there that basically just the same thing that you would have seen intensive
Gulf of poets words using mobile Madden building of an application around mobilenet. But if you look even running in the emulator, I'm getting in France X in like that hundred hundred seventy-five scale models and then sort of pull them down to a manageable size on a mobile and or an embedded devices huge and really excited to see what time's our flight does this year. So we're working on a bunch of new tutorials. Those are out there. If you take a look at that GitHub page, you'll see that there's example of their there's details and how it's built.
Let me just go back on here. So for example, if I say Explorer on Android, you'll see this details and how it's billed how you can put it all together how you can compile it. You don't need to build tensorflow in order to use tensorflow Lite that was one bit of confusion that folks had in the past. Now, it's just a case of what you learn iOS Developer the pods that you had that kind of stuff can go on like start kicking the tires on these applications for yourself. All righty, and we will have more code labs and I would love to get Pete stencil loaf of poets
Buy this talk
Access to all the recordings of the event
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.