Pocket AutoML: Tutorial for Creating an Android App for Image Classification with Deep Learning

Translations

• English (this document)
• Русский
• 简体中文

Overview

This document will walk you through the steps for creating your Android app that runs a deep learning image classification model trained in Pocket AutoML and exported in LiteRT (formerly TensorFlow Lite) .tflite format. The app will continuously classify whatever it sees from the device's back camera.

This tutorial is based on the Google AI Edge LiteRT image classification Android example.

Note: This app was modernized in 2026 — it is now written in Kotlin and uses LiteRT (formerly TensorFlow Lite; the CompiledModel API) for inference, CameraX for the camera, and a current Gradle/AGP build. The previous version (Java, Camera2, and the legacy TensorFlow Lite Task Library) is preserved at the pre-modernization tag.

If you have any issues following this tutorial please contact me (the creator of Pocket AutoML) via email at or by creating a GitHub issue.

Requirements

• Android Studio Panda 4 (2025.3.4) or newer (installed on a Linux, Mac or Windows machine)

• [if not using an Android emulator] an Android device in developer mode with USB debugging enabled and a USB cable (to connect an Android device to your computer)

Step 1. Train a model in Pocket AutoML

• Install Pocket AutoML from Google Play Store and open it

• Create a task e.g. Kittens or Puppies by pressing + button

• Create a class e.g. Kittens

  

• Add example images of the class by taking photos with a camera or picking them from a storage

  

• Go back to the task view by pressing <- and repeat these steps for each class

• Go back to the task view by pressing <-, switch to the MODEL tab and press TRAIN

  

Step 2. Export a model in TF Lite format from Pocket AutoML

• Press EXPORT IN TENSORFLOW LITE FORMAT

• The model is downloaded to your device's Downloads folder by the system download manager. Swipe down from the top of the screen to watch the download progress in the notification drawer. The export takes a few minutes.

  

• When the download is done, open the Downloads folder (e.g. with the Files app), find <your_task_name>.tflite and transfer it to your PC — for example by sending it to yourself via an email app like GMail or uploading it to cloud storage like Google Drive or Dropbox.

Step 3. Clone the Pocket AutoML example source code

Run the following command to get the demo application.

git clone https://github.com/OutSorcerer/pocket-automl-android-tutorial

Open the example source code in Android Studio. To do this, open Android Studio and select Open an existing project, setting the folder to pocket-automl-android-tutorial

This app runs inference with the LiteRT CompiledModel API. On the GPU it forces full FP32 precision (LiteRT's GpuOptions), which is what lets the EfficientNet-based Pocket AutoML model run on the GPU without producing NaN scores. Class labels and the input normalization (mean/std) are read directly from the metadata embedded in the .tflite model, so there is no separate labels file to manage at runtime (see Step 6).

Step 4. Build the Android Studio project

Select Build -> Assemble Project (you can also click the hammer 🔨 icon in the toolbar or press Ctrl+F9 / Cmd+F9) and check that the project builds successfully. Android Studio will prompt you to download any missing components (such as the Android SDK and build tools).

Step 5. Install and run the app

Follow this step to make sure that the example runs successfully in your environment using its built-in models. The following step will demonstrate how to add your custom model from Pocket AutoML into the example app.

Run on a device

If you are willing to test the app on an Android device, connect the device to the computer and be sure to approve any ADB permission prompts that appear on your phone. Select your device in the target device dropdown at the top of Android Studio (connected devices also appear in the Device Manager). Then click Run -> Run 'app' from the main menu of Android Studio to build and install the app on the selected device.

Run on an an emulator

If you are willing to test the app on an Android emulator

• open Tools -> Device Manager and click + -> Create Virtual Device
• choose a device definition e.g. Medium Phone (this controls its screen resolution and density) and click Next
• select the API level — Android 16 (API level 36) or newer is recommended to match the app's target SDK (the app requires API level 24 or higher)
• select a system image, preferably the one marked with a star, which is recommended for your system
• click Finish
• select the newly created device in the Device Manager and click Run -> Run 'app' from the main menu of Android Studio

If you want to know more, see Create and manage virtual devices in Android documentation.

To test the app, open the app called Pocket AutoML Predictor on your device or emulator. When you run the app the first time, the app will request permission to access the camera. Re-installing the app may require you to uninstall the previous installations.

Step 6. Add your model from Pocket AutoML into the example app

• At this point you must have your <your_task_name>.tflite model exported from Pocket AutoML. The class labels are embedded in the model's metadata, which the LiteRT Metadata library reads directly at runtime — there is nothing else to prepare.

• Copy <your_task_name>.tflite into pocket-automl-android-tutorial/app/src/main/assets

• Open ImageClassifierHelper.kt (by clicking Navigate -> Search Everywhere or pressing Shift twice and typing its name) and, in setupImageClassifier(), point the MODEL_POCKET_AUTOML branch of the when (currentModel) block at your file: MODEL_POCKET_AUTOML -> "<your_task_name>.tflite"

• Run the app. The Pocket AutoML model is selected by default; you can also swipe up the bottom sheet and pick it from the ML Model dropdown menu.

• You will see the predicted class and its probability, with the other classes below. Well done!

  

Next steps

Applications

Do you have a task at hand that can be solved with a help of an image classification model running in a mobile app? It could be sorting lego bricks or controlling a robot with hand gestures.

I will be excited to know what you have built with the help of Pocket AutoML and this tutorial and will add links to Play Store or GitHub into this document.

Other platforms

TF Lite can run not only on Android but on other platforms as well including iOS, embedded Linux devices like Raspberry Pi or Coral and microcontrollers.

Other model training methods

You can try other no-code or low-code deep learning solutions like Teachable Machine, Roboflow, Create ML, Google Cloud AutoML, Azure Custom Vision, LandingLens, Edge Impulse or MediaPipe Model Maker.

Pocket AutoML uses transfer learning approach, you can also implement it yourself using a tutorial Transfer learning and fine-tuning in Google Colab.

Learning deep learning

If you want to learn how to train better models and have a systematic understanding or deep learning I recommend Deep Learning Specialization and Machine Learning Engineering for Production (MLOps) Specialization on Coursera.

License

Apache License 2.0