react-native-tensorflow-lite

A react native library for running Tensorflow Lite Image Recognition on Android app.

Installing

$ npm install react-native-tensorflow-lite --save

Linking

$ react-native link react-native-tensorflow-lite

Converting your model to tflite format

Follow this guide: https://www.tensorflow.org/lite/convert/cmdline_examples

Android

• Open up android/app/src/main/java/[...]/MainActivity.java
• Add import com.reactlibrary.RNTensorflowLitePackage; to the imports at the top of the file
• Add new RNTensorflowLitePackage() to the list returned by the getPackages() method
• Add the following lines to your app's build.gradle(android/app/build.gradle):

  android {
  aaptOptions {
  noCompress 'tflite'
  noCompress 'lite'
  }
  }
  

Usage

Place your tflite model file and labels.txt in your app's asset folder.

import{TFLiteImageRecognition}from'react-native-tensorflow-lite';classMyImageClassifierextendsComponent{constructor(){super()this.state={}try{// Initialize Tensorflow Lite Image Recognizerthis.classifier=newTFLiteImageRecognition({model: "mymodel.tflite",// Your tflite model in assets folder.labels: "label.txt"// Your label file})}catch(err){alert(err)}}componentWillMount(){this.classifyImage("apple.jpg")// Your image path.}asyncclassifyImage(imagePath){try{constresults=awaitthis.classifier.recognize({image: imagePath,// Your image path.inputShape: 224,// the input shape of your model. If none given, it will be default to 224.})constresultObj={name: "Name: "+results[0].name,confidence: "Confidence: "+results[0].confidence,inference: "Inference: "+results[0].inference+"ms"};this.setState(resultObj)}catch(err){alert(err)}}componentWillUnmount(){this.classifier.close()// Must close the classifier when destroying or unmounting component to release object.}render(){return(<Viewstyle={styles.container}><View><Textstyle={styles.results}>{this.state.name}</Text><Textstyle={styles.results}>{this.state.confidence}</Text><Textstyle={styles.results}>{this.state.inference}</Text></View></View>);}}

Things to note

• Sometimes, when using the float model the tensorflow lite inference is slower than using the ordinary tensorflow mobile as discussed in this issue tensorflow/tensorflow#21787