/flutter_tflite

Flutter plugin for TensorFlow Lite

Primary LanguageJavaMIT LicenseMIT

tflite

A Flutter plugin for accessing TensorFlow Lite API. Supports image classification, object detection (SSD and YOLO), Pix2Pix and Deeplab and PoseNet on both iOS and Android.

Table of Contents

Breaking changes since 1.0.0:

  1. Updated to TensorFlow Lite API v1.12.0.
  2. No longer accepts parameter inputSize and numChannels. They will be retrieved from input tensor.
  3. numThreads is moved to Tflite.loadModel.

Installation

Add tflite as a dependency in your pubspec.yaml file.

Android

In android/app/build.gradle, add the following setting in android block.

    aaptOptions {
        noCompress 'tflite'
        noCompress 'lite'
    }

iOS

If you get error like "'vector' file not found", please open ios/Runner.xcworkspace in Xcode, click Runner > Tagets > Runner > Build Settings, search Compile Sources As, change the value to Objective-C++;

Usage

  1. Create a assets folder and place your label file and model file in it. In pubspec.yaml add:
  assets:
   - assets/labels.txt
   - assets/mobilenet_v1_1.0_224.tflite
  1. Import the library:
import 'package:tflite/tflite.dart';
  1. Load the model and labels:
String res = await Tflite.loadModel(
  model: "assets/mobilenet_v1_1.0_224.tflite",
  labels: "assets/labels.txt",
  numThreads: 1 // defaults to 1
);
  1. See the section for the respective model below.

  2. Release resources:

await Tflite.close();

Image Classification

  • Output fomart:
{
  index: 0,
  label: "person",
  confidence: 0.629
}
  • Run on image:
var recognitions = await Tflite.runModelOnImage(
  path: filepath,   // required
  imageMean: 0.0,   // defaults to 117.0
  imageStd: 255.0,  // defaults to 1.0
  numResults: 2,    // defaults to 5
  threshold: 0.2,   // defaults to 0.1
  asynch: true      // defaults to true
);
  • Run on binary:
var recognitions = await Tflite.runModelOnBinary(
  binary: imageToByteListFloat32(image, 224, 127.5, 127.5),// required
  numResults: 6,    // defaults to 5
  threshold: 0.05,  // defaults to 0.1
  asynch: true      // defaults to true
);

Uint8List imageToByteListFloat32(
    img.Image image, int inputSize, double mean, double std) {
  var convertedBytes = Float32List(1 * inputSize * inputSize * 3);
  var buffer = Float32List.view(convertedBytes.buffer);
  int pixelIndex = 0;
  for (var i = 0; i < inputSize; i++) {
    for (var j = 0; j < inputSize; j++) {
      var pixel = image.getPixel(j, i);
      buffer[pixelIndex++] = (img.getRed(pixel) - mean) / std;
      buffer[pixelIndex++] = (img.getGreen(pixel) - mean) / std;
      buffer[pixelIndex++] = (img.getBlue(pixel) - mean) / std;
    }
  }
  return convertedBytes.buffer.asUint8List();
}

Uint8List imageToByteListUint8(img.Image image, int inputSize) {
  var convertedBytes = Uint8List(1 * inputSize * inputSize * 3);
  var buffer = Uint8List.view(convertedBytes.buffer);
  int pixelIndex = 0;
  for (var i = 0; i < inputSize; i++) {
    for (var j = 0; j < inputSize; j++) {
      var pixel = image.getPixel(j, i);
      buffer[pixelIndex++] = img.getRed(pixel);
      buffer[pixelIndex++] = img.getGreen(pixel);
      buffer[pixelIndex++] = img.getBlue(pixel);
    }
  }
  return convertedBytes.buffer.asUint8List();
}
  • Run on image stream (video frame):

Works with camera plugin 4.0.0. Video format: (iOS) kCVPixelFormatType_32BGRA, (Android) YUV_420_888.

var recognitions = await Tflite.runModelOnFrame(
  bytesList: img.planes.map((plane) {return plane.bytes;}).toList(),// required
  imageHeight: img.height,
  imageWidth: img.width,
  imageMean: 127.5,   // defaults to 127.5
  imageStd: 127.5,    // defaults to 127.5
  rotation: 90,       // defaults to 90, Android only
  numResults: 2,      // defaults to 5
  threshold: 0.1,     // defaults to 0.1
  asynch: true        // defaults to true
);

Object Detection

  • Output fomart:

x, y, w, h are between [0, 1]. You can scale x, w by the width and y, h by the height of the image.

{
  detectedClass: "hot dog",
  confidenceInClass: 0.123,
  rect: {
    x: 0.15,
    y: 0.33,
    w: 0.80,
    h: 0.27
  }
}

SSD MobileNet:

  • Run on image:
var recognitions = await Tflite.detectObjectOnImage(
  path: filepath,       // required
  model: "SSDMobileNet",
  imageMean: 127.5,     
  imageStd: 127.5,      
  threshold: 0.4,       // defaults to 0.1
  numResultsPerClass: 2,// defaults to 5
  asynch: true          // defaults to true
);
  • Run on binary:
var recognitions = await Tflite.detectObjectOnBinary(
  binary: imageToByteListUint8(resizedImage, 300), // required
  model: "SSDMobileNet",  
  threshold: 0.4,                                  // defaults to 0.1
  numResultsPerClass: 2,                           // defaults to 5
  asynch: true                                     // defaults to true
);
  • Run on image stream (video frame):

Works with camera plugin 4.0.0. Video format: (iOS) kCVPixelFormatType_32BGRA, (Android) YUV_420_888.

var recognitions = await Tflite.detectObjectOnFrame(
  bytesList: img.planes.map((plane) {return plane.bytes;}).toList(),// required
  model: "SSDMobileNet",  
  imageHeight: img.height,
  imageWidth: img.width,
  imageMean: 127.5,   // defaults to 127.5
  imageStd: 127.5,    // defaults to 127.5
  rotation: 90,       // defaults to 90, Android only
  numResults: 2,      // defaults to 5
  threshold: 0.1,     // defaults to 0.1
  asynch: true        // defaults to true
);

Tiny YOLOv2:

  • Run on image:
var recognitions = await Tflite.detectObjectOnImage(
  path: filepath,       // required
  model: "YOLO",      
  imageMean: 0.0,       
  imageStd: 255.0,      
  threshold: 0.3,       // defaults to 0.1
  numResultsPerClass: 2,// defaults to 5
  anchors: anchors,     // defaults to [0.57273,0.677385,1.87446,2.06253,3.33843,5.47434,7.88282,3.52778,9.77052,9.16828]
  blockSize: 32,        // defaults to 32
  numBoxesPerBlock: 5,  // defaults to 5
  asynch: true          // defaults to true
);
  • Run on binary:
var recognitions = await Tflite.detectObjectOnBinary(
  binary: imageToByteListFloat32(resizedImage, 416, 0.0, 255.0), // required
  model: "YOLO",  
  threshold: 0.3,       // defaults to 0.1
  numResultsPerClass: 2,// defaults to 5
  anchors: anchors,     // defaults to [0.57273,0.677385,1.87446,2.06253,3.33843,5.47434,7.88282,3.52778,9.77052,9.16828]
  blockSize: 32,        // defaults to 32
  numBoxesPerBlock: 5,  // defaults to 5
  asynch: true          // defaults to true
);
  • Run on image stream (video frame):

Works with camera plugin 4.0.0. Video format: (iOS) kCVPixelFormatType_32BGRA, (Android) YUV_420_888.

var recognitions = await Tflite.detectObjectOnFrame(
  bytesList: img.planes.map((plane) {return plane.bytes;}).toList(),// required
  model: "YOLO",  
  imageHeight: img.height,
  imageWidth: img.width,
  imageMean: 0,         // defaults to 127.5
  imageStd: 255.0,      // defaults to 127.5
  numResults: 2,        // defaults to 5
  threshold: 0.1,       // defaults to 0.1
  numResultsPerClass: 2,// defaults to 5
  anchors: anchors,     // defaults to [0.57273,0.677385,1.87446,2.06253,3.33843,5.47434,7.88282,3.52778,9.77052,9.16828]
  blockSize: 32,        // defaults to 32
  numBoxesPerBlock: 5,  // defaults to 5
  asynch: true          // defaults to true
);

Pix2Pix

Thanks to RP from Green Appers

  • Output format:

    The output of Pix2Pix inference is Uint8List type. Depending on the outputType used, the output is:

    • (if outputType is png) byte array of a png image

    • (otherwise) byte array of the raw output

  • Run on image:

var result = await runPix2PixOnImage(
  path: filepath,       // required
  imageMean: 0.0,       // defaults to 0.0
  imageStd: 255.0,      // defaults to 255.0
  asynch: true      // defaults to true
);
  • Run on binary:
var result = await runPix2PixOnBinary(
  binary: binary,       // required
  asynch: true      // defaults to true
);
  • Run on image stream (video frame):
var result = await runPix2PixOnFrame(
  bytesList: img.planes.map((plane) {return plane.bytes;}).toList(),// required
  imageHeight: img.height, // defaults to 1280
  imageWidth: img.width,   // defaults to 720
  imageMean: 127.5,   // defaults to 0.0
  imageStd: 127.5,    // defaults to 255.0
  rotation: 90,       // defaults to 90, Android only
  asynch: true        // defaults to true
);

Deeplab

Thanks to RP from see-- for Android implementation.

  • Output format:

    The output of Deeplab inference is Uint8List type. Depending on the outputType used, the output is:

    • (if outputType is png) byte array of a png image

    • (otherwise) byte array of r, g, b, a values of the pixels

  • Run on image:

var result = await runSegmentationOnImage(
  path: filepath,     // required
  imageMean: 0.0,     // defaults to 0.0
  imageStd: 255.0,    // defaults to 255.0
  labelColors: [...], // defaults to https://github.com/shaqian/flutter_tflite/blob/master/lib/tflite.dart#L219
  outputType: "png",  // defaults to "png"
  asynch: true        // defaults to true
);
  • Run on binary:
var result = await runSegmentationOnBinary(
  binary: binary,     // required
  labelColors: [...], // defaults to https://github.com/shaqian/flutter_tflite/blob/master/lib/tflite.dart#L219
  outputType: "png",  // defaults to "png"
  asynch: true        // defaults to true
);
  • Run on image stream (video frame):
var result = await runSegmentationOnFrame(
  bytesList: img.planes.map((plane) {return plane.bytes;}).toList(),// required
  imageHeight: img.height, // defaults to 1280
  imageWidth: img.width,   // defaults to 720
  imageMean: 127.5,        // defaults to 0.0
  imageStd: 127.5,         // defaults to 255.0
  rotation: 90,            // defaults to 90, Android only
  labelColors: [...],      // defaults to https://github.com/shaqian/flutter_tflite/blob/master/lib/tflite.dart#L219
  outputType: "png",       // defaults to "png"
  asynch: true             // defaults to true
);

PoseNet

Model is from StackOverflow thread.

  • Output format:

x, y are between [0, 1]. You can scale x by the width and y by the height of the image.

[ // array of poses/persons
  { // pose #1
    score: 0.6324902,
    keypoints: {
      0: {
        x: 0.250,
        y: 0.125,
        part: nose,
        score: 0.9971070
      },
      1: {
        x: 0.230,
        y: 0.105,
        part: leftEye,
        score: 0.9978438
      }
      ......
    }
  },
  { // pose #2
    score: 0.32534285,
    keypoints: {
      0: {
        x: 0.402,
        y: 0.538,
        part: nose,
        score: 0.8798978
      },
      1: {
        x: 0.380,
        y: 0.513,
        part: leftEye,
        score: 0.7090239
      }
      ......
    }
  },
  ......
]
  • Run on image:
var result = await runPoseNetOnImage(
  path: filepath,     // required
  imageMean: 125.0,   // defaults to 125.0
  imageStd: 125.0,    // defaults to 125.0
  numResults: 2,      // defaults to 5
  threshold: 0.7,     // defaults to 0.5
  nmsRadius: 10,      // defaults to 20
  asynch: true        // defaults to true
);
  • Run on binary:
var result = await runPoseNetOnBinary(
  binary: binary,     // required
  numResults: 2,      // defaults to 5
  threshold: 0.7,     // defaults to 0.5
  nmsRadius: 10,      // defaults to 20
  asynch: true        // defaults to true
);
  • Run on image stream (video frame):
var result = await runPoseNetOnFrame(
  bytesList: img.planes.map((plane) {return plane.bytes;}).toList(),// required
  imageHeight: img.height, // defaults to 1280
  imageWidth: img.width,   // defaults to 720
  imageMean: 125.0,        // defaults to 125.0
  imageStd: 125.0,         // defaults to 125.0
  rotation: 90,            // defaults to 90, Android only
  numResults: 2,           // defaults to 5
  threshold: 0.7,          // defaults to 0.5
  nmsRadius: 10,           // defaults to 20
  asynch: true             // defaults to true
);

Example

Prediction in Static Images

Refer to the example.

Real-time detection

Refer to flutter_realtime_Detection.