dwat27 / tflite_flutter_plugin Goto Github PK

TensorFlow Lite Flutter plugin provides a flexible and fast solution for accessing TensorFlow Lite interpreter and performing inference. The API is similar to the TFLite Java and Swift APIs. It directly binds to TFLite C API making it efficient (low-latency). Offers acceleration support using NNAPI, GPU delegates on Android, and Metal delegate on iOS.

• Flexibility to use any TFLite Model.
• Acceleration using multi-threading and delegate support.
• Similar structure as TensorFlow Lite Java API.
• Inference speeds close to native Android Apps built using the Java API.
• You can choose to use any TensorFlow version by building binaries locally.
• Run inference in different isolates to prevent jank in UI thread.

1. Place the script install.sh (Linux/Mac) or install.bat (Windows) at the root of your project.

2. Execute sh install.sh (Linux) / install.bat (Windows) at the root of your project to automatically download and place binaries at appropriate folders.

   Note: The binaries installed will not include support for GpuDelegateV2 and NnApiDelegate however InterpreterOptions().useNnApiForAndroid can still be used.

3. Use sh install.sh -d (Linux) or install.bat -d (Windows) instead if you wish to use these GpuDelegateV2 and NnApiDelegate.

These scripts install pre-built binaries based on latest stable tensorflow release. For info about using other tensorflow versions refer to this part of readme.

Title	Code	Demo	Blog
Text Classification App	Code		Blog/Tutorial
Image Classification App	Code		-
Object Detection App	Code		Blog/Tutorial

import 'package:tflite_flutter/tflite_flutter.dart';

• From asset

  Place your_model.tflite in assets directory. Make sure to include assets in pubspec.yaml.

  final interpreter = await tfl.Interpreter.fromAsset('your_model.tflite');

Refer to the documentation for info on creating interpreter from buffer or file.

See TFLite Flutter Helper Library for easy processing of input and output.

• For single input and output

  Use void run(Object input, Object output).

  // For ex: if input tensor shape [1,5] and type is float32
  var input = [[1.23, 6.54, 7.81. 3.21, 2.22]];
  
  // if output tensor shape [1,2] and type is float32
  var output = List(1*2).reshape([1,2]);
  
  // inference
  interpreter.run(input, output);
  
  // print the output
  print(output);

• For multiple inputs and outputs

  Use void runForMultipleInputs(List<Object> inputs, Map<int, Object> outputs).

  var input0 = [1.23];  
  var input1 = [2.43];  
  
  // input: List<Object>
  var inputs = [input0, input1, input0, input1];  
  
  var output0 = List<double>(1);  
  var output1 = List<double>(1);
  
  // output: Map<int, Object>
  var outputs = {0: output0, 1: output1};
  
  // inference  
  interpreter.runForMultipleInputs(inputs, outputs);
  
  // print outputs
  print(outputs)

interpreter.close();

Note: This feature is under testing and could be unstable with some builds and on some devices.

• NNAPI delegate for Android

  var interpreterOptions = InterpreterOptions()..useNnApiForAndroid = true;
  final interpreter = await Interpreter.fromAsset('your_model.tflite',
      options: interpreterOptions);
  

  or

  var interpreterOptions = InterpreterOptions()..addDelegate(NnApiDelegate());
  final interpreter = await Interpreter.fromAsset('your_model.tflite',
      options: interpreterOptions);
  

• GPU delegate for Android and iOS

  • Android GpuDelegateV2

    final gpuDelegateV2 = GpuDelegateV2(
            options: GpuDelegateOptionsV2(
            false,
            TfLiteGpuInferenceUsage.fastSingleAnswer,
            TfLiteGpuInferencePriority.minLatency,
            TfLiteGpuInferencePriority.auto,
            TfLiteGpuInferencePriority.auto,
        ));
    
    var interpreterOptions = InterpreterOptions()..addDelegate(gpuDelegateV2);
    final interpreter = await Interpreter.fromAsset('your_model.tflite',
        options: interpreterOptions);

  • iOS Metal Delegate (GpuDelegate)

    final gpuDelegate = GpuDelegate(
          options: GpuDelegateOptions(true, TFLGpuDelegateWaitType.active),
        );
    var interpreterOptions = InterpreterOptions()..addDelegate(gpuDelegate);
    final interpreter = await Interpreter.fromAsset('your_model.tflite',
        options: interpreterOptions);

Refer Tests to see more example code for each method.

The pre-built binaries are updated with each stable tensorflow release. However, you many want to use latest unstable tf releases or older tf versions, for that proceed to build locally, if you are unable to find the required version in release assets.

Make sure you have required version of bazel installed. (Check TF_MIN_BAZEL_VERSION, TF_MAX_BAZEL_VERSION in configure.py)

• Android

Configure your workspace for android builds as per these instructions.

For TensorFlow >= v2.2

    bazel build -c opt --cxxopt=--std=c++11 --config=android_arm //tensorflow/lite/c:tensorflowlite_c

    // similarily for arm64 use --config=android_arm64

For TensorFlow <= v2.1

    bazel build -c opt --cxxopt=--std=c++11 --config=android_arm //tensorflow/lite/experimental/c:libtensorflowlite_c.so

    // similarily for arm64 use --config=android_arm64

• iOS

Refer instructions on TensorFlow Lite website to build locally for iOS.

Note: You must use macOS for building iOS.

tflite_flutter dynamically links to C APIs which are supplied in the form of libtensorflowlite_c.so on Android and TensorFlowLiteC.framework on iOS.

For Android, We need to manually download these binaries from release assets and place the libtensorflowlite_c.so files in the <root>/android/app/src/main/jniLibs/ directory for each arm, arm64, x86, x86_64 architecture as done here in the example app.  

• Enabling support for Flutter Desktop Applications.
• Better and more precise error handling.

• Tian LIN, Jared Duke, Andrew Selle, YoungSeok Yoon, Shuangfeng Li from the TensorFlow Lite Team for their invaluable guidance.
• Authors of dart-lang/tflite_native.