On-Device Activity Classifier (Expo + TensorFlow Lite)

This project demonstrates how to train a machine-learning model on accelerometer data and deploy it on a mobile phone for real-time, offline inference using React Native, Expo Dev Client, and TensorFlow Lite.

The model runs entirely on the device without sending any data to a server.

---

1. Project Overview

• Platform: React Native (Expo)

• Sensors: Android Accelerometer

• ML Framework (training): TensorFlow (Python)

• ML Runtime (mobile): TensorFlow Lite

• Inference: On-device, offline

• Build system: Expo Dev Client + Android Studio

---

2. Initial App Setup (Expo)

1. Create an Expo project:

   npx create-expo-app ActivityClassifier cd ActivityClassifier

2. Run the app with Expo Go to confirm setup:

   npx expo start

At this stage, the app is JavaScript-only and cannot run native ML code.

---

3. Sensor Data Collection (Accelerometer)

Inside the main screen:

• Used expo-sensors to access the accelerometer

• Collected (x, y, z) values at ~20 Hz

• Maintained a rolling buffer of the last 128 samples

• Added UI buttons to:

  • Start / Stop recording

  • Record labeled windows (Still, Walking, Running)

Each labeled sample stored:

• 128 accelerometer samples

• Activity label

• Timestamp

Data was saved locally as:

activity_dataset.json

This file was exported from the phone and used for training.

---

4. Model Training (Python + TensorFlow)

Environment

• Python 3.10 (virtual environment)

• TensorFlow 2.15

Training Steps

1. Load activity_dataset.json

2. Convert samples into:

   • Input shape: (128, 3)

   • Labels: Still / Walking / Running

3. Train a simple neural network classifier

4. Save the trained model as a TensorFlow SavedModel

Output:

saved_model_activity/

---

5. Convert TensorFlow Model to TensorFlow Lite

Using TensorFlow Lite for mobile inference:

import tensorflow as tf converter = tf.lite.TFLiteConverter.from_saved_model("saved_model_activity") converter.optimizations = [tf.lite.Optimize.DEFAULT] tflite_model = converter.convert() with open("activity.tflite", "wb") as f: f.write(tflite_model)

Output:

activity.tflite

---

6. Add Native Support (Expo Dev Client)

TensorFlow Lite requires native code, so Expo Go cannot be used.

Steps:

1. Install dev client:

   npx expo install expo-dev-client

2. Prebuild native project:

   npx expo prebuild

3. Build and install app:

   npx expo run:android

This generated an Android app with native modules enabled.

---

7. Install TensorFlow Lite Runtime

Installed a native TFLite binding for React Native:

npm install react-native-fast-tflite npx expo prebuild --clean npx expo run:android

This library provides:

• Model loading

• Synchronous inference

• Access to output tensors

---

8. Bundle the TFLite Model

1. Place the model inside the project:

   assets/images/models/activity.tflite

2. Configure Metro to allow .tflite assets:

metro.config.js

const { getDefaultConfig } = require("expo/metro-config"); const config = getDefaultConfig(__dirname); config.resolver.assetExts.push("tflite"); module.exports = config;

3. Restart Metro:

   npx expo start -c

---

9. Load the Model in React Native

In the main screen:

import { loadTensorflowModel } from "react-native-fast-tflite"; const model = await loadTensorflowModel( require("../../assets/images/models/activity.tflite") );

A successful load confirms that the model is available on the device.

---

10. Run Real-Time Inference

Inference Pipeline

1. Collect accelerometer samples continuously

2. Keep the latest 128 samples

3. Every ~1 second:

   • Flatten samples into a Float32Array

   • Run inference:

     const outputs = model.runSync([input]);

   • Select highest probability output

   • Map index to label (Still / Walking / Running)

4. Update UI with predicted activity and confidence

The model now runs fully offline on the phone.

---

11. Result

• The app correctly detects:

  • Still

  • Walking

  • Running

• Inference happens locally

• No internet connection required

• No sensor data leaves the device

---

12. What This Demonstrates

• End-to-end ML pipeline

• Real sensor data collection

• Model training and optimization

• Native mobile inference

• Privacy-preserving, offline AI

This is the same architecture used in real fitness, health, and activity-tracking applications.

---

13. Possible Extensions

• Prediction smoothing / majority voting

• Activity history logging

• Background tracking

• Fall detection

• BLE health device integration

• Medical or fitness dashboards

---

14. Summary

This project shows how to go from:

• Raw sensor data
  to

• A trained ML model
  to

• On-device inference in a mobile app

without relying on cloud services or external APIs.