TensorFlow Lite: Reference app gallery (TF Dev Summit '20)

Key Takeaways

[Music] hi everyone welcome I'm ever Matt Eska and I'm a technical program manager on the tensorflow team I'm Luke and a software engineer from Tesla fell light what are you going to be showing us today Luke and I like to show guys the tests for light example apps tests of all I build more than 10 example apps that shows not only how to use model but for end-to-end code that a developer will need to write where can I find the source code for the for these models you can check out all the examples in the code from our website tensa folder or / like / examples and those examples are works for cross-platform for example Android iOS raspberry pi and even as GPU okay awesome I'm excited to see some of these absolutely so I'm going to show you three an example apps today before I going deep into each of them I like to point out that the devices right now is on airplane mode which means the chip I model is running purely on device without talking to any servers that's great so it does that mean I don't have to worry about privacy exactly so the user privacy is really protected Oh fantastic let's go into the first app it's a question and answer and joy app what motor does it use it's powered by mobile bird so mobile Birds recent article and it can answer a given question based on articles would you like to try an article first sure let's try the Amazon rainforest one and would you like to ask question there okay what is the size of the Amazon rainforest let's see if it figures that out five million five hundred thousand square kilometers that's pretty big so mobile bird is able to return the correct answer in just the 70 milliseconds that's awesome and why did you choose this after to create as a model yeah so it's demonstrated it actually demonstrate the power of TF light so TF lightest has been spent a significant effort to support all those state of the art models such as bird more birds mr. bird Albert and efficient ad do you have any apps to show me that work on iOS yeah the second app I like to show is the PostNet it actually available for both Android and iOS so the version we're looking at here is the iOS app so PostNet it continuously detected up to 17 key points of a person in real time would you like to try it out sure you'll need to point this at you okay hey I see circus I see lines wow how is it doing it so fast in real time it runs tfy interpreter with Koriyama delegate which really accelerate the performance okay cool that's pretty fast yes let me show you a third app which the stout transfer it takes accountant image and their style image and then outputs are sterilised image based on the specified styles that's pretty neat yeah let me try it yeah you can try to take a picture of me let me take a picture right I have a picture which style and going to pick for me what style would you like you want starry night yeah ah there you are cool that's pretty neat it's good let's go how can I find out more yeah if you like to see more examples feel free to visit our website at tensile org slash light slash examples there's a lot more code documentation and examples available and also if you like to download approaching models feel free to visit TF hub okay awesome thank you so much and thank you for joining us thanks [Music] you [Music]