Update: 26 April, 2023
This repo is a TensorFlow managed fork of the tflite_flutter_plugin project by the amazing Amish Garg. The goal of this project is to support our Flutter community in creating machine-learning backed apps with the TensorFlow Lite framework.
This project is currently a work-in-progress as we update it to create a working plugin that meets the latest and greatest Flutter and TensorFlow Lite standards. That said, pull requests and contributions are more than welcome and will be reviewed by TensorFlow or Flutter team members. We thank you for your understanding as we make progress on this update.
Feel free to reach out to us by posting in the issues or discussion areas.
Thanks!
- PaulTR
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, Metal and CoreML delegates on iOS, and XNNPack delegate on Desktop platforms.
- Multi-platform Support for Android and iOS
- Flexibility to use any TFLite Model.
- Acceleration using multi-threading.
- Similar structure as TensorFlow Lite Java API.
- Inference speeds close to native Android Apps built using the Java API.
- Run inference in different isolates to prevent jank in UI thread.
Examples and support now support dynamic library downloads! iOS samples can be run with the commands
flutter build ios
& flutter install ios
from their respective iOS folders.
Android can be run with the commands
flutter build android
& flutter install android
while devices are plugged in.
Note: This requires a device with a minimum API level of 26.
Note: TFLite may not work in the iOS simulator. It's recommended that you test with a physical device.
When creating a release archive (IPA), the symbols are stripped by Xcode, so the command flutter build ipa
may throw a Failed to lookup symbol ... symbol not found
error. To work around this:
- In Xcode, go to Target Runner > Build Settings > Strip Style
- Change from All Symbols to Non-Global Symbols
For MacOS a TensorFlow Lite dynamic library needs to be added to the project manually.
For this, first a .dylib
needs to be built. You can follow the Bazel build guide or the CMake build guide to build the libraries.
CMake Note:
-
cross compiling in CMake can be achieved using:
-DCMAKE_OSX_ARCHITECTURES=x86_64|arm64
-
bundling two architectures (arm / x86) using lipo:
lipo -create arm64/libtensorflowlite_c.dylib x86/libtensorflowlite_c.dylib -output libtensorflowlite_c.dylib
As a second step, the library needs to be added to your application's XCode project. For this, you can follow the step 1 and 2 of the official Flutter guide on adding dynamic libraries.
For Linux a TensorFlow Lite dynamic library needs to be added to the project manually.
For this, first a .so
needs to be built. You can follow the Bazel build guide or the CMake build guide to build the libraries.
As a second step, the library needs to be added to your application's project. This is a simple procedure
- Create a folder called
blobs
in the top level of your project - Copy the
libtensorflowlite_c-linux.so
to this folder - Append following lines to your
linux/CMakeLists.txt
...
# get tf lite binaries
install(
FILES ${PROJECT_BUILD_DIR}/../blobs/libtensorflowlite_c-linux.so
DESTINATION ${INSTALL_BUNDLE_DATA_DIR}/../blobs/
)
For Windows a TensorFlow Lite dynamic library needs to be added to the project manually.
For this, first a .dll
needs to be built. You can follow the Bazel build guide or the CMake build guide to build the libraries.
As a second step, the library needs to be added to your application's project. This is a simple procedure
- Create a folder called
blobs
in the top level of your project - Copy the
libtensorflowlite_c-win.dll
to this folder - Append following lines to your
windows/CMakeLists.txt
...
# get tf lite binaries
install(
FILES ${PROJECT_BUILD_DIR}/../blobs/libtensorflowlite_c-win.dll
DESTINATION ${INSTALL_BUNDLE_DATA_DIR}/../blobs/
)
The helper library has been deprecated. New development underway for a replacement at https://github.com/google/flutter-mediapipe. Current timeline is to have wide support by the end of August, 2023.
import 'package:tflite_flutter/tflite_flutter.dart';
In the dependency section of pubspec.yaml
file, add tflite_flutter: ^0.10.1
(adjust the version accordingly based on the latest release)
-
From asset
Place
your_model.tflite
inassets
directory. Make sure to include assets inpubspec.yaml
.final interpreter = await tfl.Interpreter.fromAsset('assets/your_model.tflite');
Refer to the documentation for info on creating interpreter from buffer or file.
-
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.filled(1*2, 0).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>.filled(1, 0); var output1 = List<double>.filled(1, 0); // output: Map<int, Object> var outputs = {0: output0, 1: output1}; // inference interpreter.runForMultipleInputs(inputs, outputs); // print outputs print(outputs)
interpreter.close();
To utilize asynchronous inference, first create your Interpreter
and then wrap it with IsolateInterpreter
.
final interpreter = await Interpreter.fromAsset('assets/your_model.tflite');
final isolateInterpreter =
await IsolateInterpreter.create(address: interpreter.address);
Both run
and runForMultipleInputs
methods of isolateInterpreter
are asynchronous:
await isolateInterpreter.run(input, output);
await isolateInterpreter.runForMultipleInputs(inputs, outputs);
By using IsolateInterpreter
, the inference runs in a separate isolate. This ensures that the main isolate, responsible for UI tasks, remains unblocked and responsive.