NumberDetector

Using ML in an Android demo app - The Good, The Bad, and the Ugly

MassTLC Machine Learning Developer Day, January 24, 2019

The app demonstrates using a TensorFlow Lite model trained on the MNIST dataset to recognize hand-drawn digits on an Android.

We use ML Kit for Firebase to host the model.

Credit and Acknowledgement: Mark Allison's blog posts and source code are the inspiration and source for this talk and demo app. https://blog.stylingandroid.com/ml-for-android-developers-part-1-2/

NOTE: The google-services.json file obtained from my Firebase account is not checked in. Here are steps to obtain your own google-services.json file, as well as to connect the demo app to Firebase.

1. Create a Firebase account: sign in at console.firebase.google.com > Add project e.g. NumberDetector > Download google-services.json > follow paged instructions to update the app project in Android Studio
2. Last step: connect the app to Firebase by running the app from IDE > Run | Run app
3. But Catch-22 is: what if app is not ready/debugged? A hack to try: use Firebase Assistant in Android Studio to pick any (benign) Firebase feature to connect the app. From Android Studio: > Tool | Firebase > e.g. Test Lab > Run (click on link) Firebase Test Lab from Android Studio > Connect app to Firebase